repo
stringlengths
2
99
file
stringlengths
13
225
code
stringlengths
0
18.3M
file_length
int64
0
18.3M
avg_line_length
float64
0
1.36M
max_line_length
int64
0
4.26M
extension_type
stringclasses
1 value
XFL
XFL-master/python/algorithm/model/densenet.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # This model contains a PyTorch implementation of the paper "Densely Connected Convolutional Networks."[1] # [1]Huang, G., Liu, Z., Weinberger, K. Q., & van der Maaten, L. (2016). Densely connected convolutional networks. arXiv preprint arXiv:1608.06993. from collections import OrderedDict import math import torch.nn as nn import torch import torch.nn.functional as F class BottleNeckBlock(nn.Module): expansion = 4 def __init__(self, in_planes, growth_rate, drop_out=0.0): super().__init__() self.conv_block1 = nn.Sequential(OrderedDict([ ("batch_norm1", nn.BatchNorm2d(in_planes, track_running_stats=True)), # setting track_running_stats as False ("relu1", nn.ReLU()), ("conv1", nn.Conv2d(in_planes, self.expansion*growth_rate, kernel_size=1, stride=1, bias=False)) ])) self.conv_block2 = nn.Sequential(OrderedDict([ ("batch_norm2", nn.BatchNorm2d(self.expansion*growth_rate, track_running_stats=True)), ("relu2", nn.ReLU()), ("conv2", nn.Conv2d(self.expansion*growth_rate, growth_rate, kernel_size=3, stride=1, padding=1, bias=False)), ])) self.drop_out = drop_out def forward(self, x): out = self.conv_block1(x) if self.drop_out: out = F.dropout(out, p=self.drop_out, training=self.training) out = self.conv_block2(out) if self.drop_out: out = F.dropout(out, p=self.drop_out, training=self.training) return torch.cat([x, out], 1) class TransitionBlock(nn.Module): def __init__(self, in_planes, out_planes, drop_out=0.0): super().__init__() self.bn = nn.BatchNorm2d(in_planes) self.relu = nn.ReLU() self.conv = nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=1, padding=0, bias=False) self.drop_out = drop_out def forward(self, x): out = self.conv(self.relu(self.bn(x))) if self.drop_out: out = F.dropout(out, p=self.drop_out, training=self.training) return F.avg_pool2d(out, 2) class Densenet(nn.Module): def __init__(self, block, block_list, num_classes, growth_rate=12, reduction=0.5, drop_out=0.0): super().__init__() self.growth_rate = growth_rate self.drop_out = drop_out in_planes = 2 * growth_rate self.conv = nn.Conv2d(3, in_planes, kernel_size=3, padding=1, bias=False) self.dense_layer1 = self._make_layers(block, block_list[0], in_planes) in_planes += block_list[0]*growth_rate self.transition1 = TransitionBlock(in_planes, int(math.floor(in_planes*reduction)), drop_out=drop_out) in_planes = int(math.floor(in_planes*reduction)) self.dense_layer2 = self._make_layers(block, block_list[1], in_planes) in_planes += block_list[1]*growth_rate self.transition2 = TransitionBlock(in_planes, int(math.floor(in_planes*reduction)), drop_out=drop_out) in_planes = int(math.floor(in_planes*reduction)) self.dense_layer3 = self._make_layers(block, block_list[2], in_planes) in_planes += block_list[2]*growth_rate self.transition3 = TransitionBlock(in_planes, int(math.floor(in_planes*reduction)), drop_out=drop_out) in_planes = int(math.floor(in_planes*reduction)) self.dense_layer4 = self._make_layers(block, block_list[3], in_planes) in_planes += block_list[3]*growth_rate self.batchnorm = nn.BatchNorm2d(in_planes) self.relu = nn.ReLU() self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) self.fc = nn.Linear(in_planes, num_classes) def forward(self, x): x = self.conv(x) x = self.transition1(self.dense_layer1(x)) x = self.transition2(self.dense_layer2(x)) x = self.transition3(self.dense_layer3(x)) x = self.dense_layer4(x) x = self.relu(self.batchnorm(x)) x = self.avgpool(x) x = x.view(x.shape[0], -1) x = self.fc(x) return x def _make_layers(self, block, blocks, in_planes): layers = [] for i in range(blocks): layers.append(block(in_planes, self.growth_rate, drop_out=self.drop_out)) in_planes += self.growth_rate return nn.Sequential(*layers) def DenseNet(num_classes, layers): if layers == 121: return Densenet(BottleNeckBlock, [6,12,24,16], num_classes, growth_rate=32) elif layers == 169: return Densenet(BottleNeckBlock, [6,12,32,32], num_classes, growth_rate=32) elif layers == 201: return Densenet(BottleNeckBlock, [6,12,48,32], num_classes, growth_rate=32) elif layers == 264: return Densenet(BottleNeckBlock, [6,12,64,48], num_classes, growth_rate=32) elif layers == 'unit_test': return Densenet(BottleNeckBlock, [2,2,2,2], num_classes, growth_rate=8) else: raise NotImplementedError("Only support DenseNet121, DenseNet169, DenseNet201, DenseNet264 currently, please change layers")
5,629
41.651515
147
py
XFL
XFL-master/python/algorithm/model/resnet_paddle.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # This model contains a PaddlePaddle implementation of the paper "Deep Residual Learning for Image Recognition."[1] # [1]He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778). from collections import OrderedDict import paddle from paddle import nn class ConvBlock(nn.Layer): expansion = 4 def __init__(self, in_channels, out_channels, downsample=None, stride=1): super().__init__() self.stem = nn.Sequential( ("conv1", nn.Conv2D(in_channels, out_channels, kernel_size=1, stride=1, padding=0)), ("batch_norm1", nn.BatchNorm2D(out_channels)), ("relu1", nn.ReLU()), ("conv2", nn.Conv2D(out_channels, out_channels, kernel_size=3, stride=stride, padding=1)), ("batch_norm2", nn.BatchNorm2D(out_channels)), ("relu2", nn.ReLU()), ("conv3", nn.Conv2D(out_channels, out_channels * self.expansion, kernel_size=1, stride=1, padding=0)), ("batch_norm3", nn.BatchNorm2D(out_channels*self.expansion)) ) self.downsample = downsample self.stride = stride self.relu = nn.ReLU() def forward(self, x): residual = x x = self.stem(x) if self.downsample is not None: residual = self.downsample(residual) x += residual x = self.relu(x) return x class Resnet(nn.Layer): def __init__(self, ResBlock, block_list, num_classes): super().__init__() self.stem = nn.Sequential( ("conv1", nn.Conv2D(3, 64, kernel_size=3, stride=1, padding=1, bias_attr=False)), ("batch_norm1", nn.BatchNorm2D(64)), ("relu", nn.ReLU()) ) self.max_pool = nn.MaxPool2D(kernel_size=3, stride=2, padding=1) self.layers1 = self._make_layers( ResBlock, block_list[0], inplanes=64, outplanes=64, stride=1) self.layers2 = self._make_layers( ResBlock, block_list[1], inplanes=256, outplanes=128, stride=2) self.layers3 = self._make_layers( ResBlock, block_list[2], inplanes=512, outplanes=256, stride=2) self.layers4 = self._make_layers( ResBlock, block_list[3], inplanes=1024, outplanes=512, stride=2) self.avgpool = nn.AdaptiveAvgPool2D((1, 1)) self.fc = nn.Linear(512*ResBlock.expansion, num_classes) def forward(self, x): x = self.stem(x) x = self.max_pool(x) x = self.layers1(x) x = self.layers2(x) x = self.layers3(x) x = self.layers4(x) x = self.avgpool(x) x = x.reshape([x.shape[0], -1]) x = self.fc(x) return x def _make_layers(self, ResBlock, blocks, inplanes, outplanes, stride=1): layers =[] downsample = None if stride != 1 or inplanes != outplanes*ResBlock.expansion: downsample = nn.Sequential( nn.Conv2D(inplanes, outplanes*ResBlock.expansion, kernel_size=1, stride=stride), nn.BatchNorm2D(outplanes*ResBlock.expansion) ) layers.append(ResBlock(inplanes, outplanes, downsample=downsample, stride=stride)) for i in range(1, blocks): layers.append(ResBlock(outplanes*ResBlock.expansion, outplanes)) return nn.Sequential(*layers) def ResNet(num_classes, layers): if layers == "unit_test": return Resnet(ConvBlock, [2, 2, 2, 2], num_classes) elif layers == 50: return Resnet(ConvBlock, [3, 4, 6, 3], num_classes) elif layers == 101: return Resnet(ConvBlock, [3, 4, 23, 3], num_classes) elif layers == 152: return Resnet(ConvBlock, [3, 8, 36, 3], num_classes) else: raise NotImplementedError("Only support ResNet50, ResNet101, ResNet152 currently, please change layers")
4,619
36.868852
192
py
XFL
XFL-master/python/algorithm/model/linear_regression.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import torch import torch.nn as nn class LinearRegression(nn.Module): def __init__(self, input_dim: int, bias: bool = False): super(LinearRegression, self).__init__() self.linear = torch.nn.Linear(input_dim, 1, bias=bias) def forward(self, x): return self.linear(x)
907
32.62963
74
py
XFL
XFL-master/python/algorithm/model/poisson_regression.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import torch import torch.nn as nn class PoissonRegression(nn.Module): def __init__(self, input_dim: int, bias: bool = False): super(PoissonRegression, self).__init__() self.linear = torch.nn.Linear(input_dim, 1, bias=bias) def forward(self, x): return torch.exp(self.linear(x))
917
33
74
py
XFL
XFL-master/python/algorithm/model/logistic_regression.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import torch.nn as nn class LogisticRegression(nn.Module): def __init__(self, input_dim: int, bias: bool = False): super().__init__() self.linear = nn.Linear(input_dim, 1, bias=bias) self.sigmoid = nn.Sigmoid() def forward(self, x): x = self.linear(x) x = self.sigmoid(x) return x
954
31.931034
74
py
XFL
XFL-master/python/algorithm/model/__init__.py
0
0
0
py
XFL
XFL-master/python/algorithm/framework/__init__.py
0
0
0
py
XFL
XFL-master/python/algorithm/framework/horizontal/__init__.py
0
0
0
py
XFL
XFL-master/python/algorithm/framework/horizontal/poisson_regression/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from functools import partial from algorithm.core.horizontal.template.agg_type import \ register_agg_type_for_label_trainer from common.utils.logger import logger from .common import Common class HorizontalPoissonRegressionLabelTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] self.register_hook( place="after_train_loop", rank=1, func=partial(self.val_loop, "train"), desc="validation on trainset" ) register_agg_type_for_label_trainer(self, 'torch', agg_type) def train_loop(self): self.model.train() train_loss = 0 lossfunc = list(self.lossfunc.values())[0] optimizer = list(self.optimizer.values())[0] lr_scheduler = list(self.lr_scheduler.values())[0] if self.lr_scheduler.values() else None for batch, (feature, label) in enumerate(self.train_dataloader): pred = self.model(feature) loss = lossfunc(pred, label) optimizer.zero_grad() loss.backward() optimizer.step() train_loss += loss.item() train_loss /= len(self.train_dataloader) if lr_scheduler: lr_scheduler.step() self.context["train_loss"] = train_loss logger.info(f"Train loss: {train_loss}")
2,035
36.018182
98
py
XFL
XFL-master/python/algorithm/framework/horizontal/poisson_regression/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from functools import partial from .common import Common from algorithm.core.horizontal.template.agg_type import register_agg_type_for_assist_trainer class HorizontalPoissonRegressionAssistTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] register_agg_type_for_assist_trainer(self, 'torch', agg_type) self.register_hook(place="after_local_epoch", rank=1, func=partial(self._save_model, False), desc="save model ") self.register_hook(place="after_local_epoch", rank=2, func=partial(self.val_loop, "val"), desc="validation on valset") self.register_hook(place="after_global_epoch", rank=1, func=partial(self._save_model, True), desc="save final model") def train_loop(self): pass
1,539
43
92
py
XFL
XFL-master/python/algorithm/framework/horizontal/poisson_regression/common.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import numpy as np import torch import torch.nn as nn from torch.utils.data import DataLoader, TensorDataset from algorithm.core.data_io import CsvReader from algorithm.core.horizontal.template.torch.base import BaseTrainer from algorithm.model.poisson_regression import PoissonRegression from common.checker.x_types import All from common.evaluation.metrics import CommonMetrics from common.utils.config_sync import ConfigSynchronizer from common.utils.logger import logger class Common(BaseTrainer): def __init__(self, train_conf: dict): sync_rule = { "model_info": { "config": All() }, "train_info": { "interaction_params": All(), "train_params": { "global_epoch": All(), "aggregation": All(), "encryption": All(), "optimizer": All(), "lr_scheduler": All(), "lossfunc": All(), "metric": All(), "early_stopping": All() } } } train_conf = ConfigSynchronizer(train_conf).sync(sync_rule) super().__init__(train_conf) def _set_model(self) -> nn.Module: model_config = self.common_config.model_info.get("config") model = PoissonRegression(input_dim=model_config["input_dim"], bias=model_config["bias"]) return model def _read_data(self, input_dataset): if len(input_dataset) == 0: return None conf = input_dataset[0] if conf["type"] == "csv": path = os.path.join(conf['path'], conf['name']) has_label = conf["has_label"] has_id = conf['has_id'] return CsvReader(path, has_id, has_label) else: return None def _set_train_dataloader(self): train_data = self._read_data(self.common_config.input_trainset) trainset = None train_dataloader = None if train_data: trainset = TensorDataset( torch.tensor(train_data.features(), dtype=torch.float32).to(self.device), torch.tensor(train_data.label(), dtype=torch.float32).unsqueeze(dim=-1).to(self.device) ) batch_size = self.common_config.train_params.get("train_batch_size") if trainset: train_dataloader = DataLoader(trainset, batch_size, shuffle=True) return train_dataloader def _set_val_dataloader(self): val_data = self._read_data(self.common_config.input_valset) valset = None val_dataloader = None if val_data: valset = TensorDataset( torch.tensor(val_data.features(), dtype=torch.float32).to(self.device), torch.tensor(val_data.label(), dtype=torch.float32).unsqueeze(dim=-1).to(self.device) ) batch_size = self.common_config.train_params.get("val_batch_size") if valset: val_dataloader = DataLoader(valset, batch_size, shuffle=True) return val_dataloader def val_loop(self, dataset_type: str = "val", context: dict = {}): self.model.eval() val_loss = 0 val_predicts = [] labels = [] lossfunc_name = list(self.lossfunc.keys())[0] lossfunc = list(self.lossfunc.values())[0] if dataset_type == "val": dataloader = self.val_dataloader elif dataset_type == "train": dataloader = self.train_dataloader else: raise ValueError(f"dataset type {dataset_type} is not valid.") for batch, (feature, label) in enumerate(dataloader): pred = self.model(feature) loss = lossfunc(pred, label) val_predicts.append(pred.detach().cpu().squeeze(-1).numpy()) val_loss += loss.item() labels.append(label.cpu().squeeze(-1).numpy()) val_loss /= len(dataloader) labels: np.ndarray = np.concatenate(labels, axis=0) val_predicts: np.ndarray = np.concatenate(val_predicts, axis=0) metrics_output = CommonMetrics._calc_metrics( metrics=self.metrics, labels=labels, val_predicts=val_predicts, lossfunc_name=lossfunc_name, loss=val_loss, dataset_type=dataset_type ) global_epoch = self.context["g_epoch"] if dataset_type == "val": local_epoch = None elif dataset_type == "train": local_epoch = self.context["l_epoch"] CommonMetrics.save_metric_csv( metrics_output=metrics_output, output_config=self.common_config.output, global_epoch=global_epoch, local_epoch=local_epoch, dataset_type=dataset_type, ) early_stop_flag = self.context["early_stop_flag"] if (self.common_config.save_frequency > 0) & \ (dataset_type == "val") & (self.earlystopping.patience > 0): early_stop_flag = self.earlystopping(metrics_output, global_epoch) if early_stop_flag: # find the saved epoch closest to the best epoch best_epoch = self.earlystopping.best_epoch closest_epoch = round(best_epoch / self.common_config.save_frequency) * \ self.common_config.save_frequency closest_epoch -= self.common_config.save_frequency \ if closest_epoch > global_epoch else 0 self.context["early_stop_flag"] = True self.context["early_stop_epoch"] = closest_epoch
6,322
36.414201
103
py
XFL
XFL-master/python/algorithm/framework/horizontal/vgg/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from algorithm.core.horizontal.template.agg_type import \ register_agg_type_for_label_trainer from common.utils.logger import logger from .common import Common from functools import partial class HorizontalVggLabelTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] self.register_hook( place="after_train_loop", rank=1, func=partial(self.val_loop, "train"), desc="validation on trainset" ) register_agg_type_for_label_trainer(self, 'torch', agg_type) def train_loop(self): self.model.train() train_loss = 0 lossfunc = list(self.lossfunc.values())[0] optimizer = list(self.optimizer.values())[0] lr_scheduler = list(self.lr_scheduler.values())[0] if self.lr_scheduler.values() else None for batch, (feature, label) in enumerate(self.train_dataloader): pred = self.model(feature) loss = lossfunc(pred, label) optimizer.zero_grad() loss.backward() optimizer.step() train_loss += loss.item() train_loss /= len(self.train_dataloader) if lr_scheduler: lr_scheduler.step() self.context["train_loss"] = train_loss logger.info(f"Train loss: {train_loss}")
2,018
37.09434
98
py
XFL
XFL-master/python/algorithm/framework/horizontal/vgg/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from functools import partial from .common import Common from algorithm.core.horizontal.template.agg_type import \ register_agg_type_for_assist_trainer class HorizontalVggAssistTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] register_agg_type_for_assist_trainer(self, 'torch', agg_type) self.register_hook(place="after_local_epoch", rank=1, func=partial(self._save_model, False), desc="save model ") self.register_hook(place="after_local_epoch", rank=2, func=partial(self.val_loop, "val"), desc="validation on valset") self.register_hook(place="after_global_epoch", rank=1, func=partial(self._save_model, True), desc="save final model") def train_loop(self): pass
1,538
41.75
91
py
XFL
XFL-master/python/algorithm/framework/horizontal/vgg/common.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import numpy as np import torch import torch.nn as nn from torch.utils.data import DataLoader, TensorDataset from PIL import Image import torchvision.transforms as transforms from algorithm.core.data_io import CsvReader, NpzReader from algorithm.model.vgg import VGG from common.utils.logger import logger from algorithm.core.horizontal.template.torch.base import BaseTrainer from common.utils.config_sync import ConfigSynchronizer from common.checker.x_types import All from common.evaluation.metrics import CommonMetrics class Common(BaseTrainer): def __init__(self, train_conf: dict) -> None: sync_rule = { "model_info": All(), "train_info": { "interaction_params": All(), "train_params": { "global_epoch": All(), "aggregation": All(), "encryption": All(), "optimizer": All(), "lr_scheduler": All(), "lossfunc": All(), "metric": All(), "early_stopping": All() } } } train_conf = ConfigSynchronizer(train_conf).sync(sync_rule) super().__init__(train_conf) def _set_model(self) -> nn.Module: model_config = self.common_config.model_info.get("config") model = VGG(num_classes=model_config["num_classes"], layers=model_config["layers"]) model = model.to(self.device) return model def _read_data(self, input_dataset): if len(input_dataset) == 0: return None conf = input_dataset[0] if conf["type"] == "csv": path = os.path.join(conf['path'], conf['name']) has_label = conf["has_label"] has_id = conf['has_id'] return CsvReader(path, has_id, has_label) elif conf["type"] == "npz": path = os.path.join(conf['path'], conf['name']) return NpzReader(path) else: return None def _set_train_dataloader(self): def img_collate_fn(batch): labels = [] imgs = [] transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))]) for feature, label in batch: img = Image.fromarray(feature.numpy().astype(np.uint8)) imgs.append(transform_train(img)) labels.append(label) return torch.stack(imgs,0).to(self.device), torch.stack(labels, 0).long().to(self.device) train_data = self._read_data(self.common_config.input_trainset) trainset = None train_dataloader = None if train_data: trainset = TensorDataset( torch.tensor(train_data.features()[:100]), torch.tensor(train_data.label()[:100]) ) batch_size = self.common_config.train_params.get("train_batch_size") if trainset: train_dataloader = DataLoader( trainset, batch_size, shuffle=True, collate_fn=img_collate_fn ) return train_dataloader def _set_val_dataloader(self): def img_collate_fn(batch): labels = [] imgs = [] transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))]) for feature, label in batch: img = Image.fromarray(feature.numpy().astype(np.uint8)) imgs.append(transform_test(img)) labels.append(label) return torch.stack(imgs,0).to(self.device), torch.stack(labels, 0).long().to(self.device) val_data = self._read_data(self.common_config.input_valset) valset = None val_dataloader = None if val_data: valset = TensorDataset( torch.tensor(val_data.features()[:100]), torch.tensor(val_data.label()[:100]) ) batch_size = self.common_config.train_params.get("val_batch_size") if valset: val_dataloader = DataLoader( valset, batch_size, shuffle=False, collate_fn=img_collate_fn ) return val_dataloader def val_loop(self, dataset_type: str = "val", context: dict = {}): self.model.eval() val_loss = 0 val_predicts = [] labels = [] lossfunc_name = list(self.lossfunc.keys())[0] lossfunc = list(self.lossfunc.values())[0] if dataset_type == "val": dataloader = self.val_dataloader elif dataset_type == "train": dataloader = self.train_dataloader else: raise ValueError(f"dataset type {dataset_type} is not valid.") for batch, (feature, label) in enumerate(dataloader): pred = self.model(feature) loss = lossfunc(pred, label) val_predicts.append(pred.detach().cpu().squeeze(-1).numpy()) val_loss += loss.item() labels.append(label.cpu().squeeze(-1).numpy()) val_loss /= len(dataloader) labels: np.ndarray = np.concatenate(labels, axis=0) val_predicts: np.ndarray = np.concatenate(val_predicts, axis=0) if len(val_predicts.shape) == 1: val_predicts = np.array(val_predicts > 0.5, dtype=np.int32) elif len(val_predicts.shape) == 2: val_predicts = val_predicts.argmax(axis=-1) metrics_output = CommonMetrics._calc_metrics( metrics=self.metrics, labels=labels, val_predicts=val_predicts, lossfunc_name=lossfunc_name, loss=val_loss, dataset_type=dataset_type ) global_epoch = self.context["g_epoch"] if dataset_type == "val": local_epoch = None elif dataset_type == "train": local_epoch = self.context["l_epoch"] CommonMetrics.save_metric_csv( metrics_output=metrics_output, output_config=self.common_config.output, global_epoch=global_epoch, local_epoch=local_epoch, dataset_type=dataset_type, ) early_stop_flag = self.context["early_stop_flag"] if (self.common_config.save_frequency > 0) & \ (dataset_type == "val") & (self.earlystopping.patience > 0): early_stop_flag = self.earlystopping(metrics_output, global_epoch) if early_stop_flag: # find the saved epoch closest to the best epoch best_epoch = self.earlystopping.best_epoch closest_epoch = round(best_epoch / self.common_config.save_frequency) * \ self.common_config.save_frequency closest_epoch -= self.common_config.save_frequency \ if closest_epoch > global_epoch else 0 self.context["early_stop_flag"] = True self.context["early_stop_epoch"] = closest_epoch
7,957
37.819512
101
py
XFL
XFL-master/python/algorithm/framework/horizontal/vgg_jax/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from algorithm.core.horizontal.template.jax.fedtype import _get_label_trainer from common.utils.logger import logger from .common import Common from jax import jit, value_and_grad class HorizontalVggJaxLabelTrainer(Common, _get_label_trainer()): def __init__(self, train_conf: dict): _get_label_trainer().__init__(self, train_conf) self._set_jit_train_step() self._set_jit_val_step() def _set_jit_train_step(self): def train_step(batch, state): loss_fn = lambda params: self.calculate_loss(params, state.batch_stats, batch, train=True) ret, grads = value_and_grad(loss_fn, has_aux=True)(state.params) loss, _, new_model_state = ret[0], *ret[1] state = state.apply_gradients(grads=grads, batch_stats=new_model_state['batch_stats']) return loss, state self.jit_train_step = jit(train_step) def train_loop(self): train_loss = 0 for batch_id, batch in enumerate(self.train_dataloader): loss, self.state = self.jit_train_step(batch, self.state) train_loss += loss.item() train_loss /= len(self.train_dataloader) self.context["train_loss"] = train_loss logger.info(f"Train loss: {train_loss}")
1,891
40.130435
102
py
XFL
XFL-master/python/algorithm/framework/horizontal/vgg_jax/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from algorithm.core.horizontal.template.jax.fedtype import _get_assist_trainer from functools import partial from .common import Common class HorizontalVggJaxAssistTrainer(Common, _get_assist_trainer()): def __init__(self, train_conf: dict): _get_assist_trainer().__init__(self, train_conf) self._set_jit_val_step() self.register_hook(place="after_local_epoch", rank=2, func=partial(self.val_loop, "val"), desc="validation on valset") self.register_hook(place="after_local_epoch", rank=3, func=self._save_model, desc="save model") def train_loop(self): pass
1,286
38
91
py
XFL
XFL-master/python/algorithm/framework/horizontal/vgg_jax/common.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import numpy as np from jax import jit, random import jax.numpy as jnp import flax.linen as nn import torch from torch.utils.data import DataLoader, TensorDataset import torchvision.transforms as transforms from PIL import Image from flax.core.frozen_dict import FrozenDict from collections import OrderedDict from algorithm.core.data_io import NpzReader from algorithm.model.vgg_jax import vggjax from common.utils.logger import logger class Common(): def _set_model(self) -> nn.Module: model = None self.init_params = None self.init_batch_stats = None self.state = None exmp_features = self.exmp_label if self.exmp_label is not None else self.exmp_assist model_config = self.model_info.get("config") model = vggjax(num_classes=model_config["num_classes"], layers=model_config["layers"]) init_rng = random.PRNGKey(0) variables = model.init(init_rng, exmp_features, train=True) self.init_params, self.init_batch_stats = variables["params"], variables["batch_stats"] # init the state_dict and keys_dict used for aggregation self.state_dict = OrderedDict() self.keys_dict = OrderedDict() for key in ["params", "batch_stats"]: self.keys_dict[key] = OrderedDict() for i, j in variables[key].unfreeze().items(): self.keys_dict[key][i] = [] for k, v in j.items(): self.keys_dict[key][i].append(k) self.state_dict[i+k] = np.asarray(v, dtype=np.float32) return model def state_to_state_dict(self): for i, j in self.state.params.unfreeze().items(): for k, v in j.items(): self.state_dict[i+k] = np.asarray(v, dtype=np.float32) for i, j in self.state.batch_stats.unfreeze().items(): for k, v in j.items(): self.state_dict[i+k] = np.asarray(v, dtype=np.float32) def state_dict_to_state(self): new_state = dict() for key in ["params", "batch_stats"]: new_state[key] = dict() for i, j in self.keys_dict[key].items(): value_dict = dict() for k in j: value_dict[k] = jnp.asarray(self.state_dict[i+k], dtype=np.float32) new_state[key][i] = value_dict new_state = FrozenDict(new_state) self.state = self.state.replace(params=new_state["params"], batch_stats=new_state["batch_stats"]) def _read_data(self, input_dataset): if len(input_dataset) == 0: return None conf = input_dataset[0] if conf["type"] == "npz": path = os.path.join(conf['path'], conf['name']) return NpzReader(path) else: return None def _set_train_dataloader(self): def img_collate_fn(batch): labels = [] imgs = [] transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), np.array ]) for feature, label in batch: img = Image.fromarray(feature.numpy().astype(np.uint8)) img = transform_train(img) img = np.transpose(img, (1, 2, 0)) imgs.append(img) # [N, C, H, W] -> [N, H, W, C] labels.append(label.numpy()) return jnp.stack(imgs, 0).astype(jnp.float32), jnp.stack(labels, 0).astype(jnp.int32) train_data = self._read_data(self.input_trainset) exmp_features = None trainset = None train_dataloader = None if train_data: trainset = TensorDataset(torch.tensor(train_data.features()[0:100]), torch.tensor(train_data.label()[0:100])) exmp_features = jnp.ones_like(jnp.stack(train_data.features()[0:2], 0)) batch_size = self.train_params.get("batch_size", 64) if trainset: train_dataloader = DataLoader(trainset, batch_size, shuffle=True, collate_fn=img_collate_fn) return train_dataloader, exmp_features def _set_val_dataloader(self): def img_collate_fn(batch): labels = [] imgs = [] transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), np.array ]) for feature, label in batch: img = Image.fromarray(feature.numpy().astype(np.uint8)) img = transform_test(img) img = np.transpose(img, (1, 2, 0)) imgs.append(img) # [N, C, H, W] -> [N, H, W, C] labels.append(label.numpy()) return jnp.stack(imgs, 0).astype(jnp.float32), jnp.stack(labels, 0).astype(jnp.int32) val_data = self._read_data(self.input_valset) exmp_features = None valset = None val_dataloader = None if val_data: valset = TensorDataset(torch.tensor(val_data.features()[0:100]), torch.tensor(val_data.label()[0:100])) exmp_features = jnp.ones_like(jnp.stack(val_data.features()[0:2], 0)) batch_size = self.train_params.get("batch_size", 64) if valset: val_dataloader = DataLoader(valset, batch_size, shuffle=True, collate_fn=img_collate_fn) return val_dataloader, exmp_features def calculate_loss(self, params, batch_stats, batch, train): features, labels = batch # Run model. During training, we need to update the BatchNorm statistics. outputs = self.model.apply( {'params': params, 'batch_stats': batch_stats}, features, train=train, mutable=['batch_stats'] if train else False ) logits, new_model_state = outputs if train else (outputs, None) loss = self.loss_func(logits, labels).mean() preds = logits.argmax(axis=-1) return loss, (preds, new_model_state) def _set_jit_val_step(self): def val_step(batch, state): loss, (preds, _) = self.calculate_loss(state.params, state.batch_stats, batch, train=False) return loss, preds self.jit_val_step = jit(val_step) def val_loop(self, dataset_type: str = "validation", context: dict = {}): val_loss = 0 val_predicts = [] labels = [] metric_output = {} if dataset_type in ["validation", "val"]: dataloader = self.val_dataloader elif dataset_type == "train": dataloader = self.train_dataloader else: raise ValueError(f"dataset type {dataset_type} is not valid.") for batch_id, (feature, label) in enumerate(dataloader): loss, preds = self.jit_val_step((feature, label), self.state) val_predicts.append(preds) val_loss += loss.item() labels.append(label) val_loss /= len(dataloader) metric_output[self.loss_func_name] = val_loss val_predicts = jnp.concatenate(val_predicts, axis=0) labels = jnp.concatenate(labels, axis=0) metrics_conf: dict = self.train_params["metric_config"] for method in self.metrics: metric_output[method] = self.metrics[method](labels, val_predicts, **metrics_conf[method]) logger.info(f"Metrics on {dataset_type} set: {metric_output}")
8,308
39.531707
121
py
XFL
XFL-master/python/algorithm/framework/horizontal/linear_regression/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from algorithm.core.horizontal.template.agg_type import register_agg_type_for_label_trainer from common.utils.logger import logger from .common import Common from functools import partial class HorizontalLinearRegressionLabelTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] self.register_hook( place="after_train_loop", rank=1, func=partial(self.val_loop, "train"), desc="validation on trainset" ) register_agg_type_for_label_trainer(self, 'torch', agg_type) def train_loop(self): self.model.train() train_loss = 0 lossfunc = list(self.lossfunc.values())[0] optimizer = list(self.optimizer.values())[0] lr_scheduler = list(self.lr_scheduler.values())[0] if self.lr_scheduler.values() else None for batch, (feature, label) in enumerate(self.train_dataloader): pred = self.model(feature) loss = lossfunc(pred, label) optimizer.zero_grad() loss.backward() optimizer.step() train_loss += loss.item() train_loss /= len(self.train_dataloader) if lr_scheduler: lr_scheduler.step() self.context["train_loss"] = train_loss logger.info(f"Train loss: {train_loss}")
2,036
36.722222
98
py
XFL
XFL-master/python/algorithm/framework/horizontal/linear_regression/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from functools import partial from .common import Common from algorithm.core.horizontal.template.agg_type import register_agg_type_for_assist_trainer class HorizontalLinearRegressionAssistTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] register_agg_type_for_assist_trainer(self, 'torch', agg_type) self.register_hook(place="after_local_epoch", rank=1, func=partial(self._save_model, False), desc="save model ") self.register_hook(place="after_local_epoch", rank=2, func=partial(self.val_loop, "val"), desc="validation on valset") self.register_hook(place="after_global_epoch", rank=1, func=partial(self._save_model, True), desc="save final model") def train_loop(self): pass
1,538
42.971429
92
py
XFL
XFL-master/python/algorithm/framework/horizontal/linear_regression/common.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import numpy as np import torch import torch.nn as nn from torch.utils.data import DataLoader, TensorDataset from algorithm.core.data_io import CsvReader from algorithm.model.linear_regression import LinearRegression from common.utils.logger import logger from algorithm.core.horizontal.template.torch.base import BaseTrainer from common.utils.config_sync import ConfigSynchronizer from common.checker.x_types import All from common.evaluation.metrics import CommonMetrics class Common(BaseTrainer): def __init__(self, train_conf: dict): sync_rule = { "model_info": { "config": All() }, "train_info": { "interaction_params": All(), "train_params": { "global_epoch": All(), "aggregation": All(), "encryption": All(), "optimizer": All(), "lr_scheduler": All(), "lossfunc": All(), "metric": All(), "early_stopping": All() } } } train_conf = ConfigSynchronizer(train_conf).sync(sync_rule) super().__init__(train_conf) def _set_model(self) -> nn.Module: model_config = self.common_config.model_info.get("config") model = LinearRegression(input_dim=model_config["input_dim"], bias=model_config["bias"]) return model def _read_data(self, input_dataset): if len(input_dataset) == 0: return None conf = input_dataset[0] if conf["type"] == "csv": path = os.path.join(conf['path'], conf['name']) has_label = conf["has_label"] has_id = conf['has_id'] return CsvReader(path, has_id, has_label) else: return None def _set_train_dataloader(self): train_data = self._read_data(self.common_config.input_trainset) trainset = None train_dataloader = None if train_data: trainset = TensorDataset( torch.tensor(train_data.features(), dtype=torch.float32).to(self.device), torch.tensor(train_data.label(), dtype=torch.float32).unsqueeze(dim=-1).to(self.device) ) batch_size = self.common_config.train_params.get("train_batch_size") if trainset: train_dataloader = DataLoader(trainset, batch_size, shuffle=True) return train_dataloader def _set_val_dataloader(self): val_data = self._read_data(self.common_config.input_valset) valset = None val_dataloader = None if val_data: valset = TensorDataset( torch.tensor(val_data.features(), dtype=torch.float32).to(self.device), torch.tensor(val_data.label(), dtype=torch.float32).unsqueeze(dim=-1).to(self.device) ) batch_size = self.common_config.train_params.get("val_batch_size") if valset: val_dataloader = DataLoader(valset, batch_size, shuffle=True) return val_dataloader def val_loop(self, dataset_type: str = "val", context: dict = {}): self.model.eval() val_loss = 0 val_predicts = [] labels = [] lossfunc_name = list(self.lossfunc.keys())[0] lossfunc = list(self.lossfunc.values())[0] if dataset_type == "val": dataloader = self.val_dataloader elif dataset_type == "train": dataloader = self.train_dataloader else: raise ValueError(f"dataset type {dataset_type} is not valid.") for batch, (feature, label) in enumerate(dataloader): pred = self.model(feature) loss = lossfunc(pred, label) val_predicts.append(pred.detach().cpu().squeeze(-1).numpy()) val_loss += loss.item() labels.append(label.cpu().squeeze(-1).numpy()) val_loss /= len(dataloader) labels: np.ndarray = np.concatenate(labels, axis=0) val_predicts: np.ndarray = np.concatenate(val_predicts, axis=0) metrics_output = CommonMetrics._calc_metrics( metrics=self.metrics, labels=labels, val_predicts=val_predicts, lossfunc_name=lossfunc_name, loss=val_loss, dataset_type=dataset_type ) global_epoch = self.context["g_epoch"] if dataset_type == "val": local_epoch = None elif dataset_type == "train": local_epoch = self.context["l_epoch"] CommonMetrics.save_metric_csv( metrics_output=metrics_output, output_config=self.common_config.output, global_epoch=global_epoch, local_epoch=local_epoch, dataset_type=dataset_type, ) early_stop_flag = self.context["early_stop_flag"] if (self.common_config.save_frequency > 0) & \ (dataset_type == "val") & (self.earlystopping.patience > 0): early_stop_flag = self.earlystopping(metrics_output, global_epoch) if early_stop_flag: # find the saved epoch closest to the best epoch best_epoch = self.earlystopping.best_epoch closest_epoch = round(best_epoch / self.common_config.save_frequency) * \ self.common_config.save_frequency closest_epoch -= self.common_config.save_frequency \ if closest_epoch > global_epoch else 0 self.context["early_stop_flag"] = True self.context["early_stop_epoch"] = closest_epoch
6,316
35.94152
103
py
XFL
XFL-master/python/algorithm/framework/horizontal/kmeans/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import torch from torch import nn from sklearn.cluster import KMeans from algorithm.core.horizontal.template.agg_type import register_agg_type_for_label_trainer from common.utils.logger import logger from .common import Common from functools import partial class HorizontalKmeansLabelTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) self.register_hook( place="after_train_loop", rank=1, func=partial(self.val_loop, "train"), desc="validation on trainset" ) register_agg_type_for_label_trainer(self, "torch", "fedavg") # logger.info("Label trainer initialized") def train_loop(self): # load centroids centroids = self.model.state_dict()['centroids'].numpy() # train one iter of KMeans kmeans_model = KMeans( n_clusters=centroids.shape[0], init=centroids, n_init=1, max_iter=10 ) train_features, _ = self.train_dataloader.dataset.tensors train_features = train_features.numpy() kmeans_model.fit(train_features) logger.info(f"K-Means score: {kmeans_model.score(train_features)}") # write centroids model_state_dict = self.model.state_dict() model_state_dict['centroids'] = torch.tensor( kmeans_model.cluster_centers_) # self.model.load_state_dict(model_state_dict) self.model.centroids = nn.Parameter( torch.tensor(kmeans_model.cluster_centers_))
2,144
35.355932
91
py
XFL
XFL-master/python/algorithm/framework/horizontal/kmeans/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from functools import partial from common.utils.logger import logger from .common import Common from algorithm.core.horizontal.template.agg_type import register_agg_type_for_assist_trainer class HorizontalKmeansAssistTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) register_agg_type_for_assist_trainer(self, 'torch', "fedavg") self.register_hook(place="after_local_epoch", rank=1, func=partial(self._save_model, False), desc="save model ") self.register_hook(place="after_local_epoch", rank=2, func=partial(self.val_loop, "val"), desc="validation on valset") self.register_hook(place="after_global_epoch", rank=1, func=partial(self._save_model, True), desc="save final model") def train_loop(self): pass
1,491
41.628571
92
py
XFL
XFL-master/python/algorithm/framework/horizontal/kmeans/common.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import numpy as np import torch from torch import nn from torch.utils.data.dataset import TensorDataset from torch.utils.data.dataloader import DataLoader from sklearn.metrics import davies_bouldin_score from sklearn.cluster import KMeans from algorithm.core.data_io import CsvReader from algorithm.model.horizontal_k_means import HorizontalKMeans from common.utils.logger import logger from algorithm.core.horizontal.template.torch.base import BaseTrainer from common.utils.config_sync import ConfigSynchronizer from common.checker.x_types import All from common.evaluation.metrics import CommonMetrics class Common(BaseTrainer): def __init__(self, train_conf: dict) -> None: sync_rule = { "model_info": All(), "train_info": { "interaction_params": All(), "train_params": { "global_epoch": All(), "aggregation": All(), "encryption": All(), } } } train_conf = ConfigSynchronizer(train_conf).sync(sync_rule) super().__init__(train_conf) def _set_model(self) -> nn.Module: model_config = self.common_config.model_info.get("config") input_dim = model_config["input_dim"] num_clusters = model_config["num_clusters"] model = HorizontalKMeans( input_dim=input_dim, num_clusters=num_clusters) return model def _read_data(self, input_dataset): if len(input_dataset) == 0: return None conf = input_dataset[0] if conf["type"] == "csv": path = os.path.join(conf['path'], conf['name']) logger.info(f"Data path: {os.path.abspath(path)}") has_label = conf["has_label"] has_id = conf['has_id'] return CsvReader(path, has_id, has_label) else: return None def _set_train_dataloader(self): train_data = self._read_data(self.common_config.input_trainset) train_dataloader = None if train_data: train_dataset = TensorDataset( torch.Tensor(train_data.features()), torch.Tensor(train_data.label()) ) train_dataloader = DataLoader(train_dataset) return train_dataloader def _set_val_dataloader(self): val_data = self._read_data(self.common_config.input_valset) val_dataloader = None if val_data: val_dataset = TensorDataset( torch.Tensor(val_data.features()), torch.Tensor(val_data.label()) ) val_dataloader = DataLoader(val_dataset) return val_dataloader def val_loop(self, dataset_type: str = "val", context: dict = {}): if dataset_type == "val": dataloader = self.val_dataloader elif dataset_type == "train": dataloader = self.train_dataloader else: raise ValueError(f"dataset type {dataset_type} is not valid.") val_features, val_label = dataloader.dataset.tensors val_features = val_features.numpy() # val_bale = val_label.numpy() centroids = self.model.state_dict()['centroids'].numpy() kmeans = KMeans( n_clusters=centroids.shape[0], init=centroids, n_init=1, max_iter=1) kmeans.fit(val_features) kmeans.cluster_centers = centroids pred_labels = kmeans.predict(val_features) score = davies_bouldin_score(val_features, pred_labels) metrics_output = CommonMetrics._calc_metrics( metrics={}, labels=val_label, val_predicts=pred_labels, lossfunc_name="davies_bouldin_score", loss=score, dataset_type=dataset_type ) global_epoch = self.context["g_epoch"] if dataset_type == "val": local_epoch = None elif dataset_type == "train": local_epoch = self.context["l_epoch"] CommonMetrics.save_metric_csv( metrics_output=metrics_output, output_config=self.common_config.output, global_epoch=global_epoch, local_epoch=local_epoch, dataset_type=dataset_type, )
4,892
34.456522
80
py
XFL
XFL-master/python/algorithm/framework/horizontal/nbafl/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import torch from common.communication.gRPC.python.channel import DualChannel from service.fed_config import FedConfig from functools import partial from algorithm.core.horizontal.template.agg_type import \ register_agg_type_for_label_trainer from common.utils.logger import logger from .common import Common class HorizontalNbaflLabelTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) register_agg_type_for_label_trainer(self, "torch", "fedavg") self.sample_size_channel = DualChannel( name="sample_size_" + FedConfig.node_id, ids=[FedConfig.get_assist_trainer(), FedConfig.node_id] ) # initialize prev params self.prev_params = [ param.data.detach().clone() for param in self.model.parameters() ] # Update sample size self.register_hook( place="before_global_epoch", rank=1, func=self._update_sample_size, desc="Update local sample size" ) # Calculate update sigma self.register_hook( place="before_global_epoch", rank=2, func=self._calc_uplink_sigma, desc="Calculate uplink sigma" ) # Update prev param self.register_hook( place="after_local_epoch", rank=1, func=self._update_prev_param, desc="Update prev param" ) # Clip norm self.register_hook( place="after_local_epoch", rank=2, func=self._clip_params, desc="Clip param norms" ) # Add noise self.register_hook( place="after_local_epoch", rank=3, func=self._add_noise, desc="Add uplink noise" ) # Validation self.register_hook( place="after_train_loop", rank=1, func=partial(self.val_loop, "train"), desc="validation on trainset" ) def _update_prev_param(self, context): self.prev_params = [ param.data.detach().clone() for param in self.model.parameters() ] def _cal_regularization(self, p=2): reg = 0.0 for w_prev, w in zip(self.prev_params, self.model.parameters()): reg += torch.pow(torch.norm(w - w_prev, p), p) return self.mu * reg / p def _clip_params(self, context): for param in self.model.parameters(): norm_ratio = torch.maximum( torch.ones(param.shape), torch.abs(param.data) / self.common_config.train_params['C'] ) param.data = param.data / norm_ratio return def _calc_uplink_sigma(self, context): delta_S_u = 2 * self.common_config.train_params['C'] / \ len(self.train_dataloader.dataset) sigma_u = self.c * delta_S_u / self.epsilon logger.info("Uplink sigma: {}".format(sigma_u)) self.sigma_u = sigma_u return def train_loop(self): self.model.train() train_loss = 0 lossfunc = list(self.lossfunc.values())[0] optimizer = list(self.optimizer.values())[0] for batch_idx, (feature, label) in enumerate(self.train_dataloader): pred = self.model(feature) loss = lossfunc(pred, label) reg = self._cal_regularization() loss += reg optimizer.zero_grad() loss.backward() optimizer.step() train_loss += loss.item() train_loss /= len(self.train_dataloader) # retain current params self.prev_params = [ param.data.detach().clone() for param in self.model.parameters() ] return train_loss def _add_noise(self, context): for param in self.model.parameters(): param.data += torch.distributions.Normal( loc=0, scale=self.sigma_u).sample(param.size()).to(self.device) return def _update_sample_size(self, context): logger.info("trainset length: {}".format( len(self.train_dataloader.dataset))) self.sample_size_channel.send(len(self.train_dataloader.dataset)) return
4,761
33.014286
79
py
XFL
XFL-master/python/algorithm/framework/horizontal/nbafl/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import numpy as np import torch from common.communication.gRPC.python.channel import DualChannel from service.fed_config import FedConfig from functools import partial from common.utils.logger import logger from .common import Common from algorithm.core.horizontal.template.agg_type import \ register_agg_type_for_assist_trainer class HorizontalNbaflAssistTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) register_agg_type_for_assist_trainer(self, "torch", "fedavg") self.load_model() self.sample_size_channel = {} # Init size channel for party_id in FedConfig.get_label_trainer(): self.sample_size_channel[party_id] = DualChannel( name="sample_size_" + party_id, ids=[FedConfig.node_id, party_id] ) # Get sample size self.register_hook( place="before_global_epoch", rank=1, func=self._get_sample_size, desc="Get sample size" ) # Calculate downlink noise self.register_hook( place="before_global_epoch", rank=2, func=self._calc_downlink_sigma, desc="Calculate downlink noise" ) # Add noise self.register_hook( place="after_local_epoch", rank=1, func=self._add_noise, desc="Add downlink noise" ) # Validation self.register_hook( place="after_local_epoch", rank=2, func=partial(self.val_loop, "val"), desc="validation on valset" ) self.register_hook( place="after_global_epoch", rank=1, func=partial(self._save_model, True), desc="save final model" ) def _calc_downlink_sigma(self, context): logger.info("Calculating downlink sigma") if self.common_config.train_params['global_epoch'] > \ self.common_config.train_params['num_client'] * \ np.sqrt(self.common_config.train_params['num_client']): sigma_d = ( 2 * self.common_config.train_params['C'] * self.c * np.sqrt( self.common_config.train_params['global_epoch'] ** 2 - \ np.power(self.common_config.train_params['num_client'], 3)) / \ (self.min_sample_num * \ self.common_config.train_params['num_client'] * \ self.common_config.train_params['epsilon']) ) else: sigma_d = 0.0 logger.info("Downlink sigma: {}".format(sigma_d)) self.sigma_d = sigma_d return def _add_noise(self, context): if self.sigma_d > 0: noise_generator = torch.distributions.Normal( loc=0, scale=self.sigma_d) for param_data in self.model.parameters(): param_data.data += noise_generator.sample(param_data.size()) return def _get_sample_size(self, context): sample_nums = [] for party_id in FedConfig.get_label_trainer(): single_sample_size = self.sample_size_channel[party_id].recv() sample_nums.append(single_sample_size) sample_num_array = np.array(sample_nums) logger.info("Sample num array: {}".format(sample_num_array)) self.min_sample_num = np.min(sample_num_array) return def train_loop(self): pass
4,053
36.192661
87
py
XFL
XFL-master/python/algorithm/framework/horizontal/nbafl/common.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import numpy as np import torch import torch.nn as nn from torch.utils.data import DataLoader, TensorDataset from algorithm.core.data_io import CsvReader from algorithm.core.horizontal.template.torch.base import BaseTrainer from common.checker.x_types import All from common.evaluation.metrics import CommonMetrics from common.utils.config_sync import ConfigSynchronizer from common.utils.logger import logger class Common(BaseTrainer): def __init__(self, train_conf: dict): sync_rule = { "model_info": { "config": All() }, "train_info": { "interaction_params": All(), "train_params": { "global_epoch": All(), "mu": All(), "epsilon": All(), "delta": All(), "C": All(), "aggregation": All(), "encryption": All(), "optimizer": All(), "lossfunc": All(), "metric": All(), "early_stopping": All() } } } train_conf = ConfigSynchronizer(train_conf).sync(sync_rule) super().__init__(train_conf) self.mu = self.common_config.train_params['mu'] self.delta = self.common_config.train_params['delta'] self.c = np.sqrt(2 * np.log(1.25 / self.delta)) self.epsilon = self.common_config.train_params['epsilon'] def _set_model(self): logger.info("Model info: {}".format(self.common_config.model_info)) model_config = self.common_config.model_info["config"] assert len(model_config['layer_dim']) == len( model_config['activation']), "Hidden layer nums must match activation nums" layer_dims = [model_config['input_dim']] + model_config['layer_dim'] layer_act = model_config['activation'] module_list = [] for input_dim, output_dim, activation_str in zip(layer_dims, layer_dims[1:], layer_act): module_list.append( nn.Linear(input_dim, output_dim, bias=model_config['bias'])) activation = getattr(nn, activation_str)() module_list.append(activation) model = nn.Sequential(*module_list) return model def load_model(self): self._load_model({}) def _read_data(self, input_dataset): if len(input_dataset) == 0: return None conf = input_dataset[0] if conf["type"] == "csv": path = os.path.join(conf['path'], conf['name']) has_label = conf["has_label"] has_id = conf['has_id'] return CsvReader(path, has_id, has_label) else: return None def _set_train_dataloader(self): train_data = self._read_data(self.common_config.input_trainset) trainset = None train_dataloader = None if train_data: trainset = TensorDataset( torch.tensor(train_data.features(), dtype=torch.float32).to(self.device), torch.tensor(train_data.label(), dtype=torch.float32).unsqueeze(dim=-1).to(self.device) ) batch_size = self.common_config.train_params.get("train_batch_size") if trainset: train_dataloader = DataLoader(trainset, batch_size, shuffle=True) return train_dataloader def _set_val_dataloader(self): val_data = self._read_data(self.common_config.input_valset) valset = None val_dataloader = None if val_data: valset = TensorDataset( torch.tensor(val_data.features(), dtype=torch.float32).to(self.device), torch.tensor(val_data.label(), dtype=torch.float32).unsqueeze(dim=-1).to(self.device) ) batch_size = self.common_config.train_params.get("val_batch_size") if valset: val_dataloader = DataLoader(valset, batch_size, shuffle=True) return val_dataloader def val_loop(self, dataset_type: str = "val", context: dict = {}): self.model.eval() val_loss = 0 val_predicts = [] labels = [] lossfunc_name = list(self.lossfunc.keys())[0] lossfunc = list(self.lossfunc.values())[0] if dataset_type == "val": dataloader = self.val_dataloader elif dataset_type == "train": dataloader = self.train_dataloader else: raise ValueError(f"dataset type {dataset_type} is not valid.") with torch.no_grad(): for batch_idx, (feature, label) in enumerate(dataloader): pred = self.model(feature) loss = lossfunc(pred, label) val_predicts.append(pred.detach().cpu().squeeze(-1).numpy()) val_loss += loss.item() labels.append(label.cpu().squeeze(-1).numpy()) val_loss /= len(dataloader) labels: np.ndarray = np.concatenate(labels, axis=0) val_predicts: np.ndarray = np.concatenate(val_predicts, axis=0) if len(val_predicts.shape) == 1: val_predicts = np.array(val_predicts > 0.5, dtype=np.int32) elif len(val_predicts.shape) == 2: val_predicts = val_predicts.argmax(axis=-1) metrics_output = CommonMetrics._calc_metrics( metrics=self.metrics, labels=labels, val_predicts=val_predicts, lossfunc_name=lossfunc_name, loss=val_loss, dataset_type=dataset_type ) global_epoch = self.context["g_epoch"] if dataset_type == "val": local_epoch = None elif dataset_type == "train": local_epoch = self.context["l_epoch"] CommonMetrics.save_metric_csv( metrics_output=metrics_output, output_config=self.common_config.output, global_epoch=global_epoch, local_epoch=local_epoch, dataset_type=dataset_type, ) early_stop_flag = self.context["early_stop_flag"] if (self.common_config.save_frequency > 0) & \ (dataset_type == "val") & (self.earlystopping.patience > 0): early_stop_flag = self.earlystopping(metrics_output, global_epoch) if early_stop_flag: # find the saved epoch closest to the best epoch best_epoch = self.earlystopping.best_epoch closest_epoch = round(best_epoch / self.common_config.save_frequency) * \ self.common_config.save_frequency closest_epoch -= self.common_config.save_frequency \ if closest_epoch > global_epoch else 0 self.context["early_stop_flag"] = True self.context["early_stop_epoch"] = closest_epoch
7,492
37.229592
103
py
XFL
XFL-master/python/algorithm/framework/horizontal/nbafl/__init__.py
0
0
0
py
XFL
XFL-master/python/algorithm/framework/horizontal/chatglm/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from transformers import Trainer from service.fed_config import FedConfig from algorithm.framework.horizontal.chatglm.common import Common from algorithm.framework.horizontal.chatglm.callback import LabelTrainerCallback class HorizontalChatglmLabelTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) def fit(self): agg_steps = self.common_config.aggregation.get("agg_steps") or self.common_config.train_params["trainer"]["save_steps"] sec_conf = self.common_config.train_params["encryption"] (peft_type, peft_config_dict), = self.common_config.train_params["peft"].items() my_callback = LabelTrainerCallback(agg_steps, sec_conf, root_id=FedConfig.get_assist_trainer(), leaf_ids=FedConfig.get_label_trainer(), init_params=not self.load_from_pretrained, peft_type=peft_type) trainer = Trainer( model=self.model, args=self.training_args, train_dataset=self.train_dataset, eval_dataset=self.val_dataset, tokenizer=self.tokenizer, data_collator=self.data_collator, callbacks=[my_callback], ) trainer.train()
2,039
40.632653
127
py
XFL
XFL-master/python/algorithm/framework/horizontal/chatglm/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from pathlib import Path from transformers import Trainer from algorithm.core.horizontal.aggregation.api import get_aggregation_root_inst from service.fed_config import FedConfig from common.utils.logger import logger from algorithm.framework.horizontal.chatglm.common import Common from algorithm.framework.horizontal.chatglm.callback import ( AssistTrainerCallback, get_adapter_state_dict, set_adapter_state_dict ) class HorizontalChatglmAssistTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) def fit(self): agg_steps = self.common_config.aggregation.get("agg_steps") or self.common_config.train_params["trainer"]["save_steps"] sec_conf = self.common_config.train_params["encryption"] (peft_type, peft_config_dict), = self.common_config.train_params["peft"].items() if len(self.common_config.input_trainset) != 0: my_callback = AssistTrainerCallback(agg_steps, sec_conf, root_id=FedConfig.get_assist_trainer(), leaf_ids=FedConfig.get_label_trainer(), init_params=not self.load_from_pretrained, peft_type=peft_type) trainer = Trainer( model=self.model, args=self.training_args, train_dataset=self.train_dataset, eval_dataset=self.val_dataset, tokenizer=self.tokenizer, data_collator=self.data_collator, callbacks=[my_callback], ) trainer.train() else: agg_inst = get_aggregation_root_inst(sec_conf, root_id=FedConfig.get_assist_trainer(), leaf_ids=FedConfig.get_label_trainer()) self.agg_steps_list = [] i = agg_steps while i < 1: self.agg_steps_list.append(round(i, 4)) i += agg_steps self.agg_steps_list.append(1) adapters_weights = get_adapter_state_dict(self.model, peft_type) agg_inst.broadcast(adapters_weights) for i in range(len(self.agg_steps_list)): logger.info(f"gather and agg, global_step={self.agg_steps_list[i]}") new_adapters_weights = agg_inst.aggregate() logger.info(f"broadcast, global_step={self.agg_steps_list[i]}") agg_inst.broadcast(new_adapters_weights) if self.training_args.output_dir and self.training_args.save_strategy != 'no': save_dir = Path(self.common_config.output["path"]) / f"checkpoint-{str(self.agg_steps_list[i])}" if peft_type == "PREFIX_TUNING": self.model.save_pretrained(save_directory=save_dir, state_dict=new_adapters_weights) else: set_adapter_state_dict(self.model, peft_type, new_adapters_weights) self.model.save_pretrained(save_directory=save_dir)
3,982
45.313953
127
py
XFL
XFL-master/python/algorithm/framework/horizontal/chatglm/common.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os from pathlib import Path import torch from peft import ( get_peft_model, PeftModel, PEFT_TYPE_TO_CONFIG_MAPPING ) from transformers import ( AutoTokenizer, AutoModel, AutoConfig, DataCollatorForSeq2Seq, TrainingArguments ) from algorithm.core.data_io import QADataset from service.fed_config import FedConfig from common.utils.config_parser import CommonConfigParser from common.utils.logger import logger from common.checker.x_types import All from common.utils.config_sync import ConfigSynchronizer def alternateSVDLinear(): import torch.nn.functional as F from peft.tuners.adalora import transpose def forward(self, x: torch.Tensor): if self.active_adapter not in self.lora_A.keys(): return F.linear(x, transpose(self.weight, self.fan_in_fan_out), bias=self.bias) if self.disable_adapters: if self.r[self.active_adapter] > 0 and self.merged: self.unmerge() result = F.linear(x, transpose(self.weight, self.fan_in_fan_out), bias=self.bias) elif self.r[self.active_adapter] > 0 and not self.merged: result = F.linear(x, transpose(self.weight, self.fan_in_fan_out), bias=self.bias) result += ( ( self.lora_dropout[self.active_adapter](x.float()) # to float() @ (self.lora_A[self.active_adapter] * self.lora_E[self.active_adapter]).T @ self.lora_B[self.active_adapter].T ) * self.scaling[self.active_adapter] / (self.ranknum[self.active_adapter] + 1e-5) ) else: result = F.linear(x, transpose(self.weight, self.fan_in_fan_out), bias=self.bias) return result from peft.tuners.adalora import SVDLinear SVDLinear.forward = forward class Common: def __init__(self, train_conf: dict): if FedConfig.get_assist_trainer(): sync_rule = { "train_info": { "train_params": { "aggregation": All(), "encryption": All(), "peft": All(), "trainer": { "learning_rate": All(), "weight_decay": All(), "adam_beta1": All(), "adam_beta2": All(), "adam_epsilon": All(), "max_grad_norm": All(), "max_steps": All(), "num_train_epochs": All(), "seed": All() }, "dataset": All() } } } train_conf = ConfigSynchronizer(train_conf).sync(sync_rule) root_path = Path(__file__).parents[4] path = train_conf.get('input', {}).get("pretrained_model", {}).get("path") if path and not os.path.isabs(path): train_conf["input"]["pretrained_model"]['path'] = os.path.abspath(os.path.join(root_path, path)) path = train_conf.get('input', {}).get("adapter_model", {}).get("path") if path and not os.path.isabs(path): train_conf["input"]["adapter_model"]['path'] = os.path.abspath(os.path.join(root_path, path)) trainset_conf = train_conf.get('input', {}).get("trainset") if trainset_conf: path = trainset_conf[0].get("path") if path and not os.path.isabs(path): train_conf["input"]["trainset"][0]['path'] = os.path.abspath(os.path.join(root_path, path)) self.common_config = CommonConfigParser(train_conf) # for adalora alternateSVDLinear() # CPU if self.common_config.train_params["trainer"].get("no_cuda"): os.environ["CUDA_VISIBLE_DEVICES"] = "-1" pretrained_model_conf = self.common_config.input.get("pretrained_model", {}) path = pretrained_model_conf.get("path") model_name_or_path = path self.tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True) # Callback logger.info(self.tokenizer) self.load_from_pretrained = False for name in os.listdir(model_name_or_path): if 'pytorch_model' in name: self.load_from_pretrained = True break (peft_type, peft_config_dict), = self.common_config.train_params["peft"].items() if self.load_from_pretrained: if peft_type == "PREFIX_TUNING": config = AutoConfig.from_pretrained(model_name_or_path, trust_remote_code=True) config.pre_seq_len = self.common_config.train_params["peft"][peft_type]["pre_seq_len"] config.prefix_projection = self.common_config.train_params["peft"][peft_type]["prefix_projection"] model = AutoModel.from_pretrained(model_name_or_path, config=config, trust_remote_code=True) else: model = AutoModel.from_pretrained(model_name_or_path, trust_remote_code=True, device_map="auto") else: if peft_type == "PREFIX_TUNING": config = AutoConfig.from_pretrained(model_name_or_path, trust_remote_code=True) config.pre_seq_len = self.common_config.train_params["peft"][peft_type]["pre_seq_len"] config.prefix_projection = self.common_config.train_params["peft"][peft_type]["prefix_projection"] model = AutoModel.from_config(config, trust_remote_code=True) else: # if peft_type == "LORA": config = AutoConfig.from_pretrained(model_name_or_path, trust_remote_code=True) model = AutoModel.from_config(config, trust_remote_code=True) logger.warning("No pretrained model founded, load from config") if self.common_config.train_params["trainer"].get("no_cuda"): model = model.float() else: model = model.half() if peft_type == "PREFIX_TUNING": self.model = model else: peft_config = PEFT_TYPE_TO_CONFIG_MAPPING[peft_type](inference_mode=False, **peft_config_dict) # model = prepare_model_for_int8_training(model) self.model = get_peft_model(model, peft_config) if self.common_config.input.get("adapter_model", {}): adapter_path = self.common_config.input.get("adapter_model")["path"] if peft_type == "PREFIX_TUNING": # P-tuning v2 prefix_state_dict = torch.load(os.path.join(adapter_path, "pytorch_model.bin")) new_prefix_state_dict = {} for k, v in prefix_state_dict.items(): if k.startswith("transformer.prefix_encoder."): new_prefix_state_dict[k[len("transformer.prefix_encoder."):]] = v self.model.transformer.prefix_encoder.load_state_dict(new_prefix_state_dict) else: self.model = PeftModel.from_pretrained(self.model, adapter_path, adapter_name="default", is_trainable=True) logger.info("Load adapter model.") if peft_type == "PREFIX_TUNING": if not self.common_config.train_params["trainer"].get("no_cuda"): self.model.transformer.prefix_encoder.float() logger.info(self.model) self.train_dataset, self.val_dataset = self._set_dataset() self.data_collator = self._set_data_collator() trainer_conf = self.common_config.train_params["trainer"] trainer_conf["max_steps"] = -1 # trainer_conf["save_strategy"] = 'steps' trainer_conf["save_steps"] = self.common_config.train_params["aggregation"]["agg_steps"] trainer_conf["output_dir"] = self.common_config.output_dir self.training_args = TrainingArguments(**trainer_conf) self.trainer_conf = trainer_conf if peft_type == "PREFIX_TUNING": self.training_args.local_rank = -1 def _set_data_collator(self): data_collator = DataCollatorForSeq2Seq( self.tokenizer, model=None, label_pad_token_id=-100, pad_to_multiple_of=None, padding=True ) return data_collator def _set_dataset(self): dataset_conf = self.common_config.train_params["dataset"] train_dataset, val_dataset = None, None if self.common_config.input_trainset: file_name_or_path = os.path.join( self.common_config.input_trainset[0].get("path") ) train_dataset = QADataset( file_name_or_path=file_name_or_path, tokenizer=self.tokenizer, max_src_length=dataset_conf["max_src_length"], max_dst_length=dataset_conf["max_dst_length"], prompt_pattern=dataset_conf["prompt_pattern"], key_query=dataset_conf.get("key_query", "input"), key_answer=dataset_conf.get("key_answer", "output") ) if self.common_config.input_valset: file_name_or_path = os.path.join( self.common_config.input_valset[0].get("path") ) val_dataset = QADataset( file_name_or_path=file_name_or_path, tokenizer=self.tokenizer, max_src_length=dataset_conf["max_src_length"], max_dst_length=dataset_conf["max_dst_length"], prompt_pattern=dataset_conf["prompt_pattern"], key_query=dataset_conf.get("key_query", "input"), key_answer=dataset_conf.get("key_answer", "output") ) return train_dataset, val_dataset
10,798
42.898374
114
py
XFL
XFL-master/python/algorithm/framework/horizontal/chatglm/callback.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math import inspect from pathlib import Path from typing import Union from copy import deepcopy import torch import torch.nn as nn import transformers from accelerate import ( dispatch_model, infer_auto_device_map ) from accelerate.utils import get_balanced_memory from accelerate.hooks import ( AlignDevicesHook, add_hook_to_module, remove_hook_from_submodules ) from transformers import ( TrainerCallback, TrainingArguments, TrainerState, TrainerControl ) from peft import PeftModel from peft.utils import ( get_peft_model_state_dict, set_peft_model_state_dict, PromptLearningConfig ) from algorithm.core.horizontal.aggregation.api import ( get_aggregation_root_inst, get_aggregation_leaf_inst ) from common.utils.logger import logger from service.fed_config import FedConfig # def is_nan_exists(state_dict): # flag = False # for k, v in state_dict.items(): # if torch.isnan(v).any(): # flag = True # logger.warning(f"Parameter {k} contains nan") # break # return flag class AssistTrainerCallback(TrainerCallback): def __init__(self, agg_steps: int, sec_conf: dict, root_id: str, leaf_ids: list[str], init_params: bool = False, peft_type: str = "LORA"): super().__init__() self.agg_steps = agg_steps self.agg_steps_list = [] assert 0 < agg_steps <= 1 self.agg_inst = get_aggregation_root_inst(sec_conf, root_id, leaf_ids) self.init_params = init_params self.peft_type = peft_type # self.latest_adapters_weights = None def on_train_begin(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, model: Union[transformers.PreTrainedModel, torch.nn.Module], **kwargs): if self.init_params: for m in model.modules(): if isinstance(m, nn.Conv2d): torch.nn.init.xavier_normal_(m.weight.data) if m.bias is not None: m.bias.data.zero_() elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.Linear): torch.nn.init.normal_(m.weight.data, 0, 0.01) if m.bias is not None: m.bias.data.zero_() elif isinstance(m, nn.Embedding): torch.nn.init.uniform_(m.weight.data) args.logging_steps = state.max_steps + 1 args.save_steps = state.max_steps + 1 adapters_weights = get_adapter_state_dict(model, self.peft_type) self.agg_inst.broadcast(adapters_weights) # self.latest_adapters_weights = adapters_weights # def on_step_begin(self, # args: TrainingArguments, # state: TrainerState, # control: TrainerControl, # model: Union[transformers.PreTrainedModel, torch.nn.Module], # **kwargs): # for k, v in model.state_dict().items(): # if torch.isnan(v).any(): # # set nan to 0 # v[v != v] = 0 # logger.warning(f"Parameter {k} contains nan, replace nan to 0") # control.should_log = False def on_step_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, model: Union[transformers.PreTrainedModel, torch.nn.Module], tokenizer, **kwargs): # Trainer saves model after check on_step_end if not self.agg_steps_list: i = self.agg_steps while i < 1: self.agg_steps_list.append(round(i, 4)) i += self.agg_steps self.agg_steps_list.append(1) self.steps_list = [math.ceil(i * state.max_steps) for i in self.agg_steps_list] assert len(self.steps_list) == len(set(self.steps_list)) logger.info(f"Aggergate model by steps: {self.agg_steps_list}") if state.global_step in self.steps_list: idx = self.steps_list.index(state.global_step) if idx == 0: factor = 1 else: factor = 1 adapters_weights = get_adapter_state_dict(model, self.peft_type) logger.info( f"gather and aggregating..., global_step={state.global_step}") new_adapters_weights = self.agg_inst.aggregate( parameters=adapters_weights, parameters_weight=factor) set_adapter_state_dict(model, self.peft_type, new_adapters_weights) logger.info(f"broadcasting..., global_step={state.global_step}") self.agg_inst.broadcast(new_adapters_weights) if args.output_dir and args.save_strategy != 'no': if self.peft_type != "PREFIX_TUNING": model.save_pretrained(save_directory=Path(args.output_dir) / f"checkpoint-{str(self.agg_steps_list[idx])}") else: model.save_pretrained(save_directory=Path(args.output_dir) / f"checkpoint-{str(self.agg_steps_list[idx])}", state_dict=get_adapter_state_dict(model, self.peft_type)) control.should_log = True class LabelTrainerCallback(TrainerCallback): def __init__(self, agg_steps: Union[float, int], sec_conf: dict, root_id: str, leaf_ids: list[str], init_params: bool = False, peft_type: str = "LORA"): super().__init__() self.agg_steps = agg_steps self.agg_steps_list = [] assert 0 < agg_steps <= 1 self.is_standalone = False if FedConfig.get_assist_trainer() else True if not self.is_standalone: self.agg_inst = get_aggregation_leaf_inst(sec_conf, root_id, leaf_ids) self.init_params = init_params self.peft_type = peft_type def on_train_begin(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, model: Union[transformers.PreTrainedModel, torch.nn.Module], train_dataloader, **kwargs): if self.init_params: for m in model.modules(): if isinstance(m, nn.Conv2d): torch.nn.init.xavier_normal_(m.weight.data) if m.bias is not None: m.bias.data.zero_() elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.Linear): torch.nn.init.normal_(m.weight.data, 0, 0.01) if m.bias is not None: m.bias.data.zero_() elif isinstance(m, nn.Embedding): torch.nn.init.uniform_(m.weight.data) args.logging_steps = state.max_steps + 1 args.save_steps = state.max_steps + 1 if not self.is_standalone: new_adapters_weights = self.agg_inst.download() set_adapter_state_dict(model, self.peft_type, new_adapters_weights) # def on_step_begin(self, # args: TrainingArguments, # state: TrainerState, # control: TrainerControl, # model: Union[transformers.PreTrainedModel, torch.nn.Module], # **kwargs): # for k, v in model.state_dict().items(): # if torch.isnan(v).any(): # # set nan to 0 # v[v != v] = 0 # logger.warning(f"Parameter {k} contains nan, replace nan to 0") # control.should_log = False def on_step_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, model: Union[transformers.PreTrainedModel, torch.nn.Module], **kwargs): # if is_nan_exists(model.state_dict()): # logger.warning(f"Nan exists!") # Trainer saves model after check on_step_end if not self.agg_steps_list: i = self.agg_steps while i < 1: self.agg_steps_list.append(round(i, 4)) i += self.agg_steps self.agg_steps_list.append(1) self.steps_list = [math.ceil(i * state.max_steps) for i in self.agg_steps_list] if len(self.steps_list) != len(set(self.steps_list)): raise ValueError(f"agg_steps is too small, try a larger one.") logger.info(f"Aggergate model by steps: {self.agg_steps_list}") if state.global_step in self.steps_list: idx = self.steps_list.index(state.global_step) if not self.is_standalone: if idx == 0: factor = 1 # factor = self.agg_steps_list[0] * \ # args.gradient_accumulation_steps * args.train_batch_size else: factor = 1 # factor = (self.agg_steps_list[idx] - self.agg_steps_list[idx-1]) * \ # args.gradient_accumulation_steps * args.train_batch_size adapters_weights = get_adapter_state_dict(model, self.peft_type) logger.info(f"uploading..., global_step={state.global_step}") self.agg_inst.upload(adapters_weights, factor) logger.info(f"downloading..., global_step={state.global_step}") new_adapters_weights = self.agg_inst.download() set_adapter_state_dict(model, self.peft_type, new_adapters_weights) if args.output_dir and args.save_strategy != 'no': if self.peft_type != "PREFIX_TUNING": model.save_pretrained(save_directory=Path(args.output_dir) / f"checkpoint-{str(self.agg_steps_list[idx])}") else: model.save_pretrained(save_directory=Path(args.output_dir) / f"checkpoint-{str(self.agg_steps_list[idx])}", state_dict=get_adapter_state_dict(model, self.peft_type)) control.should_log = True def get_adapter_state_dict(model: PeftModel, peft_type: str, **kwargs): if peft_type == "PREFIX_TUNING": state_dict = model.state_dict() adapters_weights = {} for k, v in model.named_parameters(): if v.requires_grad: adapters_weights[k] = deepcopy(state_dict[k]).to('cpu') else: # if peft_type == 'LORA': adapter_name = model.active_adapter adapters_weights = get_peft_model_state_dict( model, state_dict=kwargs.get("state_dict", None), adapter_name=adapter_name ) for k, v in adapters_weights.items(): adapters_weights[k] = deepcopy(v).to('cpu') return adapters_weights def set_adapter_state_dict(model: PeftModel, peft_type: str, adapters_weights: dict, **kwargs): if peft_type == "PREFIX_TUNING": state_dict = model.state_dict() state_dict.update(adapters_weights) model.load_state_dict(state_dict) else: adapter_name = model.active_adapter # load the weights into the model set_peft_model_state_dict(model, adapters_weights, adapter_name=adapter_name) if ( (getattr(model, "hf_device_map", None) is not None) and (len(set(model.hf_device_map.values()).intersection({"cpu", "disk"})) > 0) and len(model.peft_config) == 1 ): device_map = kwargs.get("device_map", "auto") max_memory = kwargs.get("max_memory", None) offload_dir = kwargs.get("offload_folder", None) offload_index = kwargs.get("offload_index", None) dispatch_model_kwargs = {} # Safety checker for previous `accelerate` versions # `offload_index` was introduced in https://github.com/huggingface/accelerate/pull/873/ if "offload_index" in inspect.signature(dispatch_model).parameters: dispatch_model_kwargs["offload_index"] = offload_index no_split_module_classes = model._no_split_modules if device_map != "sequential": max_memory = get_balanced_memory( model, max_memory=max_memory, no_split_module_classes=no_split_module_classes, low_zero=(device_map == "balanced_low_0"), ) if isinstance(device_map, str): device_map = infer_auto_device_map( model, max_memory=max_memory, no_split_module_classes=no_split_module_classes ) dispatch_model( model, device_map=device_map, offload_dir=offload_dir, **dispatch_model_kwargs, ) hook = AlignDevicesHook(io_same_device=True) if isinstance(model.peft_config[adapter_name], PromptLearningConfig): remove_hook_from_submodules(model.prompt_encoder) add_hook_to_module(model.get_base_model(), hook) # Set model in evaluation mode to deactivate Dropout modules by default # model.eval()
14,501
41.527859
127
py
XFL
XFL-master/python/algorithm/framework/horizontal/bert/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from algorithm.core.horizontal.template.agg_type import \ register_agg_type_for_label_trainer from common.utils.logger import logger from .common import Common from functools import partial class HorizontalBertLabelTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] self.register_hook( place="after_train_loop", rank=1, func=partial(self.val_loop, "train"), desc="validation on trainset" ) register_agg_type_for_label_trainer(self, 'torch', agg_type) def train_loop(self): self.model.train() train_loss = 0 lossfunc = list(self.lossfunc.values())[0] optimizer = list(self.optimizer.values())[0] lr_scheduler = list(self.lr_scheduler.values())[0] if self.lr_scheduler.values() else None for batch_id, (input_ids, token_type_ids, attention_masks, labels) in enumerate(self.train_dataloader): _,_,pred = self.model(input_ids, token_type_ids, attention_masks, labels) loss = lossfunc(pred, labels) optimizer.zero_grad() loss.backward() optimizer.step() train_loss += loss.item() train_loss /= len(self.train_dataloader) if lr_scheduler: lr_scheduler.step() self.context["train_loss"] = train_loss logger.info(f"Train loss: {train_loss}")
2,106
38.754717
111
py
XFL
XFL-master/python/algorithm/framework/horizontal/bert/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from functools import partial from .common import Common from algorithm.core.horizontal.template.agg_type import register_agg_type_for_assist_trainer class HorizontalBertAssistTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] register_agg_type_for_assist_trainer(self, 'torch', agg_type) self.register_hook(place="after_local_epoch", rank=1, func=partial(self._save_model, False), desc="save model ") self.register_hook(place="after_local_epoch", rank=2, func=partial(self.val_loop, "val"), desc="validation on valset") self.register_hook(place="after_global_epoch", rank=1, func=partial(self._save_model, True), desc="save final model") def train_loop(self): pass
1,529
42.714286
92
py
XFL
XFL-master/python/algorithm/framework/horizontal/bert/common.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import torch import pandas as pd import numpy as np from transformers import BertTokenizer import torch.nn as nn from torch.utils.data import DataLoader, TensorDataset from python.algorithm.model.bert import BertForSst2Torch from common.utils.logger import logger from algorithm.core.horizontal.template.torch.base import BaseTrainer from common.utils.config_sync import ConfigSynchronizer from common.checker.x_types import All from common.evaluation.metrics import CommonMetrics class Common(BaseTrainer): def __init__(self, train_conf: dict) -> None: sync_rule = { "model_info": All(), "train_info": { "interaction_params": All(), "train_params": { "global_epoch": All(), "aggregation": All(), "encryption": All(), "optimizer": All(), "lr_scheduler": All(), "lossfunc": All(), "metric": All(), "early_stopping": All() } } } train_conf = ConfigSynchronizer(train_conf).sync(sync_rule) super().__init__(train_conf) def _set_model(self) -> nn.Module: model_config = self.common_config.model_info.get("config") model = BertForSst2Torch(**model_config) return model def _read_data(self, input_dataset): if len(input_dataset) == 0: return None conf = input_dataset[0] path = os.path.join(conf['path'], conf['name']) raw_data = pd.read_csv(path, sep='\t') data = raw_data["sentence"].values, raw_data["label"].values return data def _encode_examples(self, data, tokenizer, max_length=512): input_ids, token_type_ids, attention_masks, labels = [], [], [], [] for feature, label in zip(*data): bert_input = tokenizer.encode_plus(feature, add_special_tokens=True, max_length=max_length, padding='max_length', return_token_type_ids=True, return_attention_mask=True) input_ids.append(bert_input['input_ids']) token_type_ids.append(bert_input['token_type_ids']) attention_masks.append(bert_input['attention_mask']) labels.append(label) return TensorDataset(torch.tensor(input_ids), torch.tensor(token_type_ids), torch.tensor(attention_masks), torch.tensor(labels)) def _set_train_dataloader(self): train_dataset = None train_dataloader = None tokenizer = BertTokenizer.from_pretrained("bert-base-uncased") train_data = self._read_data(self.common_config.input_trainset) if train_data is not None: train_dataset = self._encode_examples(train_data, tokenizer) batch_size = self.common_config.train_params.get( "train_batch_size") train_dataloader = DataLoader( train_dataset, batch_size, shuffle=True) return train_dataloader def _set_val_dataloader(self): val_dataset = None val_dataloader = None tokenizer = BertTokenizer.from_pretrained("bert-base-uncased") val_data = self._read_data(self.common_config.input_valset) if val_data is not None: val_dataset = self._encode_examples(val_data, tokenizer) batch_size = self.common_config.train_params.get("val_batch_size") val_dataloader = DataLoader(val_dataset, batch_size, shuffle=False) return val_dataloader def val_loop(self, dataset_type: str = "val", context: dict = {}): self.model.eval() val_loss = 0 val_predicts = [] labels = [] lossfunc_name = list(self.lossfunc.keys())[0] lossfunc = list(self.lossfunc.values())[0] if dataset_type == "val": dataloader = self.val_dataloader elif dataset_type == "train": dataloader = self.train_dataloader else: raise ValueError(f"dataset type {dataset_type} is not valid.") for batch_id, (input_ids, token_type_ids, attention_masks, label) in enumerate(dataloader): _, _, pred = self.model( input_ids, token_type_ids, attention_masks, label) loss = lossfunc(pred, label) val_predicts.append(pred.detach().cpu().squeeze(-1).numpy()) val_loss += loss.item() labels.append(label.cpu().squeeze(-1).numpy()) val_loss /= len(dataloader) labels: np.ndarray = np.concatenate(labels, axis=0) val_predicts: np.ndarray = np.concatenate(val_predicts, axis=0) if len(val_predicts.shape) == 1: val_predicts = np.array(val_predicts > 0.5, dtype=np.int32) elif len(val_predicts.shape) == 2: val_predicts = val_predicts.argmax(axis=-1) metrics_output = CommonMetrics._calc_metrics( metrics=self.metrics, labels=labels, val_predicts=val_predicts, lossfunc_name=lossfunc_name, loss=val_loss, dataset_type=dataset_type ) global_epoch = self.context["g_epoch"] if dataset_type == "val": local_epoch = None elif dataset_type == "train": local_epoch = self.context["l_epoch"] CommonMetrics.save_metric_csv( metrics_output=metrics_output, output_config=self.common_config.output, global_epoch=global_epoch, local_epoch=local_epoch, dataset_type=dataset_type, ) early_stop_flag = self.context["early_stop_flag"] if (self.common_config.save_frequency > 0) & \ (dataset_type == "val") & (self.earlystopping.patience > 0): early_stop_flag = self.earlystopping(metrics_output, global_epoch) if early_stop_flag: # find the saved epoch closest to the best epoch best_epoch = self.earlystopping.best_epoch closest_epoch = round(best_epoch / self.common_config.save_frequency) * \ self.common_config.save_frequency closest_epoch -= self.common_config.save_frequency \ if closest_epoch > global_epoch else 0 self.context["early_stop_flag"] = True self.context["early_stop_epoch"] = closest_epoch
7,291
39.966292
99
py
XFL
XFL-master/python/algorithm/framework/horizontal/resnet_paddle/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from algorithm.core.horizontal.template.paddle.fedavg.label_trainer import FedAvgLabelTrainer from common.utils.logger import logger import numpy as np from .common import Common class HorizontalResnetPaddleLabelTrainer(Common, FedAvgLabelTrainer): def __init__(self, train_conf: dict): FedAvgLabelTrainer.__init__(self, train_conf) def train_loop(self): self.model.train() train_losses = [] loss_func = list(self.loss_func.values())[0] lr_scheduler = list(self.lr_scheduler.values())[0] if self.lr_scheduler.values() else None optimizer = list(self.optimizer.values())[0] if lr_scheduler: optimizer for batch, (feature, label) in enumerate(self.train_dataloader): pred = self.model(feature) loss = loss_func(pred, label) optimizer.clear_grad() loss.backward() optimizer.step() train_losses.append(loss.numpy()) train_loss = np.mean(train_losses) if lr_scheduler: lr_scheduler.step() self.context["train_loss"] = train_loss logger.info(f"Train loss: {train_loss}")
1,796
36.4375
98
py
XFL
XFL-master/python/algorithm/framework/horizontal/resnet_paddle/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from algorithm.core.horizontal.template.paddle.fedavg.assist_trainer import FedAvgAssistTrainer from functools import partial from .common import Common class HorizontalResnetPaddleAssistTrainer(Common, FedAvgAssistTrainer): def __init__(self, train_conf: dict): FedAvgAssistTrainer.__init__(self, train_conf) self.register_hook(place="after_local_epoch", rank=2, func=partial(self.val_loop, "val"), desc="validation on valset") self.register_hook(place="after_local_epoch", rank=3, func=self._save_model, desc="save model") def train_loop(self): pass
1,272
38.78125
95
py
XFL
XFL-master/python/algorithm/framework/horizontal/resnet_paddle/common.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import paddle import numpy as np from algorithm.core.data_io import CsvReader, NpzReader from algorithm.model.resnet_paddle import ResNet from common.utils.logger import logger from paddle.io import TensorDataset, DataLoader import paddle.vision.transforms as transforms import paddle.nn as nn from PIL import Image class Common(): def _set_model(self) -> nn.Layer: paddle.device.set_device(self.device) model_config = self.model_info.get("config") model = ResNet(num_classes=model_config["num_classes"], layers=model_config["layers"]) return model def _read_data(self, input_dataset): if len(input_dataset) == 0: return None conf = input_dataset[0] if conf["type"] == "csv": path = os.path.join(conf['path'], conf['name']) has_label = conf["has_label"] has_id = conf['has_id'] return CsvReader(path, has_id, has_label) elif conf["type"] == "npz": path = os.path.join(conf['path'], conf['name']) return NpzReader(path) else: return None def _set_train_dataloader(self): def img_collate_fn(batch): labels = [] imgs = [] transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))]) for feature, label in batch: img = Image.fromarray(feature.numpy().astype(np.uint8)) imgs.append(transform_train(img)) labels.append(label) return paddle.stack(imgs,0), paddle.stack(labels, 0) train_data = self._read_data(self.input_trainset) trainset = None train_dataloader = None if train_data: trainset = TensorDataset([paddle.to_tensor(train_data.features(),dtype="float64"), paddle.to_tensor(train_data.label(), dtype="int64")]) batch_size = self.train_params.get("batch_size", 64) if trainset: train_dataloader = DataLoader(trainset, batch_size=batch_size, shuffle=True, collate_fn=img_collate_fn) return train_dataloader def _set_val_dataloader(self): def img_collate_fn(batch): labels = [] imgs = [] transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))]) for feature, label in batch: img = Image.fromarray(feature.numpy().astype(np.uint8)) imgs.append(transform_test(img)) labels.append(label) return paddle.stack(imgs,0), paddle.stack(labels, 0) val_data = self._read_data(self.input_valset) valset = None val_dataloader = None if val_data: valset = TensorDataset([paddle.to_tensor(val_data.features(),dtype="float64"), paddle.to_tensor(val_data.label(), dtype="int64")]) batch_size = self.train_params.get("batch_size", 64) if valset: val_dataloader = DataLoader(valset, batch_size=batch_size, shuffle=True, collate_fn=img_collate_fn) return val_dataloader def val_loop(self, dataset_type: str = "validation", context: dict = {}): self.model.eval() val_losses = [] val_predicts = [] labels = [] metric_output = {} loss_func_name = list(self.loss_func.keys())[0] loss_func = list(self.loss_func.values())[0] if dataset_type in ["validation", "val"]: dataloader = self.val_dataloader elif dataset_type == "train": dataloader = self.train_dataloader else: raise ValueError(f"dataset type {dataset_type} is not valid.") for batch, (feature, label) in enumerate(dataloader): pred = self.model(feature) loss = loss_func(pred, label) val_predicts.append(pred.squeeze(-1).numpy()) val_losses.append(loss.numpy()) labels.append(label.squeeze(-1).numpy()) val_loss = np.mean(val_losses) metric_output[loss_func_name] = val_loss val_predicts = np.concatenate(val_predicts, axis=0) labels = np.concatenate(labels, axis=0) if len(val_predicts.shape) == 1: val_predicts = np.array(val_predicts > 0.5, dtype=np.int32) elif len(val_predicts.shape) == 2: val_predicts = val_predicts.argmax(axis=-1) metrics_conf: dict = self.train_params["metric_config"] for method in self.metrics: metric_output[method] = self.metrics[method](labels, val_predicts, **metrics_conf[method]) logger.info(f"Metrics on {dataset_type} set: {metric_output}")
5,692
38.262069
148
py
XFL
XFL-master/python/algorithm/framework/horizontal/gcn_mol/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from algorithm.core.horizontal.template.agg_type import register_agg_type_for_label_trainer from common.utils.logger import logger from .common import Common from functools import partial class HorizontalGcnMolLabelTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] self.register_hook( place="after_train_loop", rank=1, func=partial(self.val_loop, "train"), desc="validation on trainset" ) register_agg_type_for_label_trainer(self, 'torch', agg_type) def train_loop(self): self.model.train() train_loss = 0 lossfunc = list(self.lossfunc.values())[0] optimizer = list(self.optimizer.values())[0] lr_scheduler = list(self.lr_scheduler.values())[ 0] if self.lr_scheduler.values() else None for batch, (smiles, bg, labels, masks) in enumerate(self.train_dataloader): node_feats = bg.ndata.pop('h') logits = self.model(bg, node_feats) labels = labels.reshape((-1,1)) loss = lossfunc(logits, labels) optimizer.zero_grad() loss.backward() optimizer.step() train_loss += loss.item() train_loss /= len(self.train_dataloader) if lr_scheduler: lr_scheduler.step() self.context["train_loss"] = train_loss logger.info(f"Train loss: {train_loss}")
2,116
36.803571
91
py
XFL
XFL-master/python/algorithm/framework/horizontal/gcn_mol/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from functools import partial from .common import Common from algorithm.core.horizontal.template.agg_type import register_agg_type_for_assist_trainer class HorizontalGcnMolAssistTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] register_agg_type_for_assist_trainer(self, 'torch', agg_type) self.register_hook(place="after_local_epoch", rank=1, func=partial(self._save_model, False), desc="save model ") self.register_hook(place="after_local_epoch", rank=2, func=partial(self.val_loop, "val"), desc="validation on valset") self.register_hook(place="after_global_epoch", rank=1, func=partial(self._save_model, True), desc="save final model") def train_loop(self): pass
1,528
42.685714
92
py
XFL
XFL-master/python/algorithm/framework/horizontal/gcn_mol/common.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from torch.utils.data import DataLoader, Dataset import dgl from dgllife.model import GCNPredictor from dgllife.utils import SMILESToBigraph, CanonicalAtomFeaturizer from algorithm.core.data_io import CsvReader, NpzReader from common.utils.logger import logger from algorithm.core.horizontal.template.torch.base import BaseTrainer from common.utils.config_sync import ConfigSynchronizer from common.checker.x_types import All from common.evaluation.metrics import CommonMetrics class SmilesDataset(Dataset): def __init__(self, smiles, graphs, labels): assert len(smiles) == len( graphs), "Inconsistent lengths of smiles and graphs" assert len(graphs) == len( labels), "Inconsistent lengths of graphs and labels" self.smiles = smiles self.graphs = graphs self.labels = labels def __len__(self): return len(self.smiles) def __getitem__(self, index): return self.smiles[index], self.graphs[index], self.labels[index] def collate_molgraphs(data): """Function from dgllife.examples.property_prediction.moleculenet.utils Batching a list of datapoints for dataloader. Parameters ---------- data : list of 3-tuples or 4-tuples. Each tuple is for a single datapoint, consisting of a SMILES, a DGLGraph, all-task labels and optionally a binary mask indicating the existence of labels. Returns ------- smiles : list List of smiles bg : DGLGraph The batched DGLGraph. labels : Tensor of dtype float32 and shape (B, T) Batched datapoint labels. B is len(data) and T is the number of total tasks. masks : Tensor of dtype float32 and shape (B, T) Batched datapoint binary mask, indicating the existence of labels. """ if len(data[0]) == 3: smiles, graphs, labels = map(list, zip(*data)) else: smiles, graphs, labels, masks = map(list, zip(*data)) bg = dgl.batch(graphs) bg.set_n_initializer(dgl.init.zero_initializer) bg.set_e_initializer(dgl.init.zero_initializer) labels = torch.stack(labels, dim=0) if len(data[0]) == 3: masks = torch.ones(labels.shape) else: masks = torch.stack(masks, dim=0) return smiles, bg, labels, masks class Common(BaseTrainer): def __init__(self, train_conf: dict) -> None: sync_rule = { "model_info": All(), "train_info": { "interaction_params": All(), "train_params": { "global_epoch": All(), "aggregation": All(), "encryption": All(), "optimizer": All(), "lr_scheduler": All(), "lossfunc": All(), "metric": All(), "early_stopping": All() } } } train_conf = ConfigSynchronizer(train_conf).sync(sync_rule) super().__init__(train_conf) def _set_model(self) -> nn.Module: model_config = self.common_config.model_info.get("config") model_params = self._prepare_model_params(model_config) model = GCNPredictor(**model_params) return model def _prepare_model_params(self, model_config): config = {} config['in_feats'] = model_config.get("input_dim", 100) num_gnn_layers = model_config.get('num_gnn_layers', 1) config['hidden_feats'] = [model_config.get( 'gnn_hidden_feats', 64)] * num_gnn_layers if model_config['activation'] == 'relu': config['activation'] = [F.relu] * num_gnn_layers elif model_config['activation'] == 'tanh': config['activation'] = [F.tanh] * num_gnn_layers else: logger.info(f"Setting gnn activation to relu") config['activation'] = [F.relu] * num_gnn_layers config['dropout'] = [model_config.get("dropout", 0.5)] * num_gnn_layers config['batchnorm'] = [model_config.get( 'batchnorm', False)] * num_gnn_layers config['residual'] = [model_config.get( "residual", False)] * num_gnn_layers config['predictor_hidden_feats'] = model_config.get( 'predictor_hidden_dim', 64) config['n_tasks'] = model_config.get('n_tasks', 1) return config def _read_data(self, input_dataset): if len(input_dataset) == 0: return None conf = input_dataset[0] if conf["type"] == "csv": path = os.path.join(conf['path'], conf['name']) has_label = conf["has_label"] has_id = conf['has_id'] return CsvReader(path, has_id, has_label) elif conf["type"] == "npz": path = os.path.join(conf['path'], conf['name']) return NpzReader(path) else: return None def _set_train_dataloader(self): train_data = self._read_data(self.common_config.input_trainset) trainset = None train_dataloader = None if train_data is None: return train_dataloader # construct smiles smiles = train_data.features(type="series").values.reshape((-1)) labels = train_data.label().astype(np.int32) smiles_to_g = SMILESToBigraph( add_self_loop=True, node_featurizer=CanonicalAtomFeaturizer() ) graph_list = [] for smile in smiles: graph_list.append(smiles_to_g(smile)) valid_ids = [] clean_graphs = [] failed_mols = [] clean_labels = [] clean_smiles = [] for i, g in enumerate(graph_list): if g is not None: valid_ids.append(i) clean_graphs.append(g) clean_labels.append(labels[i]) clean_smiles.append(smiles[i]) else: failed_mols.append((i, smiles[i])) # construct dataset if train_data: trainset = SmilesDataset( clean_smiles, clean_graphs, torch.Tensor(clean_labels)) # construct dataloader batch_size = self.common_config.train_params.get("train_batch_size") if trainset: train_dataloader = DataLoader( trainset, batch_size=batch_size, shuffle=True, collate_fn=collate_molgraphs ) return train_dataloader def _set_val_dataloader(self): val_data = self._read_data(self.common_config.input_valset) valset = None val_dataloader = None if val_data is None: return val_dataloader # construct smiles smiles = val_data.features(type="series").values.reshape((-1)) labels = val_data.label().astype(np.int32) smiles_to_g = SMILESToBigraph( add_self_loop=True, node_featurizer=CanonicalAtomFeaturizer() ) graph_list = [] for smile in smiles: graph_list.append(smiles_to_g(smile)) valid_ids = [] clean_graphs = [] failed_mols = [] clean_labels = [] clean_smiles = [] for i, g in enumerate(graph_list): if g is not None: valid_ids.append(i) clean_graphs.append(g) clean_labels.append(labels[i]) clean_smiles.append(smiles[i]) else: failed_mols.append((i, smiles[i])) # construct dataset if val_data: valset = SmilesDataset( clean_smiles, clean_graphs, torch.Tensor(clean_labels)) # construct dataloader batch_size = self.common_config.train_params.get("val_batch_size") if valset: val_dataloader = DataLoader( valset, batch_size=batch_size, shuffle=True, collate_fn=collate_molgraphs ) return val_dataloader def val_loop(self, dataset_type: str = "val", context: dict = {}): self.model.eval() val_loss = 0 val_predicts = [] labels = [] lossfunc_name = list(self.lossfunc.keys())[0] lossfunc = list(self.lossfunc.values())[0] if dataset_type == "val": dataloader = self.val_dataloader elif dataset_type == "train": dataloader = self.train_dataloader else: raise ValueError(f"dataset type {dataset_type} is not valid.") for batch, (smiles, bg, label, masks) in enumerate(dataloader): node_feats = bg.ndata.pop('h') logits = self.model(bg, node_feats) label = label.reshape((-1, 1)) loss = lossfunc(logits, label) val_predicts.append(logits.detach().cpu().squeeze(-1).numpy()) val_loss += loss.item() labels.append(label.cpu().squeeze(-1).numpy()) val_loss /= len(dataloader) labels: np.ndarray = np.concatenate(labels, axis=0) val_predicts: np.ndarray = np.concatenate(val_predicts, axis=0) if len(val_predicts.shape) == 1: val_predicts = np.array(val_predicts > 0.5, dtype=np.int32) elif len(val_predicts.shape) == 2: val_predicts = val_predicts.argmax(axis=-1) metrics_output = CommonMetrics._calc_metrics( metrics=self.metrics, labels=labels, val_predicts=val_predicts, lossfunc_name=lossfunc_name, loss=val_loss, dataset_type=dataset_type ) global_epoch = self.context["g_epoch"] if dataset_type == "val": local_epoch = None elif dataset_type == "train": local_epoch = self.context["l_epoch"] CommonMetrics.save_metric_csv( metrics_output=metrics_output, output_config=self.common_config.output, global_epoch=global_epoch, local_epoch=local_epoch, dataset_type=dataset_type, ) early_stop_flag = self.context["early_stop_flag"] if (self.common_config.save_frequency > 0) & \ (dataset_type == "val") & (self.earlystopping.patience > 0): early_stop_flag = self.earlystopping(metrics_output, global_epoch) if early_stop_flag: # find the saved epoch closest to the best epoch best_epoch = self.earlystopping.best_epoch closest_epoch = round(best_epoch / self.common_config.save_frequency) * \ self.common_config.save_frequency closest_epoch -= self.common_config.save_frequency \ if closest_epoch > global_epoch else 0 self.context["early_stop_flag"] = True self.context["early_stop_epoch"] = closest_epoch
11,572
34.069697
89
py
XFL
XFL-master/python/algorithm/framework/horizontal/densenet/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from algorithm.core.horizontal.template.agg_type import register_agg_type_for_label_trainer from common.utils.logger import logger from .common import Common from functools import partial class HorizontalDensenetLabelTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] self.register_hook( place="after_train_loop", rank=1, func=partial(self.val_loop, "train"), desc="validation on trainset" ) register_agg_type_for_label_trainer(self, 'torch', agg_type) def train_loop(self): self.model.train() train_loss = 0 lossfunc = list(self.lossfunc.values())[0] optimizer = list(self.optimizer.values())[0] lr_scheduler = list(self.lr_scheduler.values())[0] if self.lr_scheduler.values() else None for batch, (feature, label) in enumerate(self.train_dataloader): pred = self.model(feature) loss = lossfunc(pred, label) optimizer.zero_grad() loss.backward() optimizer.step() train_loss += loss.item() train_loss /= len(self.train_dataloader) if lr_scheduler: lr_scheduler.step() self.context["train_loss"] = train_loss logger.info(f"Train loss: {train_loss}")
2,017
37.807692
98
py
XFL
XFL-master/python/algorithm/framework/horizontal/densenet/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from functools import partial from .common import Common from algorithm.core.horizontal.template.agg_type import register_agg_type_for_assist_trainer class HorizontalDensenetAssistTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] register_agg_type_for_assist_trainer(self, 'torch', agg_type) self.register_hook(place="after_local_epoch", rank=1, func=partial(self._save_model, False), desc="save model ") self.register_hook(place="after_local_epoch", rank=2, func=partial(self.val_loop, "val"), desc="validation on valset") self.register_hook(place="after_global_epoch", rank=1, func=partial(self._save_model, True), desc="save final model") def train_loop(self): pass
1,537
42.942857
92
py
XFL
XFL-master/python/algorithm/framework/horizontal/densenet/common.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import numpy as np import torch import torch.nn as nn from torch.utils.data import DataLoader, TensorDataset from PIL import Image import torchvision.transforms as transforms from algorithm.core.data_io import CsvReader, NpzReader from algorithm.model.densenet import DenseNet from common.utils.logger import logger from algorithm.core.horizontal.template.torch.base import BaseTrainer from common.utils.config_sync import ConfigSynchronizer from common.checker.x_types import All from common.evaluation.metrics import CommonMetrics class Common(BaseTrainer): def __init__(self, train_conf: dict) -> None: sync_rule = { "model_info": All(), "train_info": { "interaction_params": All(), "train_params": { "global_epoch": All(), "aggregation": All(), "encryption": All(), "optimizer": All(), "lr_scheduler": All(), "lossfunc": All(), "metric": All(), "early_stopping": All() } } } train_conf = ConfigSynchronizer(train_conf).sync(sync_rule) super().__init__(train_conf) def _set_model(self) -> nn.Module: model_config = self.common_config.model_info.get("config") model = DenseNet(num_classes=model_config["num_classes"], layers=model_config["layers"]) model = model.to(self.device) return model def _read_data(self, input_dataset): if len(input_dataset) == 0: return None conf = input_dataset[0] if conf["type"] == "csv": path = os.path.join(conf['path'], conf['name']) has_label = conf["has_label"] has_id = conf['has_id'] return CsvReader(path, has_id, has_label) elif conf["type"] == "npz": path = os.path.join(conf['path'], conf['name']) return NpzReader(path) else: return None def _set_train_dataloader(self): def img_collate_fn(batch): labels = [] imgs = [] transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))]) for feature, label in batch: img = Image.fromarray(feature.numpy().astype(np.uint8)) imgs.append(transform_train(img)) labels.append(label) return torch.stack(imgs,0).to(self.device), torch.stack(labels, 0).long().to(self.device) train_data = self._read_data(self.common_config.input_trainset) trainset = None train_dataloader = None if train_data: trainset = TensorDataset( torch.tensor(train_data.features()), torch.tensor(train_data.label()) ) batch_size = self.common_config.train_params.get("train_batch_size") if trainset: train_dataloader = DataLoader( trainset, batch_size, shuffle=True, collate_fn=img_collate_fn ) return train_dataloader def _set_val_dataloader(self): def img_collate_fn(batch): labels = [] imgs = [] transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))]) for feature, label in batch: img = Image.fromarray(feature.numpy().astype(np.uint8)) imgs.append(transform_test(img)) labels.append(label) return torch.stack(imgs,0).to(self.device), torch.stack(labels, 0).long().to(self.device) val_data = self._read_data(self.common_config.input_valset) valset = None val_dataloader = None if val_data: valset = TensorDataset( torch.tensor(val_data.features()), torch.tensor(val_data.label()) ) batch_size = self.common_config.train_params.get("val_batch_size") if valset: val_dataloader = DataLoader( valset, batch_size, shuffle=True, collate_fn=img_collate_fn ) return val_dataloader def val_loop(self, dataset_type: str = "val", context: dict = {}): self.model.eval() val_loss = 0 val_predicts = [] labels = [] lossfunc_name = list(self.lossfunc.keys())[0] lossfunc = list(self.lossfunc.values())[0] if dataset_type == "val": dataloader = self.val_dataloader elif dataset_type == "train": dataloader = self.train_dataloader else: raise ValueError(f"dataset type {dataset_type} is not valid.") for batch, (feature, label) in enumerate(dataloader): pred = self.model(feature) loss = lossfunc(pred, label) val_predicts.append(pred.detach().cpu().squeeze(-1).numpy()) val_loss += loss.item() labels.append(label.cpu().squeeze(-1).numpy()) val_loss /= len(dataloader) labels: np.ndarray = np.concatenate(labels, axis=0) val_predicts: np.ndarray = np.concatenate(val_predicts, axis=0) if len(val_predicts.shape) == 1: val_predicts = np.array(val_predicts > 0.5, dtype=np.int32) elif len(val_predicts.shape) == 2: val_predicts = val_predicts.argmax(axis=-1) metrics_output = CommonMetrics._calc_metrics( metrics=self.metrics, labels=labels, val_predicts=val_predicts, lossfunc_name=lossfunc_name, loss=val_loss, dataset_type=dataset_type ) global_epoch = self.context["g_epoch"] if dataset_type == "val": local_epoch = None elif dataset_type == "train": local_epoch = self.context["l_epoch"] CommonMetrics.save_metric_csv( metrics_output=metrics_output, output_config=self.common_config.output, global_epoch=global_epoch, local_epoch=local_epoch, dataset_type=dataset_type, ) early_stop_flag = self.context["early_stop_flag"] if (self.common_config.save_frequency > 0) & \ (dataset_type == "val") & (self.earlystopping.patience > 0): early_stop_flag = self.earlystopping(metrics_output, global_epoch) if early_stop_flag: # find the saved epoch closest to the best epoch best_epoch = self.earlystopping.best_epoch closest_epoch = round(best_epoch / self.common_config.save_frequency) * \ self.common_config.save_frequency closest_epoch -= self.common_config.save_frequency \ if closest_epoch > global_epoch else 0 self.context["early_stop_flag"] = True self.context["early_stop_epoch"] = closest_epoch
7,943
37.75122
101
py
XFL
XFL-master/python/algorithm/framework/horizontal/logistic_regression/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from functools import partial from algorithm.core.horizontal.template.agg_type import register_agg_type_for_label_trainer from common.utils.logger import logger from .common import Common class HorizontalLogisticRegressionLabelTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] register_agg_type_for_label_trainer(self, 'torch', agg_type) self.register_hook( place="after_train_loop", rank=1, func=partial(self.val_loop, "train"), desc="validation on trainset" ) def train_loop(self): self.model.train() train_loss = 0 lossfunc = list(self.lossfunc.values())[0] optimizer = list(self.optimizer.values())[0] lr_scheduler = list(self.lr_scheduler.values())[0] if self.lr_scheduler.values() else None for batch, (feature, label) in enumerate(self.train_dataloader): pred = self.model(feature) loss = lossfunc(pred, label) optimizer.zero_grad() loss.backward() optimizer.step() train_loss += loss.item() train_loss /= len(self.train_dataloader) if lr_scheduler: lr_scheduler.step() self.context["train_loss"] = train_loss logger.info(f"Train loss: {train_loss}")
2,042
36.145455
98
py
XFL
XFL-master/python/algorithm/framework/horizontal/logistic_regression/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from functools import partial from .common import Common from algorithm.core.horizontal.template.agg_type import register_agg_type_for_assist_trainer class HorizontalLogisticRegressionAssistTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] register_agg_type_for_assist_trainer(self, 'torch', agg_type) self.register_hook(place="after_local_epoch", rank=1, func=partial(self._save_model, False), desc="save model ") self.register_hook(place="after_local_epoch", rank=2, func=partial(self.val_loop, "val"), desc="validation on valset") self.register_hook(place="after_global_epoch", rank=1, func=partial(self._save_model, True), desc="save final model") def train_loop(self): pass
1,543
43.114286
92
py
XFL
XFL-master/python/algorithm/framework/horizontal/logistic_regression/common.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import numpy as np import torch import torch.nn as nn from torch.utils.data import DataLoader, TensorDataset from algorithm.core.data_io import CsvReader from algorithm.core.horizontal.template.torch.base import BaseTrainer from algorithm.model.logistic_regression import LogisticRegression from common.checker.x_types import All from common.evaluation.metrics import CommonMetrics from common.utils.config_sync import ConfigSynchronizer class Common(BaseTrainer): def __init__(self, train_conf: dict): sync_rule = { "model_info": { "config": All() }, "train_info": { "interaction_params": All(), "train_params": { "global_epoch": All(), "aggregation": All(), "encryption": All(), "optimizer": All(), "lr_scheduler": All(), "lossfunc": All(), "metric": All(), "early_stopping": All() } } } train_conf = ConfigSynchronizer(train_conf).sync(sync_rule) super().__init__(train_conf) def _set_model(self) -> nn.Module: model_config = self.common_config.model_info.get("config") model = LogisticRegression(input_dim=model_config["input_dim"], bias=model_config["bias"]) return model def _read_data(self, input_dataset): if len(input_dataset) == 0: return None conf = input_dataset[0] if conf["type"] == "csv": path = os.path.join(conf['path'], conf['name']) has_label = conf["has_label"] has_id = conf['has_id'] return CsvReader(path, has_id, has_label) else: return None def _set_train_dataloader(self): train_data = self._read_data(self.common_config.input_trainset) trainset = None train_dataloader = None if train_data: trainset = TensorDataset( torch.tensor(train_data.features(), dtype=torch.float32).to(self.device), torch.tensor(train_data.label(), dtype=torch.float32).unsqueeze( dim=-1).to(self.device) ) batch_size = self.common_config.train_params.get("train_batch_size") if trainset: train_dataloader = DataLoader(trainset, batch_size, shuffle=True) return train_dataloader def _set_val_dataloader(self): val_data = self._read_data(self.common_config.input_valset) valset = None val_dataloader = None if val_data: valset = TensorDataset( torch.tensor(val_data.features(), dtype=torch.float32).to(self.device), torch.tensor(val_data.label(), dtype=torch.float32).unsqueeze( dim=-1).to(self.device) ) batch_size = self.common_config.train_params.get("val_batch_size") if valset: val_dataloader = DataLoader(valset, batch_size, shuffle=True) return val_dataloader def val_loop(self, dataset_type: str = "val", context: dict = {}): self.model.eval() val_loss = 0 val_predicts = [] labels = [] lossfunc_name = list(self.lossfunc.keys())[0] lossfunc = list(self.lossfunc.values())[0] if dataset_type == "val": dataloader = self.val_dataloader elif dataset_type == "train": dataloader = self.train_dataloader else: raise ValueError(f"dataset type {dataset_type} is not valid.") for batch, (feature, label) in enumerate(dataloader): pred = self.model(feature) loss = lossfunc(pred, label) val_predicts.append(pred.detach().cpu().squeeze(-1).numpy()) val_loss += loss.item() labels.append(label.cpu().squeeze(-1).numpy()) val_loss /= len(dataloader) labels: np.ndarray = np.concatenate(labels, axis=0) val_predicts: np.ndarray = np.concatenate(val_predicts, axis=0) if len(val_predicts.shape) == 1: val_predicts = np.array(val_predicts > 0.5, dtype=np.int32) elif len(val_predicts.shape) == 2: val_predicts = val_predicts.argmax(axis=-1) metrics_output = CommonMetrics._calc_metrics( metrics=self.metrics, labels=labels, val_predicts=val_predicts, lossfunc_name=lossfunc_name, loss=val_loss, dataset_type=dataset_type ) global_epoch = self.context["g_epoch"] if dataset_type == "val": local_epoch = None elif dataset_type == "train": local_epoch = self.context["l_epoch"] CommonMetrics.save_metric_csv( metrics_output=metrics_output, output_config=self.common_config.output, global_epoch=global_epoch, local_epoch=local_epoch, dataset_type=dataset_type, ) early_stop_flag = self.context["early_stop_flag"] if (self.common_config.save_frequency > 0) & \ (dataset_type == "val") & (self.earlystopping.patience > 0): early_stop_flag = self.earlystopping(metrics_output, global_epoch) if early_stop_flag: # find the saved epoch closest to the best epoch best_epoch = self.earlystopping.best_epoch closest_epoch = round(best_epoch / self.common_config.save_frequency) * \ self.common_config.save_frequency closest_epoch -= self.common_config.save_frequency \ if closest_epoch > global_epoch else 0 self.context["early_stop_flag"] = True self.context["early_stop_epoch"] = closest_epoch
6,601
36.089888
89
py
XFL
XFL-master/python/algorithm/framework/horizontal/logistic_regression/__init__.py
0
0
0
py
XFL
XFL-master/python/algorithm/framework/horizontal/resnet/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from algorithm.core.horizontal.template.agg_type import register_agg_type_for_label_trainer from common.utils.logger import logger from .common import Common from functools import partial class HorizontalResnetLabelTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] self.register_hook( place="after_train_loop", rank=1, func=partial(self.val_loop, "train"), desc="validation on trainset" ) register_agg_type_for_label_trainer(self, 'torch', agg_type) def train_loop(self): self.model.train() train_loss = 0 lossfunc = list(self.lossfunc.values())[0] optimizer = list(self.optimizer.values())[0] lr_scheduler = list(self.lr_scheduler.values())[0] if self.lr_scheduler.values() else None for batch, (feature, label) in enumerate(self.train_dataloader): pred = self.model(feature) loss = lossfunc(pred, label) optimizer.zero_grad() loss.backward() optimizer.step() train_loss += loss.item() train_loss /= len(self.train_dataloader) if lr_scheduler: lr_scheduler.step() self.context["train_loss"] = train_loss logger.info(f"Train loss: {train_loss}")
2,015
37.769231
98
py
XFL
XFL-master/python/algorithm/framework/horizontal/resnet/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from functools import partial from .common import Common from algorithm.core.horizontal.template.agg_type import register_agg_type_for_assist_trainer class HorizontalResnetAssistTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf) agg_type = list(self.common_config.aggregation["method"].keys())[0] register_agg_type_for_assist_trainer(self, 'torch', agg_type) self.register_hook(place="after_local_epoch", rank=1, func=partial(self._save_model, False), desc="save model ") self.register_hook(place="after_local_epoch", rank=2, func=partial(self.val_loop, "val"), desc="validation on valset") self.register_hook(place="after_global_epoch", rank=1, func=partial(self._save_model, True), desc="save final model") def train_loop(self): pass
1,531
42.771429
92
py
XFL
XFL-master/python/algorithm/framework/horizontal/resnet/common.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import numpy as np import torch import torch.nn as nn from torch.utils.data import DataLoader, TensorDataset from PIL import Image import torchvision.transforms as transforms from algorithm.core.data_io import CsvReader, NpzReader from algorithm.model.resnet import ResNet from common.utils.logger import logger from algorithm.core.horizontal.template.torch.base import BaseTrainer from common.utils.config_sync import ConfigSynchronizer from common.checker.x_types import All from common.evaluation.metrics import CommonMetrics class Common(BaseTrainer): def __init__(self, train_conf: dict) -> None: sync_rule = { "model_info": All(), "train_info": { "interaction_params": All(), "train_params": { "global_epoch": All(), "aggregation": All(), "encryption": All(), "optimizer": All(), "lr_scheduler": All(), "lossfunc": All(), "metric": All(), "early_stopping": All() } } } train_conf = ConfigSynchronizer(train_conf).sync(sync_rule) super().__init__(train_conf) def _set_model(self) -> nn.Module: model_config = self.common_config.model_info.get("config") model = ResNet(num_classes=model_config["num_classes"], layers=model_config["layers"]) model = model.to(self.device) # model = torch.nn.DataParallel(model) # torch.backends.cudnn.benchmark = True return model def _read_data(self, input_dataset): if len(input_dataset) == 0: return None conf = input_dataset[0] if conf["type"] == "csv": path = os.path.join(conf['path'], conf['name']) has_label = conf["has_label"] has_id = conf['has_id'] return CsvReader(path, has_id, has_label) elif conf["type"] == "npz": path = os.path.join(conf['path'], conf['name']) return NpzReader(path) else: return None def _set_train_dataloader(self): def img_collate_fn(batch): labels = [] imgs = [] transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))]) for feature, label in batch: img = Image.fromarray(feature.numpy().astype(np.uint8)) imgs.append(transform_train(img)) labels.append(label) return torch.stack(imgs,0).to(self.device), torch.stack(labels, 0).long().to(self.device) train_data = self._read_data(self.common_config.input_trainset) trainset = None train_dataloader = None if train_data: trainset = TensorDataset( torch.tensor(train_data.features()), torch.tensor(train_data.label()) ) batch_size = self.common_config.train_params.get("train_batch_size") if trainset: train_dataloader = DataLoader( trainset, batch_size, shuffle=True, collate_fn=img_collate_fn ) return train_dataloader def _set_val_dataloader(self): def img_collate_fn(batch): labels = [] imgs = [] transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))]) for feature, label in batch: img = Image.fromarray(feature.numpy().astype(np.uint8)) imgs.append(transform_test(img)) labels.append(label) return torch.stack(imgs,0).to(self.device), torch.stack(labels, 0).long().to(self.device) val_data = self._read_data(self.common_config.input_valset) valset = None val_dataloader = None if val_data: valset = TensorDataset( torch.tensor(val_data.features()), torch.tensor(val_data.label()) ) batch_size = self.common_config.train_params.get("val_batch_size") if valset: val_dataloader = DataLoader( valset, batch_size, shuffle=True, collate_fn=img_collate_fn ) return val_dataloader def val_loop(self, dataset_type: str = "val", context: dict = {}): self.model.eval() val_loss = 0 val_predicts = [] labels = [] lossfunc_name = list(self.lossfunc.keys())[0] lossfunc = list(self.lossfunc.values())[0] if dataset_type == "val": dataloader = self.val_dataloader elif dataset_type == "train": dataloader = self.train_dataloader else: raise ValueError(f"dataset type {dataset_type} is not valid.") for batch, (feature, label) in enumerate(dataloader): pred = self.model(feature) loss = lossfunc(pred, label) val_predicts.append(pred.detach().cpu().squeeze(-1).numpy()) val_loss += loss.item() labels.append(label.cpu().squeeze(-1).numpy()) val_loss /= len(dataloader) labels: np.ndarray = np.concatenate(labels, axis=0) val_predicts: np.ndarray = np.concatenate(val_predicts, axis=0) if len(val_predicts.shape) == 1: val_predicts = np.array(val_predicts > 0.5, dtype=np.int32) elif len(val_predicts.shape) == 2: val_predicts = val_predicts.argmax(axis=-1) metrics_output = CommonMetrics._calc_metrics( metrics=self.metrics, labels=labels, val_predicts=val_predicts, lossfunc_name=lossfunc_name, loss=val_loss, dataset_type=dataset_type ) global_epoch = self.context["g_epoch"] if dataset_type == "val": local_epoch = None elif dataset_type == "train": local_epoch = self.context["l_epoch"] CommonMetrics.save_metric_csv( metrics_output=metrics_output, output_config=self.common_config.output, global_epoch=global_epoch, local_epoch=local_epoch, dataset_type=dataset_type, ) early_stop_flag = self.context["early_stop_flag"] if (self.common_config.save_frequency > 0) & \ (dataset_type == "val") & (self.earlystopping.patience > 0): early_stop_flag = self.earlystopping(metrics_output, global_epoch) if early_stop_flag: # find the saved epoch closest to the best epoch best_epoch = self.earlystopping.best_epoch closest_epoch = round(best_epoch / self.common_config.save_frequency) * \ self.common_config.save_frequency closest_epoch -= self.common_config.save_frequency \ if closest_epoch > global_epoch else 0 self.context["early_stop_flag"] = True self.context["early_stop_epoch"] = closest_epoch
8,036
37.826087
101
py
XFL
XFL-master/python/algorithm/framework/transfer/base.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import torch import torch.nn as nn from typing import OrderedDict from functools import partial from common.utils.config_parser import CommonConfigParser from algorithm.core.loss.torch_loss import get_lossfunc from algorithm.core.metrics import get_metric from algorithm.core.optimizer.torch_optimizer import get_optimizer from algorithm.core.lr_scheduler.torch_lr_scheduler import get_lr_scheduler class BaseTrainer: def __init__(self, train_conf: dict): self.common_config = CommonConfigParser(train_conf) if self.common_config.random_seed is not None: self.set_seed(self.common_config.random_seed) self.device = self.common_config.device self._parse_config() self.model = self._set_model() self._set_train_dataloader() self._set_val_dataloader() self.optimizer = self._set_optimizer() self.lr_scheduler = self._set_lr_scheduler(self.optimizer) self.lossfunc = self._set_lossfunc() self.optimizer = self._set_optimizer() self.lr_scheduler = self._set_lr_scheduler(self.optimizer) self.metrics = self._set_metrics() @staticmethod def set_seed(seed): torch.manual_seed(seed) torch.cuda.manual_seed(seed) def _parse_config(self) -> nn.Module: raise NotImplementedError("The _parse_config method is not implemented.") def _set_optimizer(self): """ Define self.optimizer """ optimizer_conf = OrderedDict(self.common_config.optimizer) optimizer = OrderedDict() for k, v in optimizer_conf.items(): optimizer[k] = get_optimizer(k)(self.model.parameters(), **v) return optimizer def _set_lossfunc(self): """ Define self.lossfunc """ lossfunc_conf = OrderedDict(self.common_config.lossfunc) lossfunc = OrderedDict() for k, v in lossfunc_conf.items(): lossfunc[k] = get_lossfunc(k)(**v) return lossfunc def _set_lr_scheduler(self, optimizer: OrderedDict): lr_scheduler_conf = OrderedDict(self.common_config.lr_scheduler) lr_scheduler = OrderedDict() for (k, v), o in zip(lr_scheduler_conf.items(), optimizer.values()): lr_scheduler[k] = get_lr_scheduler(k)(o, **v) return lr_scheduler def _set_metrics(self): """ Define metric """ metrics = {} metrics_conf: dict = self.common_config.metric for k, v in metrics_conf.items(): metric = get_metric(k) metrics[k] = partial(metric, **v) return metrics def _set_model(self) -> nn.Module: raise NotImplementedError("The _set_model method is not implemented.") def _set_train_dataloader(self): raise NotImplementedError( "The _set_train_dataloader method is not implemented.") def _set_val_dataloader(self): raise NotImplementedError( "The _set_val_dataloader method is not implemented.")
3,594
34.95
81
py
XFL
XFL-master/python/algorithm/framework/transfer/transfer_model_base.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from pathlib import Path import torch from functools import partial from common.utils.config_parser import TrainConfigParser from algorithm.core.metrics import get_metric class TransferModelBase(TrainConfigParser): def __init__(self, train_conf: dict, label: bool = False): super().__init__(train_conf) self.train_conf = train_conf self.model_conf = train_conf["model_info"].get("config") self.label = label def _parse_config(self) -> None: self.save_dir = Path(self.output.get("path", "")) self.metric_dir = self.save_dir # interaction_params self.model_name = self.model_info.get("name") self.save_model_name = self.output.get("model", {}).get("name", {}) self.pretrain_model_path = self.input.get("pretrained_model", {}).get("path") self.num_features = self.model_conf.get("num_features") self.hidden_features = self.model_conf.get("hidden_features") self.constant_k = 1 / self.hidden_features self.alpha = self.model_conf.get("alpha") self.global_epoch = self.train_params.get("global_epoch") self.local_epoch = self.train_params.get("local_epoch") self.batch_size = self.train_params.get("batch_size") self.shuffle_seed = self.train_params.get("shuffle_seed") self.random_seed = self.train_params.get("random_seed") @staticmethod def set_seed(seed): torch.manual_seed(seed) torch.cuda.manual_seed(seed) def _set_metrics(self): """ Define metric """ metrics = {} metrics_conf: dict = self.train_params.get("metric", {}) for k, v in metrics_conf.items(): metric = get_metric(k) metrics[k] = partial(metric, **v) return metrics
2,407
36.046154
85
py
XFL
XFL-master/python/algorithm/framework/transfer/__init__.py
0
0
0
py
XFL
XFL-master/python/algorithm/framework/transfer/logistic_regression/base.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import pandas as pd import numpy as np import torch import torch.nn as nn from torch.utils.data import DataLoader, TensorDataset from algorithm.framework.transfer.transfer_model_base import TransferModelBase from common.utils.logger import logger from common.utils.model_preserver import ModelPreserver class TransferLogisticRegression(nn.Module): def __init__(self, input_dim: int, output_dim: int, bias: bool = False): super().__init__() self.linear = torch.nn.Linear(input_dim, output_dim, bias=bias) def forward(self, x): return self.linear(x) class TransferLogisticRegressionBase(TransferModelBase): def __init__(self, train_conf: dict, label: bool = False, *args, **kwargs): """_summary_ Args: train_conf (dict): _description_ label (bool, optional): _description_. Defaults to False. """ super().__init__(train_conf) self._parse_config() self.label = label self.model = None self.phi = None # phi will be saved in the checkpoint of label_trainer self.overlap_y, self.non_overlap_y = None, None self.overlap_train_dataloader, self.non_overlap_train_dataloader = None, None self.eval_dataloader = None self.metric_functions = {} self._set_train_dataloader() self._set_val_dataloader() def _init_model(self, bias: bool = False) -> None: """ Init logistic regression model. Returns: None """ logger.info("Init model start.") self.model = TransferLogisticRegression( input_dim=self.num_features, output_dim=self.hidden_features, bias=bias ) # Load pretrained model if needed. if self.pretrain_model_path is not None and self.pretrain_model_path != "": checkpoint = ModelPreserver.load( os.path.join(self.pretrain_model_path, self.input.get("pretrained_model").get("name"))) state_dict = checkpoint["state_dict"] if "phi" in state_dict.keys(): self.phi = state_dict.pop("phi") self.model.load_state_dict(state_dict) self.model = self.model.to(self.device) logger.info("Init model completed.") def _read_data(self, input_dataset, is_train=True): if len(input_dataset) == 0: return None conf = input_dataset[0] if conf["type"] == "csv": path = os.path.join(conf['path'], conf['name']) has_id = conf['has_id'] index_col = 0 if has_id else False train_data = pd.read_csv(path, index_col=index_col) if is_train: index_name = "overlap_index.npy" index_path = os.path.join(conf['path'], index_name) overlap_index = np.load(index_path) return train_data, overlap_index else: return train_data else: raise NotImplementedError( "Dataset load method {} does not Implemented.".format(conf["type"]) ) def _set_train_dataloader(self): train_data, overlap_index = self._read_data(self.input_trainset) self.sample_num = train_data.shape[0] overlap_train_data = train_data.loc[overlap_index] if self.label: non_overlap_index = np.array([]) for i in train_data.index: if i not in overlap_index: non_overlap_index = np.append(non_overlap_index, i) non_overlap_train_data = train_data.loc[non_overlap_index] # init overlap_y and non_overlap_y self.overlap_y = torch.tensor(overlap_train_data.iloc[:, 0].to_numpy(), dtype=torch.float32).unsqueeze(1) self.non_overlap_y = torch.tensor(non_overlap_train_data.iloc[:, 0].to_numpy(), dtype=torch.float32).unsqueeze(1) # init train_dataloader overlap_x = torch.tensor(overlap_train_data.iloc[:, 1:].to_numpy(), dtype=torch.float32) overlap_trainset = TensorDataset(overlap_x, self.overlap_y) self.overlap_train_dataloader = DataLoader(overlap_trainset, batch_size=self.batch_size, shuffle=False) non_overlap_x = torch.tensor(non_overlap_train_data.iloc[:, 1:].to_numpy(), dtype=torch.float32) non_overlap_trainset = TensorDataset(non_overlap_x, self.non_overlap_y) self.non_overlap_train_dataloader = DataLoader(non_overlap_trainset, batch_size=self.batch_size, shuffle=False) else: # init train_dataloader overlap_x = torch.tensor(overlap_train_data.to_numpy(), dtype=torch.float32) overlap_trainset = TensorDataset(overlap_x) self.overlap_train_dataloader = DataLoader(overlap_trainset, batch_size=self.batch_size, shuffle=False) def _set_val_dataloader(self): val_data = self._read_data(self.input_valset, is_train=False) if self.label: # init val_dataloader labels = torch.tensor(val_data.iloc[:, 0].to_numpy(), dtype=torch.float32).unsqueeze(dim=-1) valset = TensorDataset(labels) self.val_dataloader = DataLoader(valset, batch_size=self.batch_size, shuffle=False) else: # init val_dataloader features = torch.tensor(val_data.to_numpy(), dtype=torch.float32) valset = TensorDataset(features) self.val_dataloader = DataLoader(valset, batch_size=self.batch_size, shuffle=False)
6,172
42.167832
125
py
XFL
XFL-master/python/algorithm/framework/transfer/logistic_regression/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import numpy as np import torch from common.communication.gRPC.python.channel import DualChannel from common.utils.logger import logger from service.fed_config import FedConfig from .common import Common from common.utils.model_io import ModelIO from common.evaluation.metrics import CommonMetrics class TransferLogisticRegressionLabelTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf, label=True) def cal_phi_and_ua(self, dataloader): phi = None # [1, hidden_features] Φ_A ua = [] for batch_idx, (x_batch, y_batch) in enumerate(dataloader): x_batch = x_batch.to(self.device) ua_batch = self.model(x_batch) # [batch_size, hidden_features] ua.append(ua_batch) phi_batch = torch.sum(y_batch * ua_batch, axis=0).unsqueeze(0) if phi is None: phi = phi_batch else: phi += phi_batch ua = torch.concat(ua, axis=0) return phi, ua def cal_parameters(self, dual_channel): overlap_phi, overlap_ua = self.cal_phi_and_ua(self.overlap_train_dataloader) non_overlap_phi, non_overlap_ua = self.cal_phi_and_ua(self.non_overlap_train_dataloader) phi = (overlap_phi + non_overlap_phi) / self.sample_num phi_2 = torch.matmul(phi.T, phi) # (Φ_A)‘(Φ_A) [hidden_features, hidden_features] overlap_y = self.overlap_y # {C(y)=y} [overlap_size, 1] overlap_y_2 = overlap_y * overlap_y # {D(y)=y^2} [overlap_size, 1] # calculate 3 components will be sent to the trainer overlap_y_2_phi_2 = 0.25 * overlap_y_2.unsqueeze(2) * phi_2 # [overlap_size, hidden_features, hidden_features] overlap_y_phi = -0.5 * overlap_y * phi # [overlap_size, hidden_features] comp_ua = -self.constant_k * overlap_ua # [overlap_size, 1] # exchange components dual_channel.send((overlap_y_2_phi_2, overlap_y_phi, comp_ua)) ub, ub_2, comp_ub = dual_channel.recv() # compute gradients to excute backward overlap_y_2_phi = (overlap_y_2 * phi).unsqueeze(1) loss_grads_const_part1 = 0.25 * torch.matmul(overlap_y_2_phi, ub_2).squeeze(1) loss_grads_const_part2 = overlap_y * ub const = torch.sum(loss_grads_const_part1, axis=0) - 0.5 * torch.sum(loss_grads_const_part2, axis=0) non_overlap_y = self.non_overlap_y non_overlap_ua_grad = self.alpha * const * non_overlap_y / self.sample_num overlap_ua_grad = self.alpha * const * overlap_y / self.sample_num + comp_ub # compute loss overlap_num = overlap_y.shape[0] overlap_loss = torch.sum(comp_ua * ub) ub_phi = torch.matmul(ub, phi.T) part1 = -0.5 * torch.sum(overlap_y * ub_phi) part2 = 1.0 / 8 * torch.sum(ub_phi * ub_phi) part3 = len(overlap_y) * np.log(2) loss_y = part1 + part2 + part3 loss = self.alpha * (loss_y / overlap_num) + overlap_loss / overlap_num ua = torch.concat((overlap_ua, non_overlap_ua), axis=0) ua_grad = torch.concat((overlap_ua_grad, non_overlap_ua_grad), axis=0) # update phi self.phi = phi return loss, ua, ua_grad def train_loop(self, optimizer, lr_scheduler, dual_channel): loss_sum = 0 for lepoch in range(1, self.local_epoch + 1): loss, ua, ua_grad = self.cal_parameters(dual_channel) optimizer.zero_grad() ua.backward(ua_grad) optimizer.step() loss_sum += loss if lr_scheduler: lr_scheduler.step() loss_sum /= self.local_epoch logger.info(f"loss: {loss_sum}") def val_loop(self, dual_channel, global_epoch: int = 0): logger.info("val_loop start") self.model.eval() labels = [] for batch_idx, [y_batch] in enumerate(self.val_dataloader): labels.append(y_batch.numpy()) ub = dual_channel.recv() predict_score = torch.matmul(ub, self.phi.T) val_predicts = torch.sigmoid(predict_score) labels: np.ndarray = np.concatenate(labels, axis=0) val_predicts = np.array(val_predicts > 0.5, dtype=np.int32) metrics_output = CommonMetrics._calc_metrics( metrics=self.metrics, labels=labels, val_predicts=val_predicts, lossfunc_name=None, loss=None, dataset_type="val", ) CommonMetrics.save_metric_csv( metrics_output=metrics_output, output_config=self.common_config.output, global_epoch=global_epoch, local_epoch=None, dataset_type="val", ) def fit(self): logger.info("Transfer logistic regression training start") dual_channel = DualChannel( name="transfer_logistic_regression_channel", ids=FedConfig.get_trainer()+[FedConfig.node_id] ) optimizer = list(self.optimizer.values())[0] lr_scheduler = list(self.lr_scheduler.values())[0] if self.lr_scheduler.values() else None for epoch in range(1, self.global_epoch + 1): self.model.train() logger.info(f"trainer's global epoch {epoch}/{self.global_epoch} start...") self.train_loop(optimizer, lr_scheduler, dual_channel) self.val_loop(dual_channel, global_epoch=epoch) ModelIO.save_torch_model( state_dict=self.model.state_dict(), save_dir=self.save_dir, model_name=self.save_model_name, meta_dict={"phi": self.phi} )
6,285
38.534591
118
py
XFL
XFL-master/python/algorithm/framework/transfer/logistic_regression/common.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import pandas as pd import numpy as np import torch import torch.nn as nn from torch.utils.data import DataLoader, TensorDataset from pathlib import Path from ..base import BaseTrainer from common.utils.logger import logger from common.utils.model_io import ModelIO from common.checker.x_types import All from common.utils.config_sync import ConfigSynchronizer class TransferLogisticRegression(nn.Module): def __init__(self, input_dim: int, output_dim: int, bias: bool = False): super().__init__() self.linear = torch.nn.Linear(input_dim, output_dim, bias=bias) def forward(self, x): return self.linear(x) class Common(BaseTrainer): def __init__(self, train_conf: dict, label: bool = False): sync_rule = { "model_info": { "config": All() }, "train_info": { "interaction_params": All(), "train_params": All() } } train_conf = ConfigSynchronizer(train_conf).sync(sync_rule) self.label = label super().__init__(train_conf) self._parse_config() def _parse_config(self) -> None: self.save_dir = Path(self.common_config.output.get("path", "")) # interaction_params self.model_name = self.common_config.model_info.get("name") self.save_model_name = self.common_config.output.get("model", {}).get("name", {}) self.pretrain_model_path = self.common_config.input.get("pretrained_model", {}).get("path") self.pretrain_model_name = self.common_config.input.get("pretrained_model", {}).get("name") self.model_conf = self.common_config.model_info.get("config", {}) self.num_features = self.model_conf.get("num_features") self.hidden_features = self.model_conf.get("hidden_features") self.constant_k = 1 / self.hidden_features self.alpha = self.model_conf.get("alpha") self.bias = self.model_conf.get("bias", False) self.global_epoch = self.common_config.train_params.get("global_epoch") self.local_epoch = self.common_config.train_params.get("local_epoch") self.train_batch_size = self.common_config.train_params.get("train_batch_size") self.val_batch_size = self.common_config.train_params.get("val_batch_size") def _set_model(self) -> None: """ Init logistic regression model. Returns: None """ logger.info("Init model start.") self.phi = None # phi will be saved in the model_info of label_trainer model = TransferLogisticRegression( input_dim=self.num_features, output_dim=self.hidden_features, bias=self.bias ) # Load pretrained model if needed. if self.pretrain_model_path is not None and self.pretrain_model_path != "": model_info = ModelIO.load_torch_model( os.path.join(self.pretrain_model_path, self.pretrain_model_name)) state_dict = model_info["state_dict"] if "phi" in state_dict.keys(): self.phi = model_info["phi"] model.load_state_dict(state_dict) model = model.to(self.device) logger.info("Init model completed.") return model def _read_data(self, input_dataset, is_train=True): if len(input_dataset) == 0: return None conf = input_dataset[0] if conf["type"] == "csv": path = os.path.join(conf['path'], conf['name']) has_id = conf['has_id'] index_col = 0 if has_id else False train_data = pd.read_csv(path, index_col=index_col) if is_train: index_name = "overlap_index.npy" index_path = os.path.join(conf['path'], index_name) overlap_index = np.load(index_path) return train_data, overlap_index else: return train_data else: raise NotImplementedError( "Dataset load method {} does not Implemented.".format(conf["type"]) ) def _set_train_dataloader(self): self.overlap_y, self.non_overlap_y = None, None self.overlap_train_dataloader, self.non_overlap_train_dataloader = None, None train_data, overlap_index = self._read_data(self.common_config.input_trainset) self.sample_num = train_data.shape[0] overlap_train_data = train_data.loc[overlap_index] if self.label: non_overlap_index = np.array([]) for i in train_data.index: if i not in overlap_index: non_overlap_index = np.append(non_overlap_index, i) non_overlap_train_data = train_data.loc[non_overlap_index] if len(non_overlap_index) == 0: raise ValueError("There is no non-overlap data in the trainset. If non_overlap_index is empty, there is no need to use transfer learning") # init overlap_y and non_overlap_y self.overlap_y = torch.tensor( overlap_train_data.iloc[:, 0].to_numpy(), dtype=torch.float32).unsqueeze(1) self.non_overlap_y = torch.tensor( non_overlap_train_data.iloc[:, 0].to_numpy(), dtype=torch.float32).unsqueeze(1) # init train_dataloader overlap_x = torch.tensor( overlap_train_data.iloc[:, 1:].to_numpy(), dtype=torch.float32) overlap_trainset = TensorDataset(overlap_x, self.overlap_y) self.overlap_train_dataloader = DataLoader( overlap_trainset, batch_size=self.train_batch_size, shuffle=False) non_overlap_x = torch.tensor( non_overlap_train_data.iloc[:, 1:].to_numpy(), dtype=torch.float32) non_overlap_trainset = TensorDataset(non_overlap_x, self.non_overlap_y) self.non_overlap_train_dataloader = DataLoader( non_overlap_trainset, batch_size=self.train_batch_size, shuffle=False) else: # init train_dataloader overlap_x = torch.tensor(overlap_train_data.to_numpy(), dtype=torch.float32) overlap_trainset = TensorDataset(overlap_x) self.overlap_train_dataloader = DataLoader( overlap_trainset, batch_size=self.train_batch_size, shuffle=False) def _set_val_dataloader(self): self.val_dataloader = None val_data = self._read_data(self.common_config.input_valset, is_train=False) if self.label: # init val_dataloader labels = torch.tensor(val_data.iloc[:, 0].to_numpy(), dtype=torch.float32).unsqueeze(dim=-1) valset = TensorDataset(labels) self.val_dataloader = DataLoader(valset, batch_size=self.val_batch_size, shuffle=False) else: # init val_dataloader features = torch.tensor(val_data.to_numpy(), dtype=torch.float32) valset = TensorDataset(features) self.val_dataloader = DataLoader(valset, batch_size=self.val_batch_size, shuffle=False)
7,724
42.644068
154
py
XFL
XFL-master/python/algorithm/framework/transfer/logistic_regression/trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import torch from common.communication.gRPC.python.channel import DualChannel from common.utils.logger import logger from service.fed_config import FedConfig from .common import Common from common.utils.model_io import ModelIO class TransferLogisticRegressionTrainer(Common): def __init__(self, train_conf: dict): super().__init__(train_conf, label=False) def cal_ub(self, dataloader): ub = [] for batch_idx, [x_batch] in enumerate(dataloader): x_batch = x_batch.to(self.device) ub_batch = self.model(x_batch) # [batch_size, hidden_features] ub.append(ub_batch) ub = torch.concat(ub, axis=0) return ub def cal_parameters(self, dual_channel): ub = self.cal_ub(self.overlap_train_dataloader) # [overlap_size, hidden_features] # calculate 3 components will be sent to the trainer ub_ex = ub.unsqueeze(1) ub_2 = torch.matmul(ub.unsqueeze(2), ub_ex) # [overlap_size, hidden_features, hidden_features] comp_ub = -self.constant_k * ub # [overlap_size, hidden_features] # exchange components overlap_y_2_phi_2, overlap_y_phi, comp_ua = dual_channel.recv() dual_channel.send((ub, ub_2, comp_ub)) # compute gradients to excute backward ub_overlap_y_2_phi_2 = torch.matmul(ub_ex, overlap_y_2_phi_2) l1_grad_b = ub_overlap_y_2_phi_2.squeeze(1) + overlap_y_phi ub_grad = self.alpha * l1_grad_b + comp_ua return ub, ub_grad def fit(self): logger.info("Transfer logistic regression training start") dual_channel = DualChannel( name="transfer_logistic_regression_channel", ids=FedConfig.get_label_trainer()+[FedConfig.node_id] ) optimizer = list(self.optimizer.values())[0] lr_scheduler = list(self.lr_scheduler.values())[0] if self.lr_scheduler.values() else None for epoch in range(1, self.global_epoch + 1): self.model.train() logger.info(f"trainer's global epoch {epoch}/{self.global_epoch} start...") self.train_loop(optimizer, lr_scheduler, dual_channel) self.val_loop(dual_channel) ModelIO.save_torch_model( state_dict=self.model.state_dict(), save_dir=self.save_dir, model_name=self.save_model_name ) def train_loop(self, optimizer, lr_scheduler, dual_channel): for lepoch in range(1, self.local_epoch + 1): ub, ub_grad = self.cal_parameters(dual_channel) optimizer.zero_grad() ub.backward(ub_grad) optimizer.step() if lr_scheduler: lr_scheduler.step() def val_loop(self, dual_channel): self.model.eval() ub = self.cal_ub(self.val_dataloader) dual_channel.send(ub)
3,462
36.236559
102
py
XFL
XFL-master/python/algorithm/framework/vertical/__init__.py
0
0
0
py
XFL
XFL-master/python/algorithm/framework/vertical/vertical_model_base.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os from pathlib import Path import pandas as pd import numpy as np from common.evaluation.metrics import BiClsMetric, DecisionTable, RegressionMetric, ThresholdCutter, LiftGainCalculator from common.utils.config_parser import TrainConfigParser from common.utils.logger import logger from sklearn.metrics import roc_curve import json from sklearn.metrics import precision_recall_curve from algorithm.core.tree.feature_importance import FeatureImportance from algorithm.core.tree.feature_importance import FeatureImportance class VerticalModelBase(TrainConfigParser): def __init__(self, train_conf: dict, label: bool = False): super().__init__(train_conf) self._parse_config() self.train_conf = train_conf self.label = label self.train_ks_metrics = None self.val_ks_metrics = None def _parse_config(self) -> None: # output_path # self.save_dir = Path(self.output.get("model", {}).get("path", "")) self.save_dir = Path(self.output.get("path", "")) self.metric_dir = self.save_dir # params self.lossfunc_conifg = self.train_params.get("lossfunc", {}) self.metric_config = self.train_params.get("metric", {}) # interaction_params self.echo_training_metrics = self.interaction_params.get( "echo_training_metrics", False) self.write_training_prediction = self.interaction_params.get( "write_training_prediction", False) self.write_validation_prediction = self.interaction_params.get( "write_validation_prediction", False) # if self.output.get("metrics"): # self.metric_path = Path(self.output["metrics"].get("path")) # else: # self.metric_path = self.save_dir # # params # self.lossfunc_conifg = self.train_params.get("lossfunc_config") # self.metric_config = self.train_params.get("metric_config") # # interaction_params # self.echo_training_metrics = self.interaction_params.get("echo_training_metrics") # self.write_training_prediction = self.interaction_params.get("write_training_prediction") # self.write_validation_prediction = self.interaction_params.get("write_validation_prediction") def _calc_metrics(self, y, p, epoch, stage="train", loss={}): if stage == "train" and not self.echo_training_metrics: return if not os.path.exists(self.metric_dir): os.makedirs(self.metric_dir) # output_file = os.path.join( # self.metric_path, "{}_metrics.csv".format(stage)) output_file = os.path.join( self.metric_dir, self.output.get("metric_" + stage)["name"]) if self.model_info["name"] not in ["vertical_linear_regression", "vertical_poisson_regression"]: if loss: evaluate = BiClsMetric(epoch, output_file, self.metric_config) else: evaluate = BiClsMetric( epoch, output_file, self.metric_config, self.lossfunc_conifg) else: evaluate = RegressionMetric(epoch, output_file, self.metric_config) evaluate.calc_metrics(y, p) for key, value in loss.items(): evaluate.metrics[key] = value if self.model_info["name"] not in ["vertical_linear_regression", "vertical_poisson_regression"]: evaluate.save() else: evaluate.save(evaluate.metrics) if "decision_table" in self.metric_config: dt = DecisionTable(self.metric_config["decision_table"]) dt.fit(y, p) dt.save(os.path.join(self.metric_dir, self.output.get( "decision_table_" + stage)["name"])) # dt.save(os.path.join(self.metric_path, # "{}_decision_table.csv".format(stage))) logger.info("{} {}".format(stage, evaluate)) return evaluate.metrics def _write_loss(self, train_loss, val_loss, epoch): # prepare write path try: file_path = Path( self.metric_dir, self.train_conf['output']['plot_loss']['name'] ) except: file_path = Path( self.metric_dir, f"{self.model_info['name']}_plot_loss.json" ) if file_path.is_file(): with open(file_path, "r") as file: prev_loss_list = json.load(file) else: prev_loss_list = [] prev_loss_list.append( {'epoch': epoch, 'loss': train_loss, 'period': 'train'}) prev_loss_list.append( {'epoch': epoch, 'loss': val_loss, 'period': 'val'}) with open(file_path, "w") as out_fp: json.dump(prev_loss_list, out_fp) def _write_prediction(self, y, p, idx=None, epoch=None, final=False, stage="train"): if stage == "train" and not self.write_training_prediction: return elif stage == "val" and not self.write_validation_prediction: return elif stage not in ("train", "val"): raise ValueError("stage must be 'train' or 'val'.") if not os.path.exists(self.save_dir): os.makedirs(self.save_dir) if final: # file_name = os.path.join(self.save_dir, "predicted_probabilities_{}.csv".format(stage)) file_name = os.path.join( self.save_dir, self.output.get("prediction_" + stage)["name"]) else: # file_name = os.path.join(self.save_dir, "predicted_probabilities_{}.epoch_{}".format(stage, epoch)) file_name = self.output.get("prediction_" + stage)["name"] file_name_list = file_name.split(".") # file_name_list.insert(-1, '_epoch_'+str(epoch)) file_name_list[-2] += '_epoch_' + str(epoch) file_name = '.'.join(file_name_list) file_name = os.path.join(self.save_dir, file_name) if idx is None: df = pd.DataFrame({"pred": p, "label": y}) df = df.reset_index().rename(columns={"index": "id"}) df.to_csv(file_name, header=True, index=False, float_format='%.6g') else: df = pd.DataFrame({'id': idx, "pred": p, "label": y}) df.to_csv(file_name, header=True, index=False, float_format='%.6g') def _write_plot_data(self): pass @staticmethod def __prune_data_size(data: np.array, n_cuts: int): size = data.size cuts = np.linspace(0, 1, n_cuts + 1) index_list = [int(size * cut) for cut in cuts] if index_list[-1] >= size: index_list = index_list[:-1] return data[index_list] def _write_roc_data(self, train_label, train_pred, val_label=None, val_pred=None, n_cuts=1000) -> None: # prepare write path try: file_path = Path( self.metric_dir, self.train_conf['output']['plot_roc']['name'] ) except: file_path = Path( self.metric_dir, f"{self.model_info['name']}_plot_roc.json" ) try: train_fpr, train_tpr, _ = roc_curve(train_label, train_pred) except: logger.error("Could not calculate train roc curve") return # pruning size = train_fpr.size if size > n_cuts: logger.info( f"Too much points in training roc curve. Cutting to {n_cuts} points") train_fpr = self.__prune_data_size(train_fpr, n_cuts) train_tpr = self.__prune_data_size(train_tpr, n_cuts) # add to roc_list roc_list = [] for fpr, tpr in zip(train_fpr, train_tpr): roc_list.append( {"fpr": round(fpr, 6), "tpr": round(tpr, 6), "period": "train"}) # If val data exists if not val_label is None: try: val_fpr, val_tpr, _ = roc_curve(val_label, val_pred) except: logger.error("Could not calculate val roc curve") with open(file_path, "w") as file: json.dump(roc_list, file) return # pruning if val_fpr.size > n_cuts: logger.info( f"Too much points in validation roc curve. Cutting to {n_cuts} points") val_fpr = self.__prune_data_size(val_fpr, n_cuts) val_tpr = self.__prune_data_size(val_tpr, n_cuts) for fpr, tpr in zip(val_fpr, val_tpr): roc_list.append( {"fpr": round(fpr, 6), "tpr": round(tpr, 6), "period": "val"}) # Sort logger.info("Sorting roc list") roc_list = sorted(roc_list, key=lambda el: el['fpr']) logger.info("Writing roc to file") with open(file_path, "w") as file: json.dump(roc_list, file) return def _write_ks_data(self, train_label, train_pred, val_label=None, val_pred=None, n_cuts=1000) -> None: # Setup file path try: file_path = Path( self.metric_dir, self.train_conf['output']['plot_ks']['name'] ) except: file_path = Path( self.metric_dir, f"{self.model_info['name']}_plot_ks.json" ) ks_list = [] tc = ThresholdCutter() # Train if self.train_ks_metrics is None: tc.sim_cut_by_value(train_label, train_pred) train_ks_df = pd.DataFrame(tc.metrics) train_bst_threshold = tc.bst_threshold train_bst_score = tc.bst_score else: logger.info(f"Using calculated train ks") train_ks_df = pd.DataFrame(self.train_ks_metrics) train_bst_threshold = self.train_ks_bst_threshold train_bst_score = self.train_ks_bst_score # pruning if train_ks_df.shape[0] > n_cuts: logger.info( f"Pruning training ks data. Before pruning: {train_ks_df.shape[0]}") prune_train_ks_df = pd.DataFrame() prune_train_ks_df['threshold'] = self.__prune_data_size( train_ks_df['threshold'].values, n_cuts) prune_train_ks_df['tpr'] = self.__prune_data_size( train_ks_df['tpr'].values, n_cuts) prune_train_ks_df['fpr'] = self.__prune_data_size( train_ks_df['fpr'].values, n_cuts) train_ks_df = prune_train_ks_df logger.info( f"After pruning, training ks data: {train_ks_df.shape[0]}") for _, row in train_ks_df.iterrows(): ks_list.append( {"thres": round(row['threshold'], 6), "value": round(row['tpr'], 6), "type_period": "tpr_train"}) ks_list.append( {"thres": round(row['threshold'], 6), "value": round(row['fpr'], 6), "type_period": "fpr_train"}) ks_list.append( {"thres": round(row['threshold'], 6), "value": round(row['ks'], 6), "type_period": "ks_train"}) # Val if not val_label is None: if self.val_ks_metrics is None: tc.sim_cut_by_value(val_label, val_pred) val_ks_df = pd.DataFrame(tc.metrics) val_bst_threshold = tc.bst_threshold val_bst_score = tc.bst_score else: logger.info(f"Using calculated val ks") val_ks_df = pd.DataFrame(self.val_ks_metrics) val_bst_threshold = self.val_ks_bst_threshold val_bst_score = self.val_ks_bst_score # pruning if val_ks_df.shape[0] > n_cuts: logger.info( f"Pruning val ks data. Before pruning: {val_ks_df.shape[0]}") prune_val_ks_df = pd.DataFrame() prune_val_ks_df['threshold'] = self.__prune_data_size( val_ks_df['threshold'].values, n_cuts) prune_val_ks_df['tpr'] = self.__prune_data_size( val_ks_df['tpr'].values, n_cuts) prune_val_ks_df['fpr'] = self.__prune_data_size( val_ks_df['fpr'].values, n_cuts) val_ks_df = prune_val_ks_df logger.info( f"After pruning, val ks data: {val_ks_df.shape[0]}") for _, row in val_ks_df.iterrows(): ks_list.append( {"thres": round(row['threshold'], 6), "value": round(row['tpr'], 6), "type_period": "tpr_val"}) ks_list.append( {"thres": round(row['threshold'], 6), "value": round(row['fpr'], 6), "type_period": "fpr_val"}) ks_list.append( {"thres": round(row['threshold'], 6), "value": round(row['ks'], 6), "type_period": "ks_val"}) # Sort lists logger.info("Sorting ks list") ks_list = sorted(ks_list, key=lambda el: el["thres"]) logger.info("Writing ks to file") with open(file_path, "w") as file: json.dump(ks_list, file) return def _write_lift_gain_data(self, train_label, train_pred, val_label=None, val_pred=None) -> None: # Setup file path try: lift_file_path = Path( self.metric_dir, self.train_conf['output']['plot_lift']['name'] ) except: lift_file_path = Path( self.metric_dir, f"{self.model_info['name']}_plot_lift.json" ) try: gain_file_path = Path( self.metric_dir, self.train_conf['output']['plot_gain']['name'] ) except: gain_file_path = Path( self.metric_dir, f"{self.model_info['name']}_plot_gain.json" ) lift_gain_cal = LiftGainCalculator() lift_gain_cal.cal_lift_gain(train_label, train_pred) train_lift_gain_df = lift_gain_cal.metrics train_lift_gain_df = train_lift_gain_df.query("lift.notnull()") # cut_list = [] gain_list = [] lift_list = [] for _, row in train_lift_gain_df.iterrows(): gain_list.append( { "bin_val": round(row['percentage_data'], 6), "gain": round(row['cum_gain'], 6), "period": "train" } ) lift_list.append( { "bin_val": round(row['percentage_data'], 6), "lift": round(row['lift'], 6), "period": "train" } ) logger.info(f"Training lift point number: {len(lift_list)}") if not val_label is None: lift_gain_cal.cal_lift_gain(val_label, val_pred) val_lift_gain_df = lift_gain_cal.metrics val_lift_gain_df = val_lift_gain_df.query("lift.notnull()") for _, row in val_lift_gain_df.iterrows(): gain_list.append( { "bin_val": round(row['percentage_data'], 6), "gain": round(row['cum_gain'], 6), "period": "val" } ) lift_list.append( { "bin_val": round(row['percentage_data'], 6), "lift": round(row['lift'], 6), "period": "val" } ) logger.info(f"Val lift point number: {len(lift_list)}") # Sort values by horizontal axis logger.info("Sorting gain and list lists") gain_list = sorted(gain_list, key=lambda el: el["bin_val"]) lift_list = sorted(lift_list, key=lambda el: el["bin_val"]) with open(lift_file_path, "w") as lift_file: json.dump(lift_list, lift_file) with open(gain_file_path, "w") as gain_file: json.dump(gain_list, gain_file) return def _write_pr_data(self, train_label, train_pred, val_label=None, val_pred=None, n_cuts=1000) -> None: # Set up file path try: pr_curve_path = Path( self.metric_dir, self.train_conf['output']['plot_precision_recall']['name'] ) except: pr_curve_path = Path( self.metric_dir, f"{self.model_info['name']}_plot_pr_curve.json" ) pr_list = [] # Add train pr data precision, recall, thres = precision_recall_curve( train_label, train_pred) # pruning if precision.size > n_cuts: logger.info(f"Too much points in training pr curve, pruning") precision = self.__prune_data_size(precision, n_cuts) recall = self.__prune_data_size(recall, n_cuts) for pr, rc in zip(precision, recall): pr_list.append( { "recall": round(rc, 6), "precision": round(pr, 6), "period": "train" } ) # Add val pr data if not val_label is None: precision, recall, thres = precision_recall_curve( val_label, val_pred) # pruning if precision.size > n_cuts: logger.info("Too much points in validation pr curve, pruning") precision = self.__prune_data_size(precision, n_cuts) recall = self.__prune_data_size(recall, n_cuts) for pr, rc in zip(precision, recall): pr_list.append( { "recall": round(rc, 6), "precision": round(pr, 6), "period": "val" } ) # Sort logger.info("Sorting pr list") pr_list = sorted(pr_list, key=lambda el: el["recall"]) # Write file with open(pr_curve_path, "w") as pr_file: json.dump(pr_list, pr_file) return def _write_feature_importance(self): # Get importance file path try: feature_importance_file_path = Path( self.metric_dir, self.train_conf['output']['plot_feature_importance']['name'] ) except: feature_importance_file_path = Path( self.metric_dir, f"{self.model_info['name']}_plot_feature_importance.json" ) logger.info(f"Feature importances: {self.feature_importances_}") feature_importance_list = [] # Get and normalize feature importance try: normalizer = np.sum([_.get() for _ in self.feature_importances_.values()]) except: normalizer = np.sum( [abs(_) for _ in self.feature_importances_.values()]) for k, v in sorted(self.feature_importances_.items(), key=lambda d: d[1], reverse=True): feature_name = "_".join(map(str, k)) if isinstance(v, FeatureImportance): feature_importance = v.get() / normalizer else: feature_importance = v / normalizer feature_importance_list.append( { "feature": feature_name, "importance": round(feature_importance, 6) } ) feature_importance_list = sorted( feature_importance_list, key=lambda d: abs(d['importance']), reverse=True ) # Write file with open(feature_importance_file_path, "w") as feature_importance_file: json.dump(feature_importance_list, feature_importance_file)
20,479
38.384615
119
py
XFL
XFL-master/python/algorithm/framework/vertical/pearson/base.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import random import string from pathlib import Path from typing import List import pandas as pd from algorithm.core.encryption_param import PlainParam, get_encryption_param from common.crypto.paillier.utils import get_core_num from common.utils.config_parser import TrainConfigParser from common.utils.logger import logger class VerticalPearsonBase(TrainConfigParser): def __init__(self, train_conf: dict, label: bool = False): super().__init__(train_conf) self.label = label self.max_num_cores = 1 self.sample_size = None self.encryption = "plain" self.encryption_param = PlainParam() self._init_config() self._init_data() self._local_summary = {} self._summary = {} self.num_embed = 10 def _init_data(self): logger.info("init data loader.") if not self.input_trainset: return None input_info = self.input_trainset[0] file_path = str(Path(input_info.get("path"), input_info.get("name"))) type_ = input_info.get("type", "None") if input_info.get("has_id", True): self.index_col = input_info.get("index_col", 'id') else: self.index_col = None if input_info.get("has_label", False): label_name = input_info.get("label_name", 'y') self.label = True else: label_name = None if type_ == "csv": if self.computing_engine == "local": df = pd.read_csv(file_path, index_col=self.index_col) else: raise NotImplementedError("Computing engine {} is not supported.".format(self.computing_engine)) else: raise NotImplementedError("Dataset type {} is not supported.".format(type_)) if self.label: feature_cols = [_ for _ in df.columns if _ != label_name] self.train_features = df[feature_cols] if label_name: self.train_label = df[label_name] else: self.train_label = None else: self.train_features = df self.train_ids = df.index def _init_config(self): params = self.train_info.get("train_params") self.col_index = params.get("col_index", -1) self.col_names = params.get("col_names", '') encryption_params = params.get("encryption", {"plain": {}}) self.encryption = list(encryption_params.keys())[0] encryption_param = encryption_params[self.encryption] self.encryption_param = get_encryption_param(self.encryption, encryption_param) self.sample_size = params.get("sample_size", None) if self.encryption == "paillier" and self.encryption_param.parallelize_on: self.max_num_cores = get_core_num(params.get("max_num_cores", 999)) else: self.max_num_cores = 1 def _select_columns(self): if self.col_index == -1: return self.train_features elif isinstance(self.col_index, list): feature_start_index = 0 if self.index_col is not None: feature_start_index += 1 if self.label: feature_start_index += 1 feature_names = self.train_features.columns.to_list() select_feature_cols = [feature_names[_ - feature_start_index] for _ in self.col_index] if self.col_names: for f in self.col_names.split(','): if f not in select_feature_cols: select_feature_cols.append(f) select_feature_cols.sort(key=lambda d: feature_names.index(d)) return self.train_features[select_feature_cols] else: raise ValueError("col_index must be -1 or a list of int.") @staticmethod def standardize(x): mu = x.mean() sigma = x.std() if sigma > 0: return (x - mu) / sigma else: return x - mu @staticmethod def string_encryption(str_list: List[str]): ret = {} for s in str_list: ret[s] = ''.join(random.sample(string.ascii_letters + string.digits, 16)) return ret
4,149
31.677165
100
py
XFL
XFL-master/python/algorithm/framework/vertical/pearson/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import pickle from pathlib import Path import numpy as np from common.checker.matcher import get_matched_config from common.checker.x_types import All from service.fed_config import FedConfig from service.fed_node import FedNode from service.fed_control import ProgressCalculator from .base import VerticalPearsonBase from common.communication.gRPC.python.channel import BroadcastChannel, DualChannel from common.utils.logger import logger from algorithm.core.encryption_param import PlainParam, PaillierParam from common.crypto.paillier.paillier import Paillier from algorithm.core.paillier_acceleration import embed, umbed class VerticalPearsonLabelTrainer(VerticalPearsonBase): def __init__(self, train_conf: dict): """ Args: train_conf: """ self.sync_channel = BroadcastChannel(name="sync") self._sync_config(train_conf) super().__init__(train_conf, label=False) self.feature_mapping = dict() self.channels = dict() self.channels["encryption_context"] = BroadcastChannel(name="encryption_context") self.channels["sample_idx_com"] = BroadcastChannel(name="sample_idx_com") self.channels["trainer_corr_com"] = dict() self.channels["trainer_feature_com"] = dict() self.node_id = FedNode.node_id for party_id in FedConfig.get_trainer(): self.channels["trainer_feature_com"][party_id] = DualChannel( name="trainer_feature_com_" + party_id, ids=[self.node_id, party_id] ) for party_id in FedConfig.get_trainer(): self.channels["trainer_corr_com"][party_id] = DualChannel( name="trainer_corr_com_" + party_id, ids=[self.node_id, party_id] ) if isinstance(self.encryption_param, (PlainParam, type(None))): self.private_context = None elif isinstance(self.encryption_param, PaillierParam): self.private_context = Paillier.context( self.encryption_param.key_bit_size, self.encryption_param.djn_on ) self.public_context = self.private_context.to_public() self.channels["encryption_context"].broadcast(self.public_context.serialize(), use_pickle=False) else: raise TypeError(f"Encryption param type {type(self.encryption_param)} not valid.") def _sync_config(self, config): sync_rule = { "train_info": All() } config_to_sync = get_matched_config(config, sync_rule) self.sync_channel.broadcast(config_to_sync) def fit(self): logger.info("vertical pearson label trainer start.") data = self._select_columns() if self.sample_size is not None and self.sample_size < len(self.train_ids): logger.info("sampled %d data." % self.sample_size) sample_ids = np.random.choice(np.arange(len(self.train_ids)), self.sample_size) self.channels["sample_idx_com"].broadcast(sample_ids) data = data.iloc[sample_ids] data = data.apply(self.standardize) n = len(data) feature_names = data.columns.to_list() self.feature_mapping = self.string_encryption(feature_names) local_corr = np.dot(data.T, data) local_corr /= n self._summary["corr"] = dict() self._summary["corr"][(self.node_id, self.node_id)] = local_corr self._summary["features"] = { self.node_id: feature_names } self._summary["num_features"] = { self.node_id: len(feature_names) } remote_corr = dict() for party_id in FedConfig.get_trainer(): remote_corr[party_id] = [] if len(feature_names): self.channels["trainer_feature_com"][party_id].send(False) else: self.channels["trainer_feature_com"][party_id].send(True) if isinstance(self.encryption_param, (PlainParam, type(None))): for idx, f in enumerate(feature_names): for party_id in FedConfig.get_trainer(): if f != feature_names[-1]: self.channels["trainer_feature_com"][party_id].send( (data[f].to_numpy(), False)) else: self.channels["trainer_feature_com"][party_id].send( (data[f].to_numpy(), True)) logger.info("label trainer encrypted %d features." % (idx + 1)) for party_id in FedConfig.get_trainer(): corr_mat = self.channels["trainer_feature_com"][party_id].recv() remote_corr[party_id].append(corr_mat) for party_id in FedConfig.get_trainer(): remote_corr[party_id] = np.array(remote_corr[party_id]) / n if isinstance(self.encryption_param, PaillierParam): for i in range(0, len(feature_names), self.num_embed): embedded_data = embed(data.iloc[:, i:(i + self.num_embed)].to_numpy().T) encrypted_data = Paillier.encrypt( context=self.private_context, data=embedded_data, obfuscation=True, num_cores=self.max_num_cores ) for party_id in FedConfig.get_trainer(): if i + self.num_embed >= len(feature_names): self.channels["trainer_feature_com"][party_id].send( (encrypted_data, True)) logger.info("label trainer encrypted %d features." % len(feature_names)) else: self.channels["trainer_feature_com"][party_id].send( (encrypted_data, False)) logger.info("label trainer encrypted %d features." % (i + self.num_embed)) for party_id in FedConfig.get_trainer(): emb_enc_corr_mat = self.channels["trainer_feature_com"][party_id].recv() emb_corr_mat = Paillier.decrypt( self.private_context, emb_enc_corr_mat, num_cores=self.max_num_cores, out_origin=True ) result = [] if i + self.num_embed >= len(feature_names): umbed_num = len(feature_names) - i else: umbed_num = self.num_embed for r in umbed(emb_corr_mat, umbed_num): result.append(r) remote_corr[party_id].append(np.array(result) / (10 ** self.encryption_param.precision)) for party_id in FedConfig.get_trainer(): if remote_corr[party_id]: remote_corr[party_id] = np.concatenate(remote_corr[party_id]) / n logger.info("label trainer encrypted all features.") for party_id in FedConfig.get_trainer(): remote_features = self.channels["trainer_feature_com"][party_id].recv() self._summary["features"][party_id] = remote_features self._summary["num_features"][party_id] = len(remote_features) self._summary["corr"][(self.node_id, party_id)] = remote_corr[party_id] logger.info("label trainer get remote correlation matrices.") logger.info("get correlation matrix between trainers.") for party_id in FedConfig.get_trainer(): other_summary = self.channels["trainer_corr_com"][party_id].recv() for k, v in other_summary.items(): if k[0] != k[1]: if self.encryption == "plain": self._summary["corr"][k] = v[0].T / n elif self.encryption == "paillier": remote_corr = [] for emb_enc_corr_mat, pack_num in zip(v[0], v[1]): emb_corr_mat = Paillier.decrypt( self.private_context, emb_enc_corr_mat, num_cores=self.max_num_cores, out_origin=True ) corr_mat = [] for r in umbed(emb_corr_mat, pack_num): corr_mat.append(r) corr_mat = np.array(corr_mat) / (10 ** self.encryption_param.precision) remote_corr.append(corr_mat) if remote_corr: self._summary["corr"][k] = np.concatenate(remote_corr).T / n else: self._summary["corr"][k] = [] else: self._summary["corr"][k] = v # update the progress of 100 to show the training is finished ProgressCalculator.finish_progress() self.save() def save(self): save_dir = str(Path(self.output.get("path"))) if not os.path.exists(save_dir): os.makedirs(save_dir) model_name = self.output.get("corr")["name"] model_path = Path(save_dir, model_name) summary = self.merge_summary() with open(model_path, 'wb') as f: pickle.dump(summary, f) logger.info("model saved as: {}.".format(model_path)) def merge_summary(self): parties = FedConfig.get_label_trainer() + FedConfig.get_trainer() cor_mat = [] features = [] sources = [] num_feature = 0 for i in range(len(parties)): row = [] for j in range(len(parties)): if (parties[i], parties[j]) in self._summary["corr"]: corr_mat = self._summary["corr"][(parties[i], parties[j])] if min(np.array(corr_mat).shape) > 0: row.append(corr_mat) else: row.append(np.zeros( (self._summary["num_features"][parties[i]], self._summary["num_features"][parties[j]]))) else: corr_mat = self._summary["corr"][(parties[j], parties[i])] if min(np.array(corr_mat).shape) > 0: row.append(corr_mat.T) else: row.append(np.zeros( (self._summary["num_features"][parties[i]], self._summary["num_features"][parties[j]]))) cor_mat.append(np.concatenate(row, axis=1)) features.extend(self._summary["features"][parties[i]]) sources.extend([parties[i]] * self._summary["num_features"][parties[i]]) num_feature += self._summary["num_features"][parties[i]] cor_mat = np.concatenate(cor_mat, axis=0) ret = { "corr": cor_mat, "features": features, "feature_source": sources, "num_features": self._summary["num_features"] } return ret
9,458
37.45122
99
py
XFL
XFL-master/python/algorithm/framework/vertical/pearson/__init__.py
0
0
0
py
XFL
XFL-master/python/algorithm/framework/vertical/pearson/trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os from pathlib import Path import pickle import numpy as np from common.utils.utils import update_dict from service.fed_config import FedConfig from service.fed_node import FedNode from .base import VerticalPearsonBase from algorithm.core.encryption_param import PlainParam, PaillierParam from common.crypto.paillier.paillier import Paillier from common.communication.gRPC.python.channel import BroadcastChannel, DualChannel from common.utils.logger import logger from algorithm.core.paillier_acceleration import embed class VerticalPearsonTrainer(VerticalPearsonBase): def __init__(self, train_conf: dict): """ Args: train_conf: """ self.sync_channel = BroadcastChannel(name="sync") conf = self._sync_config() update_dict(train_conf, conf) super().__init__(train_conf, label=False) self.channels = {} self.node_id = FedNode.node_id self.channels = dict() self.channels["encryption_context"] = BroadcastChannel(name="encryption_context") self.channels["sample_idx_com"] = BroadcastChannel(name="sample_idx_com") self.channels["trainer_feature_com"] = DualChannel( name="trainer_feature_com_" + self.node_id, ids=[self.node_id] + FedConfig.get_label_trainer() ) self.channels["trainer_corr_com"] = DualChannel( name="trainer_corr_com_" + self.node_id, ids=[self.node_id] + FedConfig.get_label_trainer() ) self.channels["trainer_com"] = dict() self.trainers = FedConfig.get_trainer() for trainer_id in self.trainers: if trainer_id != self.node_id: if self.node_id not in self.trainers or trainer_id not in self.trainers: continue elif self.trainers.index(trainer_id) < self.trainers.index(self.node_id): self.channels["trainer_com"][trainer_id] = DualChannel( name="trainer_com_" + trainer_id + "_" + self.node_id, ids=[self.node_id, trainer_id] ) else: self.channels["trainer_com"][trainer_id] = DualChannel( name="trainer_com_" + self.node_id + "_" + trainer_id, ids=[self.node_id, trainer_id] ) if isinstance(self.encryption_param, (PlainParam, type(None))): self.public_context = None elif isinstance(self.encryption_param, PaillierParam): self.public_context = self.channels["encryption_context"].recv(use_pickle=False) self.public_context = Paillier.context_from(self.public_context) else: raise TypeError(f"Encryption param type {type(self.encryption_param)} not valid.") self.feature_mapping = dict() def _sync_config(self): config = self.sync_channel.recv() return config def fit(self): logger.info("vertical pearson trainer start.") data = self._select_columns() if self.sample_size is not None and self.sample_size < len(data): logger.info("sampled %d data." % self.sample_size) sample_ids = self.channels["sample_idx_com"].recv() data = data.iloc[sample_ids] data = data.apply(self.standardize) n = len(data) feature_names = data.columns.to_list() self.feature_mapping = self.string_encryption(feature_names) local_corr = np.dot(data.T, data) local_corr /= n self._local_summary["corr"] = local_corr self._local_summary["features"] = feature_names self._local_summary["num_features"] = { self.node_id: len(feature_names) } self._local_summary["feature_source"] = [self.node_id] * len(feature_names) self._local_summary["feature_mapping"] = self.feature_mapping self._summary[(self.node_id, self.node_id)] = local_corr feature_flag = self.channels["trainer_feature_com"].recv() # remote_corr = pd.DataFrame() j = 0 local_mat = np.array([]) if isinstance(self.encryption_param, PlainParam): local_mat = data.to_numpy() elif isinstance(self.encryption_param, PaillierParam): local_mat = np.array(data.to_numpy() * 10 ** self.encryption_param.precision, dtype=int) while not feature_flag: other, feature_flag = self.channels["trainer_feature_com"].recv() remote_corr = np.dot(local_mat.T, other) self.channels["trainer_feature_com"].send(remote_corr) logger.info("trainer calculated {} feature from label_trainer.".format(j + 1)) j += 1 self.channels["trainer_feature_com"].send([self.feature_mapping[f] for f in feature_names]) for i in range(self.trainers.index(self.node_id)): trainer_id = self.trainers[i] flag = self.channels["trainer_com"][trainer_id].recv() j = 0 corr_mat = [] pack_nums = [] while not flag: other, pack_num, flag = self.channels["trainer_com"][trainer_id].recv() remote_corr = np.dot(local_mat.T, other) self.channels["trainer_com"][trainer_id].send(True) corr_mat.append(remote_corr) pack_nums.append(pack_num) logger.info("trainer {} calculated {} feature from trainer {}.".format(self.node_id, j + 1, trainer_id)) j += 1 self._summary[(self.node_id, trainer_id)] = (np.array(corr_mat), pack_nums) for j in range(self.trainers.index(self.node_id) + 1, len(self.trainers)): trainer_id = self.trainers[j] if len(feature_names): self.channels["trainer_com"][trainer_id].send(False) else: self.channels["trainer_com"][trainer_id].send(True) if isinstance(self.encryption_param, (PlainParam, type(None))): cnt = 0 for f in feature_names: encrypted_data = data[f].to_numpy() if f != feature_names[-1]: self.channels["trainer_com"][trainer_id].send((encrypted_data, 1, False)) else: self.channels["trainer_com"][trainer_id].send((encrypted_data, 1, True)) cnt += 1 logger.info( "trainer {} encrypted {} features to trainer {}.".format( self.node_id, cnt, trainer_id)) self.channels["trainer_com"][trainer_id].recv() elif isinstance(self.encryption_param, PaillierParam): for i in range(0, len(feature_names), self.num_embed): embedded_data = embed(data.iloc[:, i:(i + self.num_embed)].to_numpy().T) encrypted_data = Paillier.encrypt( context=self.public_context, data=embedded_data, obfuscation=True, num_cores=self.max_num_cores ) if i + self.num_embed >= len(feature_names): self.channels["trainer_com"][trainer_id].send( (encrypted_data, len(feature_names) - i, True)) logger.info( "trainer {} encrypted {} features to trainer {}.".format( self.node_id, len(feature_names), trainer_id)) else: self.channels["trainer_com"][trainer_id].send( (encrypted_data, self.num_embed, False)) logger.info( "trainer {} encrypted {} features to trainer {}.".format( self.node_id, i + self.num_embed, trainer_id)) self.channels["trainer_com"][trainer_id].recv() self.channels["trainer_corr_com"].send(self._summary) self.save() def save(self): save_dir = str(Path(self.output.get("path"))) if not os.path.exists(save_dir): os.makedirs(save_dir) model_name = self.output.get("corr")["name"] model_path = Path(save_dir, model_name) with open(model_path, 'wb') as f: pickle.dump(self._local_summary, f) logger.info("model saved as: {}.".format(model_path))
7,614
36.885572
108
py
XFL
XFL-master/python/algorithm/framework/vertical/feature_selection/base.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import json import pickle from pathlib import Path import pandas as pd from common.utils.config_parser import TrainConfigParser from common.utils.logger import logger from common.model.python.feature_model_pb2 import WOEModel from google.protobuf import json_format from service.fed_job import FedJob from service.fed_node import FedNode class VerticalFeatureSelectionBase(TrainConfigParser): def __init__(self, train_conf: dict, label: bool = False): super().__init__(train_conf) self.label = label self.feature_info = [] self.iv_result = None self.corr_result = None self.transform_stages = [] self.train_id, self.train_label, self.train_features = None, None, None self.val_id, self.val_label, self.val_features = None, None, None self._init_config() self._load_feature_info() self._init_data() def _init_data(self): logger.info("init data loader.") if not self.input_trainset: return None self.train_id, self.train_label, self.train_features = self._load_data(self.input_trainset[0]) self.transform_stages.append("train") if not self.input_valset: return None self.val_id, self.val_label, self.val_features = self._load_data(self.input_valset[0]) self.transform_stages.append("valid") def _load_data(self, input_info): file_path = str(Path(input_info.get("path"), input_info.get("name"))) type_ = input_info.get("type", "None") if input_info.get("has_id", True): index_col = input_info.get("index_col", 'id') else: index_col = None if input_info.get("has_label", False): label_name = input_info.get("label_name", 'y') self.label = True else: label_name = None if type_ == "csv": if self.computing_engine == "local": df = pd.read_csv(file_path, index_col=index_col) # elif self.computing_engine == "spark": # df = ps.read_csv(file_path, index_col=index_col) else: raise NotImplementedError("Computing engine {} is not supported.".format(self.computing_engine)) else: raise NotImplementedError("Dataset type {} is not supported.".format(type_)) if self.label: feature_cols = [_ for _ in df.columns if _ != label_name] features = df[feature_cols] if label_name: label = df[label_name] else: label = None else: features = df label = None ids = df.index return ids, label, features def _init_config(self): params = self.train_info.get("train_params") self.filter = params.get("filter", {}) def _parse_from_iv(self, params): path = params["path"].replace("[JOB_ID]", str(FedJob.job_id)).replace("[NODE_ID]", str(FedNode.node_id)) with open(Path(path, params["name"])) as f: res = json.load(f) if self.label: res = res.get("iv") for k, v in res.items(): self.feature_info.append({ "feature_id": k, "iv": v }) return res @staticmethod def _parse_corr_result(params): path = params["path"].replace("[JOB_ID]", str(FedJob.job_id)).replace("[NODE_ID]", str(FedNode.node_id)) with open(Path(path, params["name"]), 'rb') as f: ret = pickle.load(f) return ret def _load_feature_info(self): if "iv_result" in self.input: self.iv_result = self._parse_from_iv(self.input.get("iv_result")) if "corr_result" in self.input: self.corr_result = self._parse_corr_result(self.input.get("corr_result")) def _rewrite_model(self, features): if self.input.get("model"): params = self.input.get("model") path = params["path"].replace("[JOB_ID]", str(FedJob.job_id)).replace("[NODE_ID]", str(FedNode.node_id)) file_path = Path(path, params["name"]) with open(file_path, 'rb') as f: byte_str = f.read() woe = WOEModel() woe.ParseFromString(byte_str) d = json_format.MessageToDict(woe, including_default_value_fields=True, preserving_proto_field_name=True) ret = dict() for k, v in d["feature_binning"].items(): if v["feature"] in features: ret[k] = v woe = WOEModel() json_format.ParseDict({"feature_binning": ret}, woe) with open(file_path, "wb") as f: f.write(woe.SerializeToString()) logger.info("rewrite model in {}.".format(file_path))
4,753
32.957143
107
py
XFL
XFL-master/python/algorithm/framework/vertical/feature_selection/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import pickle from pathlib import Path import pandas as pd from service.fed_control import ProgressCalculator from common.checker.matcher import get_matched_config from common.checker.x_types import All from common.communication.gRPC.python.channel import BroadcastChannel from common.utils.logger import logger from .base import VerticalFeatureSelectionBase class VerticalFeatureSelectionLabelTrainer(VerticalFeatureSelectionBase): def __init__(self, train_conf: dict): """ Args: train_conf: """ self.sync_channel = BroadcastChannel(name="sync") self._sync_config(train_conf) super().__init__(train_conf, label=True) self.progress_calculator = ProgressCalculator(len(self.filter)) self.channels = dict() self.channels["feature_id_com"] = BroadcastChannel(name="feature_id_com") def _common_filter(self, params): logger.info("common_filter") metrics = params.get("metrics", 'iv') filter_method = params.get("filter_method", "threshold") if isinstance(metrics, str): metrics = [metrics] elif isinstance(params.get("metrics"), list): pass else: raise NotImplementedError("param metrics must be a string or a list.") for metric in metrics: if filter_method == "threshold": self.feature_info = self._threshold_filter( metric, params.get("threshold", 1) ) def _sync_config(self, config): sync_rule = { "train_info": All() } config_to_sync = get_matched_config(config, sync_rule) self.sync_channel.broadcast(config_to_sync) def _threshold_filter(self, metric, threshold): ret = [] if metric == "iv": for r in self.feature_info: if r["iv"] < threshold: logger.info("filter feature {} < threshold {}".format(r["feature_id"], threshold)) continue ret.append(r) else: raise NotImplementedError("metric {} is not supported".format(metric)) return ret def _correlation_filter(self, params): logger.info("correlation_filter") sort_metric = params.get("sort_metric", 'iv') correlation_threshold = params.get("correlation_threshold", 0.1) self.channels["feature_id_com"].broadcast([_["feature_id"] for _ in self.feature_info]) mapping = {} for d in self.channels["feature_id_com"].collect(): for k, v in d.items(): mapping[k] = v corr = self.corr_result["corr"] features = self.corr_result["features"] res = [] filtered_features = [] self.feature_info.sort(key=lambda x: x[sort_metric], reverse=True) for r in self.feature_info: f = mapping.get(r["feature_id"], r["feature_id"]) if f in features: i = features.index(f) else: continue if features[i] in filtered_features: continue for f, s in zip(features, corr[i]): if f == features[i]: continue elif f in filtered_features: continue elif abs(s) > correlation_threshold: logger.info( "current feature {}, filtered feature {}.".format( features[i], f ) ) filtered_features.append(f) res.append({ "feature_id": mapping.get(r["feature_id"], r["feature_id"]), "iv": r["iv"] }) self.feature_info = res def fit(self): logger.info("feature selection label trainer start.") # iter_ is used to calculate the progress of the training iter_ = 0 for k, v in self.filter.items(): iter_ += 1 if k == "common": self._common_filter(v) elif k == "correlation": self._correlation_filter(v) else: raise NotImplementedError("method {} is not implemented.".format(k)) # calculate and update the progress of the training self.progress_calculator.cal_custom_progress(iter_) self.channels["feature_id_com"].broadcast([_["feature_id"] for _ in self.feature_info]) self.save() self.transform() ProgressCalculator.finish_progress() def transform(self): if not self.transform_stages: return None selected_features = [_["feature_id"] for _ in self.feature_info] selected_features = [_ for _ in self.train_features.columns if _ in selected_features] if "train" in self.transform_stages: self.train_features = self.train_features[selected_features] if self.train_label is not None: df = pd.concat([self.train_label, self.train_features], axis=1).set_index(self.train_id) else: df = self.train_features.set_index(self.train_id) output_train_path = Path(self.output.get("path"), self.output["trainset"].get("name")) if not os.path.exists(os.path.dirname(output_train_path)): os.makedirs(os.path.dirname(output_train_path)) df.to_csv(output_train_path, header=True, index=True, index_label="id", float_format="%.6g") if "valid" in self.transform_stages: self.val_features = self.val_features[selected_features] if self.val_label is not None: df = pd.concat([self.val_label, self.val_features], axis=1).set_index(self.val_id) else: df = self.val_features.set_index(self.val_id) output_val_path = Path(self.output.get("path"), self.output["valset"].get("name")) if not os.path.exists(os.path.dirname(output_val_path)): os.makedirs(os.path.dirname(output_val_path)) df.to_csv(output_val_path, header=True, index=True, index_label="id", float_format="%.6g") def save(self): save_dir = str(Path(self.output.get("path"))) if not os.path.exists(save_dir): os.makedirs(save_dir) model_name = self.output.get("model")["name"] model_path = Path(save_dir, model_name) features = [_["feature_id"] for _ in self.feature_info] output = { "features": features, "num_of_features": len(self.feature_info) } with open(model_path, 'wb') as f: pickle.dump(output, f) logger.info("model saved as: {}.".format(model_path)) self._rewrite_model(features)
6,269
32.174603
95
py
XFL
XFL-master/python/algorithm/framework/vertical/feature_selection/__init__.py
0
0
0
py
XFL
XFL-master/python/algorithm/framework/vertical/feature_selection/trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import pickle from pathlib import Path import pandas as pd from common.communication.gRPC.python.channel import BroadcastChannel from common.utils.logger import logger from common.utils.utils import update_dict from .base import VerticalFeatureSelectionBase class VerticalFeatureSelectionTrainer(VerticalFeatureSelectionBase): def __init__(self, train_conf: dict): """ Args: train_conf: """ self.sync_channel = BroadcastChannel(name="sync") conf = self._sync_config() update_dict(train_conf, conf) super().__init__(train_conf, label=False) self.channels = dict() self.channels["feature_id_com"] = BroadcastChannel(name="feature_id_com") self.feature_mapping = dict() def _sync_config(self): config = self.sync_channel.recv() return config def _common_filter(self, params): metrics = params.get("metrics", 'iv') if isinstance(metrics, str): metrics = [metrics] elif isinstance(params.get("metrics"), list): pass else: raise NotImplementedError("param metrics must be a string or a list.") for metric in metrics: if metric == "iv": for k, v in self.iv_result["feature_mapping"].items(): self.feature_mapping[k] = v def _correlation_filter(self): feature_id_list = self.channels["feature_id_com"].recv() mapping = {} reversed_feature_mapping = dict() for k, v in self.feature_mapping.items(): reversed_feature_mapping[v] = k for feature_id in feature_id_list: if feature_id in reversed_feature_mapping: local_feature_name = reversed_feature_mapping.get(feature_id) if local_feature_name in self.corr_result["feature_mapping"]: mapping[feature_id] = self.corr_result["feature_mapping"][local_feature_name] else: continue self.channels["feature_id_com"].send(mapping) self.feature_mapping = self.corr_result["feature_mapping"] def fit(self): logger.info("feature selection trainer start.") for k, v in self.filter.items(): if k == "common": self._common_filter(v) elif k == "correlation": self._correlation_filter() else: raise NotImplementedError("method {} is not implemented.".format(k)) remain_id_list = self.channels["feature_id_com"].recv() res = dict() for k, v in self.feature_mapping.items(): if v in remain_id_list: res[k] = v self.feature_mapping = res self.save() self.transform() def transform(self): if not self.transform_stages: return None selected_features = [_ for _ in self.feature_mapping] if "train" in self.transform_stages: self.train_features = self.train_features[selected_features] if self.train_label is not None: df = pd.concat([self.train_label, self.train_features], axis=1).set_index(self.train_id) else: df = self.train_features.set_index(self.train_id) output_train_path = Path(self.output.get("path"), self.output["trainset"].get("name")) if not os.path.exists(os.path.dirname(output_train_path)): os.makedirs(os.path.dirname(output_train_path)) df.to_csv(output_train_path, header=True, index=True, index_label="id", float_format="%.6g") if "valid" in self.transform_stages: self.val_features = self.val_features[selected_features] if self.val_label is not None: df = pd.concat([self.val_label, self.val_features], axis=1).set_index(self.val_id) else: df = self.val_features.set_index(self.val_id) output_val_path = Path(self.output.get("path"), self.output["valset"].get("name")) if not os.path.exists(os.path.dirname(output_val_path)): os.makedirs(os.path.dirname(output_val_path)) df.to_csv(output_val_path, header=True, index=True, index_label="id", float_format="%.6g") def save(self): save_dir = str(Path(self.output.get("path"))) if not os.path.exists(save_dir): os.makedirs(save_dir) model_name = self.output.get("model")["name"] model_path = Path(save_dir, model_name) features = list(self.feature_mapping.keys()) with open(model_path, 'wb') as f: pickle.dump(self.feature_mapping, f) logger.info("model saved as: {}.".format(model_path)) self._rewrite_model(features)
4,683
33.441176
95
py
XFL
XFL-master/python/algorithm/framework/vertical/poisson_regression/base.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import torch import torch.nn as nn from torch.utils.data import DataLoader, TensorDataset from algorithm.core.data_io import CsvReader from algorithm.framework.vertical.vertical_model_base import VerticalModelBase from common.utils.logger import logger from common.utils.model_preserver import ModelPreserver from service.fed_config import FedConfig class VerticalPoissonRegression(nn.Module): def __init__(self, input_dim: int, bias: bool = False): super(VerticalPoissonRegression, self).__init__() self.linear = torch.nn.Linear(input_dim, 1, bias=bias) self.linear.requires_grad_(False) def forward(self, x): return self.linear(x) class VerticalPoissonRegressionBase(VerticalModelBase): def __init__(self, train_conf: dict, label: bool = False, *args, **kwargs): """_summary_ Args: train_conf (dict): _description_ label (bool, optional): _description_. Defaults to False. """ super().__init__(train_conf) self._parse_config() self.train_conf = train_conf self.label = label self.data_dim = None self.model = None self.train_dataloader, self.eval_dataloader = None, None if FedConfig.node_id != "assist_trainer": self._init_dataloader() def _parse_config(self) -> None: super()._parse_config() self.model_name = self.model_info.get("name") self.save_model_name = self.output.get("model", {}).get("name") if not os.path.exists(self.save_dir): os.makedirs(self.save_dir) self.global_epoch = self.train_params.get("global_epoch") self.batch_size = self.train_params.get("batch_size") self.encryption_config = self.train_params.get("encryption") self.optimizer_config = self.train_params.get("optimizer") self.pretrain_model_path = self.input.get("pretrained_model", {}).get("path") self.random_seed = self.train_params.get("random_seed", None) self.early_stopping_config = self.train_params.get("early_stopping") self.save_frequency = self.interaction_params.get("save_frequency") @staticmethod def set_seed(seed): torch.manual_seed(seed) torch.cuda.manual_seed(seed) def _init_model(self, bias: bool = False) -> None: """ Init poisson regression model. Returns: None """ logger.info("Init model start.") self.model = VerticalPoissonRegression(input_dim=self.data_dim, bias=bias) # Load pretrained model if needed. if self.pretrain_model_path is not None and self.pretrain_model_path != "": checkpoint = ModelPreserver.load(os.path.join(self.pretrain_model_path, self.input.get( "pretrained_model").get("name", None))) self.model.load_state_dict(checkpoint["state_dict"]) logger.info("Init model completed.") def __load_data(self, config) -> CsvReader: config = config[0] if config["type"] == "csv": data_reader = CsvReader(path=os.path.join(config["path"], config["name"]), has_id=config["has_id"], has_label=config["has_label"]) else: raise NotImplementedError("Dataset type {} is not supported.".format(config["type"])) return data_reader def _init_data(self) -> None: if len(self.input_trainset) > 0: data: CsvReader = self.__load_data(self.input_trainset) self.train = data.features() self.train_label = data.label() self.train_ids = list(range(len(data.ids))) else: raise NotImplementedError("Trainset was not configured.") if self.label: assert len(self.train) == len(self.train_label) if len(self.input_valset) > 0: data: CsvReader = self.__load_data(self.input_valset) self.val = data.features() self.val_label = data.label() self.val_ids = list(range(len(data.ids))) if self.label: assert len(self.val) == len(self.val_label) def _init_dataloader(self) -> None: """ Load raw data. Returns: """ logger.info("Dataloader initiation start.") self._init_data() if self.label: self.train_dataloader = DataLoader( dataset=TensorDataset(torch.tensor(self.train, dtype=torch.float32), torch.unsqueeze(torch.tensor(self.train_label), dim=-1), torch.unsqueeze(torch.tensor(self.train_ids), dim=-1)), batch_size=self.batch_size, shuffle=True ) self.val_dataloader = DataLoader( dataset=TensorDataset(torch.tensor(self.val, dtype=torch.float32), torch.unsqueeze(torch.tensor(self.val_label), dim=-1), torch.unsqueeze(torch.tensor(self.val_ids), dim=-1)), batch_size=self.batch_size, shuffle=False ) self.data_dim = torch.tensor(self.train).shape[-1] logger.info("Train data shape: {}.".format(list(torch.tensor(self.train).shape))) else: self.train_dataloader = DataLoader( dataset=TensorDataset(torch.tensor(self.train, dtype=torch.float32), torch.unsqueeze(torch.tensor(self.train_ids), dim=-1)), batch_size=self.batch_size, shuffle=True ) self.val_dataloader = DataLoader( dataset=TensorDataset(torch.tensor(self.val, dtype=torch.float32), torch.unsqueeze(torch.tensor(self.val_ids), dim=-1)), batch_size=self.batch_size, shuffle=False ) self.data_dim = torch.tensor(self.train).shape[-1] logger.info("Train data shape: {}.".format(list(torch.tensor(self.train).shape))) logger.info("Dataloader initiation completed.")
6,716
41.783439
111
py
XFL
XFL-master/python/algorithm/framework/vertical/poisson_regression/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import copy import secrets import random from functools import reduce from pathlib import Path from common.checker.x_types import All import numpy as np import pandas as pd import tenseal as ts import torch from common.checker.matcher import get_matched_config from common.communication.gRPC.python.channel import BroadcastChannel, DualChannel from common.crypto.paillier.paillier import Paillier from common.utils.algo_utils import earlyStopping from common.utils.logger import logger from common.utils.model_preserver import ModelPreserver from common.utils.utils import save_model_config from service.fed_config import FedConfig from service.fed_node import FedNode from service.fed_control import ProgressCalculator from .base import VerticalPoissonRegressionBase class VerticalPoissonRegressionLabelTrainer(VerticalPoissonRegressionBase): def __init__(self, train_conf: dict, *args, **kwargs): """ Vertical Poisson Regression Args: train_conf: training parameters *args: **kwargs: """ self.sync_channel = BroadcastChannel(name="sync") self._sync_config(train_conf) super().__init__(train_conf, label=True, *args, **kwargs) if self.random_seed is None: self.random_seed = random.randint(-(1 << 32), 1 << 32) self.sync_channel.broadcast(self.random_seed) self.set_seed(self.random_seed) self.progress_calculator = ProgressCalculator(self.global_epoch, len(self.train_dataloader)) self._init_model(bias=True) self.export_conf = [{ "class_name": "VerticalPoissonRegression", "identity": self.identity, "filename": self.save_model_name, "input_dim": self.data_dim, "bias": True, "version": "1.4.0" }] self.es = earlyStopping(key=self.early_stopping_config["key"], patience=self.early_stopping_config["patience"], delta=self.early_stopping_config["delta"]) self.best_model = None self.broadcast_channel = BroadcastChannel(name="Public keys", root_id=FedConfig.get_assist_trainer()) self.dual_channels = {"intermediate_label_trainer": {}, "gradients_loss": None} self.trainers = FedConfig.get_trainer() for party_id in self.trainers: self.dual_channels["intermediate_label_trainer"][party_id] = DualChannel( name="intermediate_label_trainer_" + party_id, ids=[FedConfig.node_id, party_id]) self.dual_channels["gradients_loss"] = DualChannel(name="gradients_loss_" + FedConfig.node_id, ids=[FedConfig.get_assist_trainer()] + [FedConfig.node_id]) self.train_result = None self.val_result = None self.dual_channels["gradients_loss"].send(len(self.train_dataloader)) self.dual_channels["gradients_loss"].send(self.global_epoch) self.dual_channels["gradients_loss"].send(self.batch_size) self.encryption_method = list(self.encryption_config.keys())[0].lower() self.dual_channels["gradients_loss"].send(self.encryption_config) self.dual_channels["gradients_loss"].send(self.encryption_method) def _sync_config(self, config): sync_rule = { "train_info": All() } config_to_sync = get_matched_config(config, sync_rule) self.sync_channel.broadcast(config_to_sync) def predict(self, input_data): pred_prob_epoch, y_epoch, pred_tmp_epoch = [], [], [] for batch_idx, (x_batch, y_batch, _) in enumerate(input_data): pred_trainer_list = [] pre_tmp = self.model(x_batch) # receive intermediate results from trainers for party_id in FedConfig.get_trainer(): pred_trainer_list.append(self.dual_channels["intermediate_label_trainer"][party_id].recv( use_pickle=True)) # calculate prediction of batch and tmp_pred of batch pred_tmp_total = pre_tmp.numpy().astype(np.float32).flatten() + reduce( lambda x, y: x + y, pred_trainer_list) pred_total = np.exp(pred_tmp_total) # calculate prediction of epoch pred_prob_epoch += pred_total.tolist() pred_tmp_epoch += pred_tmp_total.tolist() y_epoch += y_batch.numpy().astype(np.float32).flatten().tolist() return y_epoch, pred_prob_epoch, pred_tmp_epoch def fit(self): self.check_data() public_context = None num_cores = -1 rng = secrets.SystemRandom() logger.info("Vertical poisson regression training start") # receive encryption key from assist trainer if self.encryption_method == "ckks": logger.info("Receive ckks public key.") public_context = self.broadcast_channel.recv(use_pickle=False) public_context = ts.context_from(public_context) logger.info("Public key received.") elif self.encryption_method == "paillier": logger.info("Receive paillier public key.") public_context = self.broadcast_channel.recv(use_pickle=False) public_context = Paillier.context_from(public_context) logger.info("Public key received.") elif self.encryption_method == "plain": pass else: raise ValueError(f"Encryption method {self.encryption_method} not supported! Valid methods are " f"'paillier', 'ckks', 'plain'.") # train for epoch in range(1, self.global_epoch + 1): loss_epoch = 0 for batch_idx, (x_batch, y_batch, _) in enumerate(self.train_dataloader): regular_loss_tmp = 0 regular_gradient_tmp = 0 enc_regular_gradient_tmp = 0 trainer_exp = None # calculate regular results if self.optimizer_config['p'] == 1: regular_loss_tmp = torch.abs(self.model.linear.weight).sum() * self.optimizer_config['alpha'] regular_gradient_tmp = self.optimizer_config['alpha'] * (torch.abs(self.model.linear.weight) / self.model.linear.weight) elif self.optimizer_config['p'] == 2: regular_loss_tmp = (self.model.linear.weight ** 2).sum() * self.optimizer_config['alpha'] / 2 regular_gradient_tmp = self.optimizer_config['alpha'] * self.model.linear.weight elif self.optimizer_config['p'] == 0: pass # receive intermediate exp results from trainers and compute total_exp logger.info("Calculate predicted exp result of all trainers.") pred_label_tmp = self.model(x_batch) pred_label_trainer = torch.exp(pred_label_tmp) if self.encryption_method == "ckks": trainer_exp = self.dual_channels["intermediate_label_trainer"][self.trainers[-1]].recv( use_pickle=False) trainer_exp = ts.ckks_vector_from(public_context, trainer_exp) elif self.encryption_method == "paillier": trainer_exp = self.dual_channels["intermediate_label_trainer"][self.trainers[-1]].recv( use_pickle=False) trainer_exp = Paillier.ciphertext_from(public_context, trainer_exp) elif self.encryption_method == "plain": trainer_exp = self.dual_channels["intermediate_label_trainer"][self.trainers[-1]].recv() total_exp = trainer_exp * pred_label_trainer.numpy().astype(np.float32).flatten() # receive immediate results from trainers logger.info("Calculate predicted result of all trainers.") if self.encryption_method == "ckks": total_sum = ts.ckks_vector(public_context, pred_label_tmp.numpy().astype(np.float32).flatten()) else: total_sum = pred_label_tmp.numpy().astype(np.float32).flatten() for party_id in self.trainers: if self.encryption_method == "ckks": total_sum = total_sum + ts.ckks_vector_from(public_context, self.dual_channels[ "intermediate_label_trainer"][party_id].recv(use_pickle=False)) trainer_regular_loss = ts.ckks_vector_from(public_context, self.dual_channels[ "intermediate_label_trainer"][party_id].recv(use_pickle=False)) elif self.encryption_method == "paillier": total_sum = total_sum + Paillier.ciphertext_from(public_context, self.dual_channels[ "intermediate_label_trainer"][party_id].recv(use_pickle=False)) trainer_regular_loss = Paillier.ciphertext_from(public_context, self.dual_channels[ "intermediate_label_trainer"][party_id].recv(use_pickle=False)) elif self.encryption_method == "plain": total_sum = total_sum + self.dual_channels["intermediate_label_trainer"][party_id].recv() trainer_regular_loss = self.dual_channels["intermediate_label_trainer"][party_id].recv() # calculate total loss logger.info("Calculate total loss.") enc_loss = total_exp - total_sum * y_batch.numpy().astype(np.float32).flatten() if self.encryption_method == "ckks": regular_loss_tmp = ts.ckks_vector(public_context, regular_loss_tmp.numpy().astype(np.float32).flatten()) else: regular_loss_tmp = regular_loss_tmp.numpy().astype(np.float32).flatten() enc_loss_batch = enc_loss + regular_loss_tmp + trainer_regular_loss # send total loss to assist_trainer logger.info("Send encrypted total loss to assist_trainer.") if self.encryption_method == "ckks": self.dual_channels["gradients_loss"].send(enc_loss_batch.serialize(), use_pickle=False) elif self.encryption_method == "paillier": self.dual_channels["gradients_loss"].send(Paillier.serialize(enc_loss_batch), use_pickle=False) elif self.encryption_method == "plain": self.dual_channels["gradients_loss"].send(enc_loss_batch) # receive decrypted loss from assist_trainer logger.info("Receive decrypted total loss from assist_trainer.") loss_batch = self.dual_channels["gradients_loss"].recv() loss_batch = loss_batch / x_batch.shape[0] logger.info("Loss of {} batch is {}".format(batch_idx, loss_batch)) loss_epoch += loss_batch * x_batch.shape[0] # calculate intermediate result d logger.info("Calculate intermediate result d.") enc_y = None if self.encryption_method == "ckks": enc_y = ts.ckks_vector(public_context, y_batch.numpy().astype(np.float32).flatten()) elif self.encryption_method == "paillier": enc_y = Paillier.encrypt(public_context, y_batch.numpy().astype(np.float32).flatten(), precision=self.encryption_config[self.encryption_method][ "precision"], obfuscation=True, num_cores=num_cores) elif self.encryption_method == "plain": enc_y = y_batch.numpy().astype(np.float32).flatten() enc_d = total_exp - enc_y # send intermediate result d to trainers logger.info("Send intermediate result d to trainers.") for party_id in self.trainers: if self.encryption_method == "ckks": self.dual_channels["intermediate_label_trainer"][party_id].send(enc_d.serialize(), use_pickle=False) elif self.encryption_method == "paillier": self.dual_channels["intermediate_label_trainer"][party_id].send(Paillier.serialize(enc_d), use_pickle=False) elif self.encryption_method == "plain": self.dual_channels["intermediate_label_trainer"][party_id].send(enc_d) # calculate gradient for label_trainer logger.info("Calculate gradients for label_trainer.") if self.encryption_method == "ckks": enc_regular_gradient_tmp = ts.ckks_vector(public_context, regular_gradient_tmp.numpy().astype(np.float32).flatten()) elif self.encryption_method == "paillier": enc_regular_gradient_tmp = Paillier.encrypt( public_context, regular_gradient_tmp.numpy().astype(np.float32).flatten(), precision=self.encryption_config[self.encryption_method]["precision"], obfuscation=True, num_cores=num_cores) elif self.encryption_method == "plain": enc_regular_gradient_tmp = regular_gradient_tmp.numpy().astype(np.float32).flatten() if self.encryption_method == "ckks": gradient_label_trainer_w = enc_d.matmul(x_batch.numpy()) + enc_regular_gradient_tmp else: gradient_label_trainer_w = np.matmul(enc_d.reshape(1, len(enc_d)), x_batch.numpy() ) + enc_regular_gradient_tmp gradient_label_trainer_b = enc_d if self.encryption_method == "ckks": # add noise to encrypted gradients and send to assist_trainer logger.info("Calculate noised gradients for label_trainer.") noise = np.array([rng.randint(1 << 24, 1 << 26) - (1 << 25) for _ in range(x_batch.shape[1])], dtype=np.float32) noise /= 100000 noise_b = np.array([rng.randint(1 << 24, 1 << 26) - (1 << 25) for _ in range(x_batch.shape[0])], dtype=np.float32) noise_b /= 100000 noised_gradient_label_trainer_w = gradient_label_trainer_w + noise noised_gradient_label_trainer_b = gradient_label_trainer_b + noise_b logger.info("Send noised gradient to assist_trainer.") self.dual_channels["gradients_loss"].send(noised_gradient_label_trainer_w.serialize(), use_pickle=False) self.dual_channels["gradients_loss"].send(noised_gradient_label_trainer_b.serialize(), use_pickle=False) # receive decrypted gradient from assist_trainer logger.info("Receive decrypted gradient from assist_trainer.") noised_decrypt_gradient = self.dual_channels["gradients_loss"].recv() noised_decrypt_gradient_label_trainer_w = noised_decrypt_gradient["noised_gradient_label_trainer_w"] noised_decrypt_gradient_label_trainer_b = noised_decrypt_gradient["noised_gradient_label_trainer_b"] gradient_label_trainer_w = noised_decrypt_gradient_label_trainer_w - noise gradient_label_trainer_b = noised_decrypt_gradient_label_trainer_b - np.sum(noise_b) elif self.encryption_method == "paillier": # add noise to encrypted gradients and send to assist_trainer logger.info("Calculate noised gradients for label_trainer.") noise = np.array([rng.randint(1 << 24, 1 << 26) - (1 << 25) for _ in range(x_batch.shape[1])], dtype=np.float32) noise /= 100000 noise_b = np.array([rng.randint(1 << 24, 1 << 26) - (1 << 25) for _ in range(x_batch.shape[0])], dtype=np.float32) noise_b /= 100000 noised_gradient_label_trainer_w = gradient_label_trainer_w + noise noised_gradient_label_trainer_b = gradient_label_trainer_b + noise_b logger.info("Send noised gradient to assist_trainer.") self.dual_channels["gradients_loss"].send(Paillier.serialize(noised_gradient_label_trainer_w), use_pickle=False) self.dual_channels["gradients_loss"].send(Paillier.serialize(noised_gradient_label_trainer_b), use_pickle=False) # receive decrypted gradient from assist_trainer logger.info("Receive decrypted gradient from assist_trainer.") noised_decrypt_gradient = self.dual_channels["gradients_loss"].recv() noised_decrypt_gradient_label_trainer_w = noised_decrypt_gradient["noised_gradient_label_trainer_w"] noised_decrypt_gradient_label_trainer_b = noised_decrypt_gradient["noised_gradient_label_trainer_b"] gradient_label_trainer_w = noised_decrypt_gradient_label_trainer_w - noise gradient_label_trainer_b = noised_decrypt_gradient_label_trainer_b - np.sum(noise_b) elif self.encryption_method == "plain": gradient_label_trainer_b = gradient_label_trainer_b.sum() # update w and b of label_trainer gradient_label_trainer_w = gradient_label_trainer_w / x_batch.shape[0] gradient_label_trainer_b = gradient_label_trainer_b / x_batch.shape[0] logger.info("Update weights of label trainer.") self.model.linear.weight -= (torch.FloatTensor(gradient_label_trainer_w) * self.optimizer_config["lr"]) self.model.linear.bias -= (gradient_label_trainer_b * self.optimizer_config["lr"]) # calculate and update the progress of the training self.progress_calculator.cal_custom_progress(epoch, batch_idx+1) loss_epoch = loss_epoch * (1 / len(self.train)) logger.info("Loss of {} epoch is {}".format(epoch, loss_epoch)) # predict train and val results for metrics logger.info("Predict train weights of label trainer.") self.train_result = self.predict(self.train_dataloader) loss_train_met = {"loss": loss_epoch} self._calc_metrics(np.array(self.train_result[1], dtype=float), np.array(self.train_result[0]), epoch, stage="train", loss=loss_train_met) logger.info("Predict val weights of label trainer.") self.val_result = self.predict(self.val_dataloader) loss_val = np.mean( np.array(self.val_result[1]) - np.array(self.val_result[0]) * np.array(self.val_result[2])) loss_val_met = {"loss": loss_val} # no regular val_metrics = self._calc_metrics(np.array(self.val_result[1], dtype=float), np.array(self.val_result[0]), epoch, stage="val", loss=loss_val_met) # early stopping val_metrics["loss"] = - val_metrics["loss"] if self.early_stopping_config["patience"] > 0: early_stop_flag, best_model_flag = self.es(val_metrics) else: early_stop_flag, best_model_flag = False, True # update best model if best_model_flag: self.best_model = copy.deepcopy(self.model) # send flags to trainers for party_id in FedConfig.get_trainer(): self.dual_channels["intermediate_label_trainer"][party_id].send( [early_stop_flag, best_model_flag, self.early_stopping_config["patience"]], use_pickle=True) # if need to save results by epoch if self.save_frequency > 0 and epoch % self.save_frequency == 0: ModelPreserver.save(save_dir=self.save_dir, model_name=self.save_model_name, state_dict=self.model.state_dict(), epoch=epoch) # if early stopping, break if early_stop_flag: # update the progress of 100 to show the training is finished ProgressCalculator.finish_progress() break # save model for infer save_model_config(stage_model_config=self.export_conf, save_path=Path(self.save_dir)) # if not early stopping, save probabilities and model self._save_prob() ModelPreserver.save(save_dir=self.save_dir, model_name=self.save_model_name, state_dict=self.best_model.state_dict(), final=True) # calculate feature importance self._save_feature_importance(self.dual_channels) def _save_prob(self): if self.interaction_params.get("write_training_prediction"): self._write_prediction(self.train_result[1], self.train_result[0], self.train_ids, stage="train", final=True) if self.interaction_params.get("write_validation_prediction"): self._write_prediction(self.val_result[1], self.val_result[0], self.val_ids, stage="val", final=True) def check_data(self): dim_channel = BroadcastChannel(name="check_data_com", ids=[FedConfig.node_id] + FedConfig.get_trainer()) n = self.data_dim dims = dim_channel.collect() for dim in dims: n += dim if n <= 0: raise ValueError("Number of the feature is zero. Stop training.") def _save_feature_importance(self, channel): res = {"owner_id": [], "fid": [], "importance": []} other_weight_list = [] for party_id in FedConfig.get_trainer(): other_weight_list.append(channel["intermediate_label_trainer"][party_id].recv(use_pickle=True)) for (owner_id, weights) in other_weight_list: for fid, weight in enumerate(weights): res["owner_id"].append(owner_id) res["fid"].append(fid) res["importance"].append(float(weight)) for fid, weight in enumerate(self.best_model.state_dict()["linear.weight"][0]): res["owner_id"].append(FedNode.node_id) res["fid"].append(fid) res["importance"].append(float(weight)) res = pd.DataFrame(res).sort_values(by="importance", key=lambda col: np.abs(col), ascending=False) res.to_csv(Path(self.save_dir, self.output["feature_importance"]["name"]), header=True, index=False, float_format="%.6g")
24,198
58.166259
120
py
XFL
XFL-master/python/algorithm/framework/vertical/poisson_regression/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import numpy as np import tenseal as ts from common.communication.gRPC.python.channel import BroadcastChannel, DualChannel from common.crypto.paillier.paillier import Paillier from common.utils.logger import logger from service.fed_config import FedConfig class VerticalPoissonRegressionAssistTrainer(object): def __init__(self, *args, **kwargs): """[summary] assist_trainer """ self.broadcast_channel = BroadcastChannel(name="Public keys", root_id=FedConfig.get_assist_trainer()) self.dual_channels = {"gradients_loss": {}} self.party_id_list = FedConfig.get_label_trainer() + FedConfig.get_trainer() for party_id in self.party_id_list: self.dual_channels["gradients_loss"][party_id] = DualChannel(name="gradients_loss_" + party_id, ids=[FedConfig.node_id, party_id]) self.batch_num = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv() self.global_epoch = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv() self.batch_size = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv() self.encryption_config = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv() self.encryption_method = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv() self.private_context = None self.public_context = None # send encryption key to all parties if self.encryption_method == "ckks": self.private_context = ts.context( ts.SCHEME_TYPE.CKKS, poly_modulus_degree=self.encryption_config[self.encryption_method]["poly_modulus_degree"], coeff_mod_bit_sizes=self.encryption_config[self.encryption_method]["coeff_mod_bit_sizes"] ) self.private_context.generate_galois_keys() self.private_context.generate_relin_keys() self.private_context.global_scale = 1 << self.encryption_config[self.encryption_method][ "global_scale_bit_size"] serialized_public_context = self.private_context.serialize( save_public_key=True, save_secret_key=False, save_galois_keys=True, save_relin_keys=True ) logger.info("Broadcast ckks public keys.") self.public_context_ser = serialized_public_context self.broadcast_channel.broadcast(self.public_context_ser, use_pickle=False) logger.info("Broadcast completed.") elif self.encryption_method == "paillier": self.num_cores = -1 if self.encryption_config[self.encryption_method]["parallelize_on"] else 1 self.private_context = Paillier.context(self.encryption_config[self.encryption_method]["key_bit_size"], djn_on=self.encryption_config[self.encryption_method]["djn_on"]) logger.info("Broadcast paillier public keys.") self.public_context_ser = self.private_context.to_public().serialize() self.broadcast_channel.broadcast(self.public_context_ser, use_pickle=False) logger.info("Broadcast completed.") elif self.encryption_method == "plain": pass else: raise ValueError(f"Encryption method {self.encryption_method} not supported! Valid methods are 'paillier', " f"'ckks', 'plain'.") def fit(self): """ train model Model parameters need to be updated before fitting. """ # send encryption key to all parties if self.encryption_method in ["ckks", "paillier"]: logger.info("Broadcast ckks public keys.") self.broadcast_channel.broadcast(self.public_context_ser, use_pickle=False) logger.info("Broadcast completed.") # train for epoch in range(1, self.global_epoch + 1): for batch_idx in range(self.batch_num): # receive and decrypt total encrypted loss and send to label_trainer logger.info("Receive and decrypted total loss and send back to label_trainer.") if self.encryption_method == "ckks": enc_loss_batch = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv( use_pickle=False) decrypted_loss_batch = ts.ckks_vector_from(self.private_context, enc_loss_batch).decrypt() elif self.encryption_method == "paillier": enc_loss_batch = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv( use_pickle=False) decrypted_loss_batch = Paillier.decrypt(self.private_context, Paillier.ciphertext_from( None, enc_loss_batch), dtype='float', num_cores=self.num_cores) elif self.encryption_method == "plain": enc_loss_batch = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv() decrypted_loss_batch = enc_loss_batch decrypted_loss_batch = np.sum(decrypted_loss_batch) self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].send(decrypted_loss_batch) logger.info( "Loss of {} batch {} epoch is {}".format(batch_idx, epoch, decrypted_loss_batch / self.batch_size)) # receive encrypted noised gradients from other parties and decrypt and send back to other parties if self.encryption_method == "ckks" or self.encryption_method == "paillier": # trainer for party_id in FedConfig.get_trainer(): en_noised_gradient_trainer_w = self.dual_channels["gradients_loss"][party_id].recv( use_pickle=False) if self.encryption_method == "ckks": noised_gradient_trainer_w = ts.ckks_vector_from(self.private_context, en_noised_gradient_trainer_w).decrypt() elif self.encryption_method == "paillier": noised_gradient_trainer_w = Paillier.decrypt(self.private_context, Paillier.ciphertext_from( None, en_noised_gradient_trainer_w), dtype='float', num_cores=self.num_cores) self.dual_channels["gradients_loss"][party_id].send(noised_gradient_trainer_w) # label_trainer en_noised_gradient_label_trainer_w = self.dual_channels["gradients_loss"][ FedConfig.get_label_trainer()[0]].recv(use_pickle=False) en_noised_gradient_label_trainer_b = self.dual_channels["gradients_loss"][ FedConfig.get_label_trainer()[0]].recv(use_pickle=False) if self.encryption_method == "ckks": noised_gradient_label_trainer_w = ts.ckks_vector_from( self.private_context, en_noised_gradient_label_trainer_w).decrypt() noised_gradient_label_trainer_b = ts.ckks_vector_from( self.private_context, en_noised_gradient_label_trainer_b).decrypt() elif self.encryption_method == "paillier": noised_gradient_label_trainer_w = Paillier.decrypt( self.private_context, Paillier.ciphertext_from(None, en_noised_gradient_label_trainer_w), dtype='float', num_cores=self.num_cores) noised_gradient_label_trainer_b = Paillier.decrypt( self.private_context, Paillier.ciphertext_from(None, en_noised_gradient_label_trainer_b), dtype='float', num_cores=self.num_cores) # calculate sum of gradient b noised_gradient_label_trainer_b = np.sum(noised_gradient_label_trainer_b) grad_send = {"noised_gradient_label_trainer_w": noised_gradient_label_trainer_w, "noised_gradient_label_trainer_b": noised_gradient_label_trainer_b} self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].send(grad_send)
9,167
61.794521
120
py
XFL
XFL-master/python/algorithm/framework/vertical/poisson_regression/__init__.py
0
0
0
py
XFL
XFL-master/python/algorithm/framework/vertical/poisson_regression/trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import copy import secrets from pathlib import Path import numpy as np import tenseal as ts import torch from common.utils.utils import update_dict from common.communication.gRPC.python.channel import BroadcastChannel, DualChannel from common.crypto.paillier.paillier import Paillier from common.utils.logger import logger from service.fed_config import FedConfig from service.fed_node import FedNode from common.utils.model_preserver import ModelPreserver from common.utils.utils import save_model_config from .base import VerticalPoissonRegressionBase class VerticalPoissonRegressionTrainer(VerticalPoissonRegressionBase): def __init__(self, train_conf: dict, *args, **kwargs): """[summary] Args: train_conf (dict): [description] """ self.sync_channel = BroadcastChannel(name="sync") conf = self._sync_config() update_dict(train_conf, conf) super().__init__(train_conf, label=False, *args, **kwargs) self._init_model() self.export_conf = [{ "class_name": "VerticalPoissonRegression", "identity": self.identity, "filename": self.save_model_name, "input_dim": self.data_dim, "bias": False }] if self.random_seed is None: self.random_seed = self.sync_channel.recv() self.set_seed(self.random_seed) self.best_model = None self.node_id = FedConfig.node_id self.broadcast_channel = BroadcastChannel(name="Public keys", root_id=FedConfig.get_assist_trainer()) self.trainers = FedConfig.get_trainer() self.dual_trainers = {} if len(FedConfig.get_trainer()) > 1: for trainer in self.trainers: if trainer != self.node_id: self.dual_trainers[trainer] = DualChannel(name="Trainer exchange", ids=[trainer, self.node_id]) self.dual_channels = { "intermediate_label_trainer": DualChannel(name="intermediate_label_trainer_" + self.node_id, ids=FedConfig.get_label_trainer() + [self.node_id]), "gradients_loss": DualChannel(name="gradients_loss_" + self.node_id, ids=[FedConfig.get_assist_trainer()] + [self.node_id]) } def _sync_config(self): config = self.sync_channel.recv() return config def predict(self, input_data): for batch_idx, x_batch in enumerate(input_data): # calculate prediction of batch pred_trainer = self.model(x_batch[0]) # send to label_trainer self.dual_channels["intermediate_label_trainer"].send(pred_trainer.numpy().astype(np.float32).flatten(), use_pickle=True) def fit(self): """ train model Model parameters need to be updated before fitting. """ self.check_data() num_cores = -1 encryption_config = self.encryption_config encryption_method = list(self.encryption_config.keys())[0].lower() logger.info("Vertical poisson regression training start") # receive encryption key from assist trainer public_context = None if encryption_method == "ckks": logger.info("Receive ckks public key.") public_context = self.broadcast_channel.recv(use_pickle=False) public_context = ts.context_from(public_context) logger.info("Public key received.") elif encryption_method == "paillier": logger.info("Receive paillier public key.") public_context = self.broadcast_channel.recv(use_pickle=False) public_context = Paillier.context_from(public_context) logger.info("Public key received.") elif encryption_method == "plain": pass else: raise ValueError( f"Encryption method {encryption_method} not supported! Valid methods are 'paillier', 'ckks', 'plain'.") rng = secrets.SystemRandom() # train for epoch in range(1, self.global_epoch + 1): for batch_idx, x_batch in enumerate(self.train_dataloader): regular_loss_tmp = 0 regular_gradient_tmp = 0 enc_regular_gradient_tmp = 0 # calculate regular results if self.optimizer_config['p'] == 1: regular_loss_tmp = torch.abs(self.model.linear.weight).sum() * self.optimizer_config['alpha'] regular_gradient_tmp = self.optimizer_config['alpha'] * (torch.abs(self.model.linear.weight) / self.model.linear.weight) elif self.optimizer_config['p'] == 2: regular_loss_tmp = (self.model.linear.weight ** 2).sum() * self.optimizer_config['alpha'] / 2 regular_gradient_tmp = self.optimizer_config['alpha'] * self.model.linear.weight elif self.optimizer_config['p'] == 0: pass regular_loss_tmp = regular_loss_tmp.numpy().astype(np.float32).flatten() # compute multiplication of exp of all trainers pred_tmp = self.model(x_batch[0]) pred_trainer = torch.exp(pred_tmp).numpy().astype(np.float32).flatten() # if node_id is the first trainer of trainers, encrypt the result if self.node_id == self.trainers[0]: if encryption_method == "ckks": enc_pred_trainer = ts.ckks_vector(public_context, pred_trainer) elif encryption_method == "paillier": enc_pred_trainer = Paillier.encrypt(public_context, pred_trainer, precision=encryption_config[encryption_method]["precision"], obfuscation=True, num_cores=num_cores) elif encryption_method == "plain": enc_pred_trainer = pred_trainer else: pass # encrypt regular loss of trainers if encryption_method == "ckks": enc_regular_loss = ts.ckks_vector(public_context, regular_loss_tmp) elif encryption_method == "paillier": enc_regular_loss = Paillier.encrypt(public_context, regular_loss_tmp, precision=encryption_config[encryption_method]["precision"], obfuscation=True, num_cores=num_cores) elif encryption_method == "plain": enc_regular_loss = regular_loss_tmp # communicate and calculate multiplication of trainers logger.info("Calculate predicted exp result of all trainers.") if self.node_id == self.trainers[0]: if len(self.trainers) > 1: if encryption_method == "ckks": self.dual_trainers[self.trainers[1]].send(enc_pred_trainer.serialize(), use_pickle=False) elif encryption_method == "paillier": self.dual_trainers[self.trainers[1]].send(Paillier.serialize(enc_pred_trainer), use_pickle=False) elif encryption_method == "plain": self.dual_trainers[self.trainers[1]].send(enc_pred_trainer) elif len(self.trainers) == 1: if encryption_method == "ckks": self.dual_channels["intermediate_label_trainer"].send(enc_pred_trainer.serialize(), use_pickle=False) elif encryption_method == "paillier": self.dual_channels["intermediate_label_trainer"].send(Paillier.serialize(enc_pred_trainer), use_pickle=False) elif encryption_method == "plain": self.dual_channels["intermediate_label_trainer"].send(enc_pred_trainer) elif len(self.trainers) > 1: train_ind = self.trainers.index(self.node_id) pred_recv = self.dual_trainers[self.trainers[train_ind - 1]] if train_ind != len(self.trainers) - 1: pred_send = self.dual_trainers[self.trainers[train_ind + 1]] if encryption_method == "ckks": pre_pred = ts.ckks_vector_from(public_context, pred_recv.recv(use_pickle=False)) multi_pred = pre_pred * pred_trainer pred_send.send(multi_pred.serialize(), use_pickle=False) elif encryption_method == "paillier": pre_pred = Paillier.ciphertext_from(public_context, pred_recv.recv(use_pickle=False)) multi_pred = pre_pred * pred_trainer pred_send.send(Paillier.serialize(multi_pred), use_pickle=False) elif encryption_method == "plain": pre_pred = pred_recv.recv() multi_pred = pre_pred * pred_trainer pred_send.send(multi_pred) elif train_ind == len(self.trainers) - 1: if encryption_method == "ckks": pre_pred = ts.ckks_vector_from(public_context, pred_recv.recv(use_pickle=False)) multi_pred = pre_pred * pred_trainer self.dual_channels["intermediate_label_trainer"].send(multi_pred.serialize(), use_pickle=False) elif encryption_method == "paillier": pre_pred = Paillier.ciphertext_from(public_context, pred_recv.recv(use_pickle=False)) multi_pred = pre_pred * pred_trainer self.dual_channels["intermediate_label_trainer"].send(Paillier.serialize(multi_pred), use_pickle=False) elif encryption_method == "plain": pre_pred = pred_recv.recv() multi_pred = pre_pred * pred_trainer self.dual_channels["intermediate_label_trainer"].send(multi_pred) # send intermediate results to label trainer. logger.info("Send intermediate result to label trainer.") if encryption_method == "ckks": enc_pred_tmp = ts.ckks_vector(public_context, pred_tmp.numpy().astype(np.float32).flatten()) self.dual_channels["intermediate_label_trainer"].send(enc_pred_tmp.serialize(), use_pickle=False) self.dual_channels["intermediate_label_trainer"].send(enc_regular_loss.serialize(), use_pickle=False) elif encryption_method == "paillier": enc_pred_tmp = Paillier.encrypt(public_context, pred_tmp.numpy().astype(np.float32).flatten(), precision=encryption_config[encryption_method]["precision"], obfuscation=True, num_cores=num_cores) self.dual_channels["intermediate_label_trainer"].send(Paillier.serialize(enc_pred_tmp), use_pickle=False) self.dual_channels["intermediate_label_trainer"].send(Paillier.serialize(enc_regular_loss), use_pickle=False) elif encryption_method == "plain": enc_pred_tmp = pred_tmp.numpy().astype(np.float32).flatten() self.dual_channels["intermediate_label_trainer"].send(enc_pred_tmp, use_pickle=True) self.dual_channels["intermediate_label_trainer"].send(enc_regular_loss, use_pickle=True) # receive intermediate result d from label_trainer logger.info("Receive intermediate result d from label_trainer.") if encryption_method == "ckks": enc_d = self.dual_channels["intermediate_label_trainer"].recv(use_pickle=False) enc_d = ts.ckks_vector_from(public_context, enc_d) elif encryption_method == "paillier": enc_d = self.dual_channels["intermediate_label_trainer"].recv(use_pickle=False) enc_d = Paillier.ciphertext_from(public_context, enc_d) elif encryption_method == "plain": enc_d = self.dual_channels["intermediate_label_trainer"].recv() # calculate gradient for trainer and send to assist_trainer logger.info("Calculate gradients for trainer.") if encryption_method == "ckks": enc_regular_gradient_tmp = ts.ckks_vector(public_context, regular_gradient_tmp.numpy().astype(np.float32).flatten()) elif encryption_method == "paillier": enc_regular_gradient_tmp = Paillier.encrypt( public_context, regular_gradient_tmp.numpy().astype(np.float32).flatten(), precision=encryption_config[encryption_method]["precision"], obfuscation=True, num_cores=num_cores) elif encryption_method == "plain": enc_regular_gradient_tmp = regular_gradient_tmp.numpy().astype(np.float32).flatten() if encryption_method == "ckks": gradient_trainer_w = enc_d.matmul(x_batch[0].numpy()) + enc_regular_gradient_tmp else: gradient_trainer_w = np.matmul(enc_d.reshape(1, len(enc_d)), x_batch[0].numpy() ) + enc_regular_gradient_tmp # add noise to encrypted gradients and send to assist_trainer if encryption_method == "ckks": logger.info("Calculate noised gradient for trainer.") noise = np.array([rng.randint(1 << 24, 1 << 26) - (1 << 25) for _ in range(x_batch[0].shape[1])], dtype=np.float32) noise /= 100000 noised_gradient_trainer_w = gradient_trainer_w + noise logger.info("Send noised gradient to assist_trainer.") self.dual_channels["gradients_loss"].send(noised_gradient_trainer_w.serialize(), use_pickle=False) # receive decrypted gradient from assist_trainer logger.info("Receive decrypted gradient from assist_trainer.") noised_gradient_trainer_w = self.dual_channels["gradients_loss"].recv() gradient_trainer_w = noised_gradient_trainer_w - noise elif encryption_method == "paillier": logger.info("Calculate noised gradient for trainer.") noise = np.array([rng.randint(1 << 24, 1 << 26) - (1 << 25) for _ in range(x_batch[0].shape[1])], dtype=np.float32) noise /= 100000 noised_gradient_trainer_w = gradient_trainer_w + noise logger.info("Send noised gradient to assist_trainer.") self.dual_channels["gradients_loss"].send(Paillier.serialize(noised_gradient_trainer_w), use_pickle=False) # receive decrypted gradient from assist_trainer logger.info("Receive decrypted gradient from assist_trainer.") noised_gradient_trainer_w = self.dual_channels["gradients_loss"].recv() gradient_trainer_w = noised_gradient_trainer_w - noise # gradient_trainer_w = torch.FloatTensor(gradient_trainer_w).unsqueeze(-1) # update w and b of trainer gradient_trainer_w = gradient_trainer_w / x_batch[0].shape[0] logger.info("Update weights of trainer.") self.model.linear.weight -= (torch.FloatTensor(gradient_trainer_w) * self.optimizer_config["lr"]) # predict train and val for metrics logger.info("Predict train weights of trainer.") self.predict(self.train_dataloader) logger.info("Predict val weights of trainer.") self.predict(self.val_dataloader) # receive flags early_stop_flag, best_model_flag, patient = self.dual_channels["intermediate_label_trainer"].recv( use_pickle=True) # update best model if best_model_flag: self.best_model = copy.deepcopy(self.model) # if need to save results by epoch if self.save_frequency > 0 and epoch % self.save_frequency == 0: ModelPreserver.save(save_dir=self.save_dir, model_name=self.save_model_name, state_dict=self.model.state_dict(), epoch=epoch) # if early stopping, break if early_stop_flag: break # save model for infer save_model_config(stage_model_config=self.export_conf, save_path=Path(self.save_dir)) # if not early stopping, save model ModelPreserver.save(save_dir=self.save_dir, model_name=self.save_model_name, state_dict=self.best_model.state_dict(), final=True) # send w to label trainer self._save_feature_importance(self.dual_channels["intermediate_label_trainer"]) def _save_feature_importance(self, channel): channel.send((FedNode.node_id, self.best_model.state_dict()["linear.weight"][0])) def check_data(self): dim_channel = BroadcastChannel(name="check_data_com", ids=[self.node_id] + FedConfig.get_trainer()) dim_channel.send(self.data_dim)
19,818
57.463127
120
py
XFL
XFL-master/python/algorithm/framework/vertical/sampler/base.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os from pathlib import Path from algorithm.core.data_io import CsvReader from common.communication.gRPC.python.channel import BroadcastChannel from common.utils.config_parser import TrainConfigParser from common.utils.logger import logger class VerticalSamplerBase(TrainConfigParser): def __init__(self, train_conf: dict): super().__init__(train_conf) self._init_data() self.broadcast_channel = BroadcastChannel(name="vertical_sampler_channel") self.save_id = self.output.get("sample_id", {}) self.save_data_path = self.output["path"] / Path(self.output["dataset"]["name"]) if not os.path.exists(os.path.dirname(self.save_data_path)): os.makedirs(os.path.dirname(self.save_data_path)) def __load_data(self, config) -> CsvReader: if len(config) > 1: logger.warning("More than one dataset is not supported.") config = config[0] if config["type"] == "csv": data_reader = CsvReader(path=os.path.join(config["path"], config["name"]), has_id=config["has_id"], has_label=config["has_label"]) else: raise NotImplementedError("Dataset type {} is not supported.".format(config["type"])) return data_reader def _init_data(self) -> None: if len(self.input["dataset"]) > 0: data: CsvReader = self.__load_data(self.input["dataset"]) self.data = data.table.set_index(data.ids) self.label_name = data.label_name() else: raise NotImplementedError("Dataset was not configured.")
2,228
40.277778
111
py
XFL
XFL-master/python/algorithm/framework/vertical/sampler/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import json import os from pathlib import Path from typing import Any import pandas as pd from sklearn.utils import resample from algorithm.framework.vertical.sampler.base import VerticalSamplerBase from common.utils.logger import logger from service.fed_control import ProgressCalculator class VerticalSamplerLabelTrainer(VerticalSamplerBase): """ local feature sampler method: str, "random" or "stratify", default: "random"; strategy: str, "downsample" or "upsample", default: "downsample"; fraction: int, float or list (sampling ratios of each category), e.g. [[0,0.1], [1,0.2]], default: {"percentage": 0.1}; random_state: int, RandomState instance or None, default=None """ def __init__(self, train_conf): super().__init__(train_conf) self.label_count = {} self.label_idset = {} self._parse_config() self._init_data() self.sample_ids = [] def _parse_config(self) -> None: self.random_state = self.train_params.get("random_state", None) self.method = self.train_params.get("method", "random") self.strategy = self.train_params.get("strategy", "downsample") self.fraction = self.train_params.get("fraction", {"number": 0.1}) # whether for new customers filter self.infer_params = self.train_params.get("marketing_specified", {}) if len(self.infer_params) > 0: self.threshold_method = self.infer_params["threshold_method"] self.threshold = self.infer_params["threshold"] def fraction_transform(self) -> Any: def fraction_num(fraction, num): frac_num = int(fraction * num) if self.strategy == "downsample": if fraction < 0 or fraction > 1: raise ValueError("Fraction should be a numeric number between 0 and 1") return max(1, frac_num) elif self.strategy == "upsample": if fraction < 0: raise ValueError("Fraction should be a numeric number larger than 0") return frac_num fraction_key = list(self.fraction.keys())[0] if fraction_key == "percentage": tmp = fraction_num(self.fraction[fraction_key], len(self.data)) return tmp elif fraction_key == "labeled_percentage": tmp = [(tup[0], fraction_num(tup[1], self.label_count[tup[0]])) for tup in self.fraction[fraction_key]] return tmp elif fraction_key == "number": return self.fraction[fraction_key] else: raise NotImplementedError("Fraction key {} is not supported.".format(fraction_key)) def random_method(self) -> Any: sample_num = self.fraction_transform() if self.strategy == "downsample": sample_ids = resample(self.data.index, replace=False, n_samples=sample_num, random_state=self.random_state) new_data = self.data.loc[sample_ids] logger.info("Downsample completed.") return sample_ids, new_data elif self.strategy == "upsample": sample_ids = resample(self.data.index, replace=True, n_samples=sample_num, random_state=self.random_state) new_data = self.data.loc[sample_ids] logger.info("Upsample completed.") return sample_ids, new_data else: raise NotImplementedError("Strategy type {} is not supported.".format(self.strategy)) def stratify_method(self) -> Any: sample_num = self.fraction_transform() sample_ids = [] new_data = pd.DataFrame() if self.strategy == "downsample": for label, label_num in sample_num: sample_ids_ = resample(self.label_idset[label], replace=False, n_samples=label_num, random_state=self.random_state) new_data_ = self.data.loc[sample_ids_] sample_ids += sample_ids_ new_data = pd.concat([new_data, new_data_]) logger.info("Downsample completed.") return sample_ids, new_data elif self.strategy == "upsample": for label, label_num in sample_num: sample_ids_ = resample(self.label_idset[label], replace=True, n_samples=label_num, random_state=self.random_state) new_data_ = self.data.loc[sample_ids_] sample_ids += sample_ids_ new_data = pd.concat([new_data, new_data_]) logger.info("Upsample completed.") return sample_ids, new_data else: raise NotImplementedError("Strategy type {} is not supported.".format(self.strategy)) def fit(self) -> None: new_data = None # for most cases if len(self.infer_params) == 0: if self.method == "random": self.sample_ids, new_data = self.random_method() logger.info("Random sampler completed.") elif self.method == "stratify": self.label_count = self.data.groupby(self.label_name)[self.label_name].count().to_dict() self.label_idset = self.data.groupby(self.label_name).apply(lambda group: list(group.index)).to_dict() self.sample_ids, new_data = self.stratify_method() logger.info("Stratify sampler completed.") else: raise NotImplementedError("Method type {} is not supported.".format(self.method)) # for new customers filter elif len(self.infer_params) > 0: key = self.data.columns[0] if self.threshold_method == "percentage": threshold_num = len(self.data) * self.threshold self.sample_ids = self.data.sort_values(by=key, ascending=False).iloc[:int(threshold_num)].index elif self.threshold_method == "number": if self.threshold > len(self.data): raise OverflowError("Threshold number {} is larger than input data size.".format(self.threshold)) else: self.sample_ids = self.data.sort_values(by=key, ascending=False).iloc[:int(self.threshold)].index elif self.threshold_method == "score": self.sample_ids = self.data[self.data[key] > self.threshold].index else: raise NotImplementedError("Method type {} is not supported.".format(self.threshold_method)) # save if new_data is not None: new_data.to_csv(self.save_data_path, index=self.input["dataset"][0]["has_id"]) logger.info("Data saved to {}.".format(self.save_data_path)) if len(self.save_id) > 0: save_id_path = self.output["path"] / Path(self.output["sample_id"]["name"]) if not os.path.exists(os.path.dirname(save_id_path)): os.makedirs(os.path.dirname(save_id_path)) with open(save_id_path, "w") as wf: json.dump(list(self.sample_ids), wf) logger.info("Sample ids saved to {}.".format(save_id_path)) # send ids to trainer self.broadcast_channel.broadcast(self.sample_ids) # update the progress of 100 to show the training is finished ProgressCalculator.finish_progress()
8,302
45.385475
118
py
XFL
XFL-master/python/algorithm/framework/vertical/sampler/__init__.py
0
0
0
py
XFL
XFL-master/python/algorithm/framework/vertical/sampler/trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import json import os from pathlib import Path from algorithm.framework.vertical.sampler.base import VerticalSamplerBase from common.utils.logger import logger class VerticalSamplerTrainer(VerticalSamplerBase): def __init__(self, train_conf): super().__init__(train_conf) def fit(self) -> None: # receive sample_ids sample_ids = self.broadcast_channel.recv() new_data = self.data.loc[sample_ids] if len(self.save_id) > 0: save_id_path = self.output["path"] / Path(self.output["sample_id"]["name"]) if not os.path.exists(os.path.dirname(save_id_path)): os.makedirs(os.path.dirname(save_id_path)) with open(save_id_path, "w") as wf: json.dump(list(sample_ids), wf) logger.info("Sample ids saved to {}.".format(save_id_path)) new_data.to_csv(self.save_data_path, index=self.input["dataset"][0]["has_id"]) logger.info("Data saved to {}.".format(self.save_data_path))
1,616
39.425
87
py
XFL
XFL-master/python/algorithm/framework/vertical/linear_regression/base.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import torch import torch.nn as nn from torch.utils.data import DataLoader, TensorDataset from algorithm.core.data_io import CsvReader from algorithm.framework.vertical.vertical_model_base import VerticalModelBase from common.utils.logger import logger from common.utils.model_preserver import ModelPreserver from service.fed_config import FedConfig class VerticalLinearRegression(nn.Module): def __init__(self, input_dim: int, bias: bool = False): super(VerticalLinearRegression, self).__init__() self.linear = torch.nn.Linear(input_dim, 1, bias=bias) self.linear.requires_grad_(False) def forward(self, x): return self.linear(x) class VerticalLinearRegressionBase(VerticalModelBase): def __init__(self, train_conf: dict, label: bool = False, *args, **kwargs): """_summary_ Args: train_conf (dict): _description_ label (bool, optional): _description_. Defaults to False. """ super().__init__(train_conf) self._parse_config() self.train_conf = train_conf self.label = label self.data_dim = None self.model = None self.train_dataloader, self.eval_dataloader = None, None if FedConfig.node_id != "assist_trainer": self._init_dataloader() def _parse_config(self) -> None: super()._parse_config() self.model_name = self.model_info.get("name") self.save_model_name = self.output.get("model", {}).get("name") self.save_onnx_model_name = self.output.get("onnx_model", {}).get("name", "") if not os.path.exists(self.save_dir): os.makedirs(self.save_dir) self.global_epoch = self.train_params.get("global_epoch") self.batch_size = self.train_params.get("batch_size") self.encryption_config = self.train_params.get("encryption") self.optimizer_config = self.train_params.get("optimizer") self.pretrain_model_path = self.input.get("pretrained_model", {}).get("path") self.random_seed = self.train_params.get("random_seed", None) self.early_stopping_config = self.train_params.get("early_stopping") self.save_frequency = self.interaction_params.get("save_frequency") @staticmethod def set_seed(seed): torch.manual_seed(seed) torch.cuda.manual_seed(seed) def _init_model(self, bias: bool = False) -> None: """ Init linear regression model. Returns: None """ logger.info("Init model start.") self.model = VerticalLinearRegression(input_dim=self.data_dim, bias=bias) # Load pretrained model if needed. if self.pretrain_model_path is not None and self.pretrain_model_path != "": checkpoint = ModelPreserver.load(os.path.join(self.pretrain_model_path, self.input.get( "pretrained_model").get("name", None))) self.model.load_state_dict(checkpoint["state_dict"]) logger.info("Init model completed.") def __load_data(self, config) -> CsvReader: config = config[0] if config["type"] == "csv": data_reader = CsvReader(path=os.path.join(config["path"], config["name"]), has_id=config["has_id"], has_label=config["has_label"]) else: raise NotImplementedError("Dataset type {} is not supported.".format(config["type"])) return data_reader def _init_data(self) -> None: if len(self.input_trainset) > 0: data: CsvReader = self.__load_data(self.input_trainset) self.train = data.features() self.train_label = data.label() self.train_ids = list(range(len(data.ids))) else: raise NotImplementedError("Trainset was not configured.") if self.label: assert len(self.train) == len(self.train_label) if len(self.input_valset) > 0: data: CsvReader = self.__load_data(self.input_valset) self.val = data.features() self.val_label = data.label() self.val_ids = list(range(len(data.ids))) if self.label: assert len(self.val) == len(self.val_label) def _init_dataloader(self) -> None: """ Load raw data. Returns: """ logger.info("Dataloader initiation start.") self._init_data() if self.label: self.train_dataloader = DataLoader( dataset=TensorDataset(torch.tensor(self.train, dtype=torch.float32), torch.unsqueeze(torch.tensor(self.train_label), dim=-1), torch.unsqueeze(torch.tensor(self.train_ids), dim=-1)), batch_size=self.batch_size, shuffle=True ) self.val_dataloader = DataLoader( dataset=TensorDataset(torch.tensor(self.val, dtype=torch.float32), torch.unsqueeze(torch.tensor(self.val_label), dim=-1), torch.unsqueeze(torch.tensor(self.val_ids), dim=-1)), batch_size=self.batch_size, shuffle=False ) self.data_dim = torch.tensor(self.train).shape[-1] logger.info("Train data shape: {}.".format(list(torch.tensor(self.train).shape))) else: self.train_dataloader = DataLoader( dataset=TensorDataset(torch.tensor(self.train, dtype=torch.float32), torch.unsqueeze(torch.tensor(self.train_ids), dim=-1)), batch_size=self.batch_size, shuffle=True ) self.val_dataloader = DataLoader( dataset=TensorDataset(torch.tensor(self.val, dtype=torch.float32), torch.unsqueeze(torch.tensor(self.val_ids), dim=-1)), batch_size=self.batch_size, shuffle=False ) self.data_dim = torch.tensor(self.train).shape[-1] logger.info("Train data shape: {}.".format(list(torch.tensor(self.train).shape))) logger.info("Dataloader initiation completed.")
6,807
41.55
111
py
XFL
XFL-master/python/algorithm/framework/vertical/linear_regression/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import copy import secrets from functools import reduce from pathlib import Path from typing import Optional from common.checker.x_types import All import numpy as np import pandas as pd import tenseal as ts import torch from common.checker.matcher import get_matched_config from common.communication.gRPC.python.channel import BroadcastChannel, DualChannel from common.crypto.paillier.paillier import Paillier from common.utils.algo_utils import earlyStopping from common.utils.logger import logger from common.utils.utils import save_model_config from common.utils.model_io import ModelIO from service.fed_config import FedConfig from service.fed_node import FedNode from service.fed_control import ProgressCalculator from .base import VerticalLinearRegressionBase class VerticalLinearRegressionLabelTrainer(VerticalLinearRegressionBase): def __init__(self, train_conf: dict, *args, **kwargs): """ Vertical Linear Regression Args: train_conf: training parameters *args: **kwargs: """ self.sync_channel = BroadcastChannel(name="sync") self._sync_config(train_conf) super().__init__(train_conf, label=True, *args, **kwargs) if self.random_seed: self.set_seed(self.random_seed) self.progress_calculator = ProgressCalculator(self.global_epoch, len(self.train_dataloader)) self._init_model(bias=True) self.export_conf = [{ "class_name": "VerticalLinearRegression", "identity": self.identity, "filename": self.save_onnx_model_name, "input_dim": self.data_dim, "bias": True, "version": "1.4.0" }] self.es = earlyStopping(key=self.early_stopping_config["key"], patience=self.early_stopping_config["patience"], delta=self.early_stopping_config["delta"]) self.best_model = None self.broadcast_channel = BroadcastChannel(name="Public keys", root_id=FedConfig.get_assist_trainer()) self.dual_channels = {"intermediate_label_trainer": {}, "gradients_loss": None} for party_id in FedConfig.get_trainer(): self.dual_channels["intermediate_label_trainer"][party_id] = DualChannel( name="intermediate_label_trainer_" + party_id, ids=[FedConfig.node_id, party_id]) self.dual_channels["gradients_loss"] = DualChannel(name="gradients_loss_" + FedConfig.node_id, ids=[FedConfig.get_assist_trainer()] + [FedConfig.node_id]) self.train_result = None self.val_result = None self.dual_channels["gradients_loss"].send(len(self.train_dataloader)) self.dual_channels["gradients_loss"].send(self.global_epoch) self.dual_channels["gradients_loss"].send(self.batch_size) self.encryption_method = list(self.encryption_config.keys())[0].lower() self.dual_channels["gradients_loss"].send(self.encryption_config) self.dual_channels["gradients_loss"].send(self.encryption_method) def _sync_config(self, config): sync_rule = { "train_info": All() } config_to_sync = get_matched_config(config, sync_rule) self.sync_channel.broadcast(config_to_sync) def predict(self, input_data): pred_prob_epoch, y_epoch = [], [] for batch_idx, (x_batch, y_batch, _) in enumerate(input_data): pred_trainer_list = [] pred_label_trainer = self.model(x_batch).numpy().astype(np.float32).flatten() for party_id in FedConfig.get_trainer(): pred_trainer_list.append(self.dual_channels["intermediate_label_trainer"][party_id].recv( use_pickle=True)) # calculate prediction of batch pred_total = pred_label_trainer + reduce(lambda x, y: x + y, pred_trainer_list) # calculate prediction of epoch pred_prob_epoch += pred_total.tolist() y_epoch += y_batch.numpy().astype(np.float32).flatten().tolist() return y_epoch, pred_prob_epoch def fit(self): self.check_data() public_context = None num_cores = -1 rng = secrets.SystemRandom() logger.info("Vertical linear regression training start") # receive encryption key from assist trainer if self.encryption_method == "ckks": logger.info("Receive ckks public key.") public_context = self.broadcast_channel.recv(use_pickle=False) public_context = ts.context_from(public_context) logger.info("Public key received.") elif self.encryption_method == "paillier": logger.info("Receive paillier public key.") public_context = self.broadcast_channel.recv(use_pickle=False) public_context = Paillier.context_from(public_context) logger.info("Public key received.") elif self.encryption_method == "plain": pass else: raise ValueError(f"Encryption method {self.encryption_method} not supported! Valid methods are " f"'paillier', 'ckks', 'plain'.") # train for epoch in range(1, self.global_epoch + 1): loss_epoch = 0 for batch_idx, (x_batch, y_batch, _) in enumerate(self.train_dataloader): pred_trainer = [] loss_trainer = [] loss_between_trainer = 0 enc_pred_residual = None enc_loss_label_trainer = None regular_loss_tmp = 0 regular_gradient_tmp = 0 enc_regular_gradient_tmp = 0 # calculate regular results if self.optimizer_config['p'] == 1: regular_loss_tmp = torch.abs(self.model.linear.weight).sum() * self.optimizer_config['alpha'] regular_gradient_tmp = self.optimizer_config['alpha'] * (torch.abs(self.model.linear.weight) / self.model.linear.weight) elif self.optimizer_config['p'] == 2: regular_loss_tmp = (self.model.linear.weight ** 2).sum() * self.optimizer_config['alpha'] / 2 regular_gradient_tmp = self.optimizer_config['alpha'] * self.model.linear.weight elif self.optimizer_config['p'] == 0: pass # compute theta_scheduler * label_trainer and loss of label_trainer logger.info("Calculate intermediate result of label trainer.") pred_label_trainer = self.model(x_batch) pred_residual = pred_label_trainer - y_batch # receive intermediate results from trainers for party_id in FedConfig.get_trainer(): if self.encryption_method == "ckks": pred_trainer.append(ts.ckks_vector_from(public_context, self.dual_channels[ "intermediate_label_trainer"][party_id].recv(use_pickle=False))) loss_trainer.append(ts.ckks_vector_from(public_context, self.dual_channels[ "intermediate_label_trainer"][party_id].recv(use_pickle=False))) elif self.encryption_method == "paillier": pred_trainer.append(Paillier.ciphertext_from(public_context, self.dual_channels[ "intermediate_label_trainer"][party_id].recv(use_pickle=False))) loss_trainer.append(Paillier.ciphertext_from(public_context, self.dual_channels[ "intermediate_label_trainer"][party_id].recv(use_pickle=False))) elif self.encryption_method == "plain": pred_trainer.append(self.dual_channels["intermediate_label_trainer"][party_id].recv()) loss_trainer.append(self.dual_channels["intermediate_label_trainer"][party_id].recv()) logger.info("Received predictions from trainers, length of collect list is {}." .format(len(pred_trainer))) # calculate total loss logger.info("Calculate total loss.") square_tmp = (pred_residual ** 2).sum() / 2 loss_label_trainer = square_tmp + regular_loss_tmp if self.encryption_method == "ckks": loss_between_label_trainer = np.sum([pred_t.matmul(pred_residual.numpy()) for pred_t in pred_trainer ]) else: loss_between_label_trainer = np.sum(pred_residual.numpy().flatten() * pred_trainer ) # calculate total loss_between_trainer when there are more than one trainer if len(pred_trainer) > 1: if self.encryption_method == "plain": loss_between_trainer = np.sum([np.sum(i * j) if ind_i != ind_j else 0 for ind_i, i in enumerate(pred_trainer) for ind_j, j in enumerate(pred_trainer)]) / 2 elif self.encryption_method == "ckks": loss_between_trainer = np.sum([i.dot(j) if ind_i != ind_j else 0 for ind_i, i in enumerate(pred_trainer) for ind_j, j in enumerate(pred_trainer)]) * 0.5 elif self.encryption_method == "paillier": loss_between_trainer = [] for party_id in FedConfig.get_trainer(): tmp = self.dual_channels["intermediate_label_trainer"][party_id].recv(use_pickle=False) tmp = Paillier.ciphertext_from(public_context, tmp) loss_between_trainer.append(tmp) loss_between_trainer = np.sum(loss_between_trainer) / 2 if self.encryption_method == "ckks": enc_loss_label_trainer = ts.ckks_vector(public_context, loss_label_trainer.numpy().astype(np.float32).flatten()) elif self.encryption_method == "paillier": enc_loss_label_trainer = Paillier.encrypt(public_context, float(loss_label_trainer), precision=self.encryption_config[self.encryption_method][ "precision"], obfuscation=True, num_cores=num_cores) elif self.encryption_method == "plain": enc_loss_label_trainer = loss_label_trainer enc_loss_batch = loss_between_trainer + loss_between_label_trainer + enc_loss_label_trainer + np.sum( loss_trainer) # send total loss to assist_trainer logger.info("Send total loss to assist_trainer.") if self.encryption_method == "ckks": self.dual_channels["gradients_loss"].send(enc_loss_batch.serialize(), use_pickle=False) elif self.encryption_method == "paillier": self.dual_channels["gradients_loss"].send(Paillier.serialize(enc_loss_batch), use_pickle=False) elif self.encryption_method == "plain": self.dual_channels["gradients_loss"].send(enc_loss_batch) # receive decrypted loss from assist_trainer logger.info("Receive total loss from assist_trainer.") loss_batch = self.dual_channels["gradients_loss"].recv() loss_batch = loss_batch / x_batch.shape[0] logger.info("Loss of {} batch is {}".format(batch_idx, loss_batch)) loss_epoch += loss_batch * x_batch.shape[0] # calculate intermediate result d logger.info("Calculate intermediate result d.") pred_rest_trainer = reduce(lambda x, y: x + y, pred_trainer) if self.encryption_method == "ckks": enc_pred_residual = ts.ckks_vector(public_context, pred_residual.numpy().astype(np.float32).flatten()) elif self.encryption_method == "paillier": enc_pred_residual = Paillier.encrypt(public_context, pred_residual.numpy().astype(np.float32).flatten(), precision=self.encryption_config[self.encryption_method][ "precision"], obfuscation=True, num_cores=num_cores) elif self.encryption_method == "plain": enc_pred_residual = pred_residual.numpy().astype(np.float32).flatten() enc_d = enc_pred_residual + pred_rest_trainer # send intermediate result d to trainer logger.info("Send intermediate result d to trainer.") for party_id in FedConfig.get_trainer(): if self.encryption_method == "ckks": self.dual_channels["intermediate_label_trainer"][party_id].send(enc_d.serialize(), use_pickle=False) elif self.encryption_method == "paillier": self.dual_channels["intermediate_label_trainer"][party_id].send(Paillier.serialize(enc_d), use_pickle=False) elif self.encryption_method == "plain": self.dual_channels["intermediate_label_trainer"][party_id].send(enc_d) # calculate gradient for label_trainer logger.info("Calculate gradients for label_trainer.") if self.encryption_method == "ckks": enc_regular_gradient_tmp = ts.ckks_vector(public_context, regular_gradient_tmp.numpy().astype(np.float32).flatten()) elif self.encryption_method == "paillier": enc_regular_gradient_tmp = Paillier.encrypt( public_context, regular_gradient_tmp.numpy().astype(np.float32).flatten(), precision=self.encryption_config[self.encryption_method]["precision"], obfuscation=True, num_cores=num_cores) elif self.encryption_method == "plain": enc_regular_gradient_tmp = regular_gradient_tmp.numpy().astype(np.float32).flatten() if self.encryption_method == "ckks": gradient_label_trainer_w = enc_d.matmul(x_batch.numpy()) + enc_regular_gradient_tmp else: gradient_label_trainer_w = np.matmul(enc_d.reshape(1, len(enc_d)), x_batch.numpy() ) + enc_regular_gradient_tmp gradient_label_trainer_b = enc_d if self.encryption_method == "ckks": # add noise to encrypted gradients and send to assist_trainer logger.info("Calculate noised gradients for label_trainer.") noise = np.array([rng.randint(1 << 24, 1 << 26) - (1 << 25) for _ in range(x_batch.shape[1])], dtype=np.float32) noise /= 100000 noise_b = np.array([rng.randint(1 << 24, 1 << 26) - (1 << 25) for _ in range(x_batch.shape[0])], dtype=np.float32) noise_b /= 100000 noised_gradient_label_trainer_w = gradient_label_trainer_w + noise noised_gradient_label_trainer_b = gradient_label_trainer_b + noise_b logger.info("Send noised gradient to assist_trainer.") self.dual_channels["gradients_loss"].send(noised_gradient_label_trainer_w.serialize(), use_pickle=False) self.dual_channels["gradients_loss"].send(noised_gradient_label_trainer_b.serialize(), use_pickle=False) # receive decrypted gradient from assist_trainer logger.info("Receive decrypted gradient from assist_trainer.") noised_decrypt_gradient = self.dual_channels["gradients_loss"].recv() noised_decrypt_gradient_label_trainer_w = noised_decrypt_gradient["noised_gradient_label_trainer_w"] noised_decrypt_gradient_label_trainer_b = noised_decrypt_gradient["noised_gradient_label_trainer_b"] gradient_label_trainer_w = noised_decrypt_gradient_label_trainer_w - noise gradient_label_trainer_b = noised_decrypt_gradient_label_trainer_b - np.sum(noise_b) elif self.encryption_method == "paillier": # add noise to encrypted gradients and send to assist_trainer logger.info("Calculate noised gradients for label_trainer.") noise = np.array([rng.randint(1 << 24, 1 << 26) - (1 << 25) for _ in range(x_batch.shape[1])], dtype=np.float32) noise /= 100000 noise_b = np.array([rng.randint(1 << 24, 1 << 26) - (1 << 25) for _ in range(x_batch.shape[0])], dtype=np.float32) noise_b /= 100000 noised_gradient_label_trainer_w = gradient_label_trainer_w + noise noised_gradient_label_trainer_b = gradient_label_trainer_b + noise_b logger.info("Send noised gradient to assist_trainer.") self.dual_channels["gradients_loss"].send(Paillier.serialize(noised_gradient_label_trainer_w), use_pickle=False) self.dual_channels["gradients_loss"].send(Paillier.serialize(noised_gradient_label_trainer_b), use_pickle=False) # receive decrypted gradient from assist_trainer logger.info("Receive decrypted gradient from assist_trainer.") noised_decrypt_gradient = self.dual_channels["gradients_loss"].recv() noised_decrypt_gradient_label_trainer_w = noised_decrypt_gradient["noised_gradient_label_trainer_w"] noised_decrypt_gradient_label_trainer_b = noised_decrypt_gradient["noised_gradient_label_trainer_b"] gradient_label_trainer_w = noised_decrypt_gradient_label_trainer_w - noise gradient_label_trainer_b = noised_decrypt_gradient_label_trainer_b - np.sum(noise_b) elif self.encryption_method == "plain": gradient_label_trainer_b = gradient_label_trainer_b.sum() # update w and b of label_trainer gradient_label_trainer_w = gradient_label_trainer_w / x_batch.shape[0] gradient_label_trainer_b = gradient_label_trainer_b / x_batch.shape[0] logger.info("Update weights of label trainer.") self.model.linear.weight -= (torch.FloatTensor(gradient_label_trainer_w) * self.optimizer_config["lr"]) self.model.linear.bias -= (gradient_label_trainer_b * self.optimizer_config["lr"]) # calculate and update the progress of the training self.progress_calculator.cal_custom_progress(epoch, batch_idx+1) loss_epoch = loss_epoch * (1 / len(self.train)) logger.info("Loss of {} epoch is {}".format(epoch, loss_epoch)) # predict train and val results for metrics logger.info("Predict train weights of label trainer.") self.train_result = self.predict(self.train_dataloader) loss_train_met = {"loss": loss_epoch} self._calc_metrics(np.array(self.train_result[1], dtype=float), np.array(self.train_result[0]), epoch, stage="train", loss=loss_train_met) logger.info("Predict val weights of label trainer.") self.val_result = self.predict(self.val_dataloader) val_residual = np.array(self.val_result[1]) - np.array(self.val_result[0]) loss_val_met = {"loss": np.mean((val_residual ** 2) / 2)} # no regular val_metrics = self._calc_metrics(np.array(self.val_result[1], dtype=float), np.array(self.val_result[0]), epoch, stage="val", loss=loss_val_met) # early stopping val_metrics["loss"] = - val_metrics["loss"] if self.early_stopping_config["patience"] > 0: early_stop_flag, best_model_flag = self.es(val_metrics) else: early_stop_flag, best_model_flag = False, True # update best model if best_model_flag: self.best_model = copy.deepcopy(self.model) # send flags to trainers for party_id in FedConfig.get_trainer(): self.dual_channels["intermediate_label_trainer"][party_id].send( [early_stop_flag, best_model_flag, self.early_stopping_config["patience"]], use_pickle=True) # if need to save results by epoch if self.save_frequency > 0 and epoch % self.save_frequency == 0: # ModelPreserver.save(save_dir=self.save_dir, model_name=self.save_model_name, # state_dict=self.model.state_dict(), epoch=epoch) self.save_model(epoch=epoch) # if early stopping, break if early_stop_flag: # update the progress of 100 to show the training is finished ProgressCalculator.finish_progress() break # save model for infer self.save_model(epoch=None) # if not early stopping, save probabilities and model self._save_prob() # calculate feature importance self._save_feature_importance(self.dual_channels) def save_model(self, epoch: Optional[int] = None): if not epoch: save_model_config(stage_model_config=self.export_conf, save_path=Path(self.save_dir)) if self.save_model_name: ModelIO.save_torch_model( state_dict=self.best_model.state_dict(), save_dir=self.save_dir, model_name=self.save_model_name, meta_dict={}, epoch=epoch ) if self.save_onnx_model_name: ModelIO.save_torch_onnx( model=self.best_model, input_dim=(self.data_dim,), save_dir=self.save_dir, model_name=self.save_onnx_model_name, epoch=epoch ) def _save_prob(self): if self.interaction_params.get("write_training_prediction"): self._write_prediction(self.train_result[1], self.train_result[0], self.train_ids, stage="train", final=True) if self.interaction_params.get("write_validation_prediction"): self._write_prediction(self.val_result[1], self.val_result[0], self.val_ids, stage="val", final=True) def check_data(self): dim_channel = BroadcastChannel(name="check_data_com", ids=[FedConfig.node_id] + FedConfig.get_trainer()) n = self.data_dim dims = dim_channel.collect() for dim in dims: n += dim if n <= 0: raise ValueError("Number of the feature is zero. Stop training.") def _save_feature_importance(self, channel): res = {"owner_id": [], "fid": [], "importance": []} other_weight_list = [] for party_id in FedConfig.get_trainer(): other_weight_list.append(channel["intermediate_label_trainer"][party_id].recv(use_pickle=True)) for (owner_id, weights) in other_weight_list: for fid, weight in enumerate(weights): res["owner_id"].append(owner_id) res["fid"].append(fid) res["importance"].append(float(weight)) for fid, weight in enumerate(self.best_model.state_dict()["linear.weight"][0]): res["owner_id"].append(FedNode.node_id) res["fid"].append(fid) res["importance"].append(float(weight)) res = pd.DataFrame(res).sort_values(by="importance", key=lambda col: np.abs(col), ascending=False) res.to_csv(Path(self.save_dir, self.output["feature_importance"]["name"]), header=True, index=False, float_format="%.6g")
26,235
57.825112
120
py
XFL
XFL-master/python/algorithm/framework/vertical/linear_regression/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import numpy as np import tenseal as ts from common.communication.gRPC.python.channel import BroadcastChannel, DualChannel from common.crypto.paillier.paillier import Paillier from common.utils.logger import logger from service.fed_config import FedConfig class VerticalLinearRegressionAssistTrainer(object): def __init__(self, *args, **kwargs): """[summary] assist_trainer """ self.broadcast_channel = BroadcastChannel(name="Public keys", root_id=FedConfig.get_assist_trainer()) self.dual_channels = {"gradients_loss": {}} self.party_id_list = FedConfig.get_label_trainer() + FedConfig.get_trainer() for party_id in self.party_id_list: self.dual_channels["gradients_loss"][party_id] = DualChannel(name="gradients_loss_" + party_id, ids=[FedConfig.node_id, party_id]) self.batch_num = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv() self.global_epoch = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv() self.batch_size = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv() self.encryption_config = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv() self.encryption_method = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv() self.private_context = None self.public_context = None # send encryption key to all parties if self.encryption_method == "ckks": self.private_context = ts.context( ts.SCHEME_TYPE.CKKS, poly_modulus_degree=self.encryption_config[self.encryption_method]["poly_modulus_degree"], coeff_mod_bit_sizes=self.encryption_config[self.encryption_method]["coeff_mod_bit_sizes"] ) self.private_context.generate_galois_keys() self.private_context.generate_relin_keys() self.private_context.global_scale = 1 << self.encryption_config[self.encryption_method][ "global_scale_bit_size"] serialized_public_context = self.private_context.serialize( save_public_key=True, save_secret_key=False, save_galois_keys=True, save_relin_keys=True ) logger.info("Broadcast ckks public keys.") self.public_context_ser = serialized_public_context self.broadcast_channel.broadcast(self.public_context_ser, use_pickle=False) logger.info("Broadcast completed.") elif self.encryption_method == "paillier": self.num_cores = -1 if self.encryption_config[self.encryption_method]["parallelize_on"] else 1 self.private_context = Paillier.context(self.encryption_config[self.encryption_method]["key_bit_size"], djn_on=self.encryption_config[self.encryption_method]["djn_on"]) logger.info("Broadcast paillier public keys.") self.public_context_ser = self.private_context.to_public().serialize() self.broadcast_channel.broadcast(self.public_context_ser, use_pickle=False) logger.info("Broadcast completed.") elif self.encryption_method == "plain": pass else: raise ValueError(f"Encryption method {self.encryption_method} not supported! Valid methods are 'paillier', " f"'ckks', 'plain'.") def fit(self): """ train model Model parameters need to be updated before fitting. """ # send encryption key to all parties if self.encryption_method in ["ckks", "paillier"]: logger.info("Broadcast ckks public keys.") self.broadcast_channel.broadcast(self.public_context_ser, use_pickle=False) logger.info("Broadcast completed.") # train for epoch in range(1, self.global_epoch + 1): for batch_idx in range(self.batch_num): # receive and decrypt total encrypted loss and send to label_trainer logger.info("Receive and decrypted total loss and send back to label_trainer.") if self.encryption_method == "ckks": enc_loss_batch = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv( use_pickle=False) decrypted_loss_batch = ts.ckks_vector_from(self.private_context, enc_loss_batch).decrypt()[0] elif self.encryption_method == "paillier": enc_loss_batch = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv( use_pickle=False) decrypted_loss_batch = Paillier.decrypt(self.private_context, Paillier.ciphertext_from( None, enc_loss_batch), dtype='float', num_cores=self.num_cores) elif self.encryption_method == "plain": enc_loss_batch = self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].recv() decrypted_loss_batch = enc_loss_batch self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].send(decrypted_loss_batch) logger.info( "Loss of {} batch {} epoch is {}".format(batch_idx, epoch, decrypted_loss_batch / self.batch_size)) # receive encrypted noised gradients from other parties and decrypt and send back to other parties if self.encryption_method == "ckks" or self.encryption_method == "paillier": # trainer for party_id in FedConfig.get_trainer(): en_noised_gradient_trainer_w = self.dual_channels["gradients_loss"][party_id].recv( use_pickle=False) if self.encryption_method == "ckks": noised_gradient_trainer_w = ts.ckks_vector_from(self.private_context, en_noised_gradient_trainer_w).decrypt() elif self.encryption_method == "paillier": noised_gradient_trainer_w = Paillier.decrypt(self.private_context, Paillier.ciphertext_from( None, en_noised_gradient_trainer_w), dtype='float', num_cores=self.num_cores) self.dual_channels["gradients_loss"][party_id].send(noised_gradient_trainer_w) # label_trainer en_noised_gradient_label_trainer_w = self.dual_channels["gradients_loss"][ FedConfig.get_label_trainer()[0]].recv(use_pickle=False) en_noised_gradient_label_trainer_b = self.dual_channels["gradients_loss"][ FedConfig.get_label_trainer()[0]].recv(use_pickle=False) if self.encryption_method == "ckks": noised_gradient_label_trainer_w = ts.ckks_vector_from( self.private_context, en_noised_gradient_label_trainer_w).decrypt() noised_gradient_label_trainer_b = ts.ckks_vector_from( self.private_context, en_noised_gradient_label_trainer_b).decrypt() elif self.encryption_method == "paillier": noised_gradient_label_trainer_w = Paillier.decrypt( self.private_context, Paillier.ciphertext_from(None, en_noised_gradient_label_trainer_w), dtype='float', num_cores=self.num_cores) noised_gradient_label_trainer_b = Paillier.decrypt( self.private_context, Paillier.ciphertext_from(None, en_noised_gradient_label_trainer_b), dtype='float', num_cores=self.num_cores) # calculate sum of gradient b noised_gradient_label_trainer_b = np.sum(noised_gradient_label_trainer_b) grad_send = {"noised_gradient_label_trainer_w": noised_gradient_label_trainer_w, "noised_gradient_label_trainer_b": noised_gradient_label_trainer_b} self.dual_channels["gradients_loss"][FedConfig.get_label_trainer()[0]].send(grad_send)
9,101
61.772414
120
py
XFL
XFL-master/python/algorithm/framework/vertical/linear_regression/__init__.py
0
0
0
py
XFL
XFL-master/python/algorithm/framework/vertical/linear_regression/trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import copy import secrets from pathlib import Path from typing import Optional import numpy as np import tenseal as ts import torch from common.utils.utils import update_dict from common.communication.gRPC.python.channel import BroadcastChannel, DualChannel from common.crypto.paillier.paillier import Paillier from common.utils.logger import logger from common.utils.model_io import ModelIO from service.fed_config import FedConfig from service.fed_node import FedNode from common.utils.utils import save_model_config from .base import VerticalLinearRegressionBase class VerticalLinearRegressionTrainer(VerticalLinearRegressionBase): def __init__(self, train_conf: dict, *args, **kwargs): """[summary] Args: train_conf (dict): [description] """ self.sync_channel = BroadcastChannel(name="sync") conf = self._sync_config() update_dict(train_conf, conf) super().__init__(train_conf, label=False, *args, **kwargs) self._init_model() self.export_conf = [{ "class_name": "VerticalLinearRegression", "identity": self.identity, "filename": self.save_onnx_model_name, "input_dim": self.data_dim, "bias": False }] if self.random_seed: self.set_seed(self.random_seed) self.best_model = None self.broadcast_channel = BroadcastChannel(name="Public keys", root_id=FedConfig.get_assist_trainer()) if len(FedConfig.get_trainer()) > 1: self.broadcast_trainer = BroadcastChannel(name="Trainer exchange", root_id=FedConfig.node_id, ids=FedConfig.get_trainer()) self.dual_channels = { "intermediate_label_trainer": DualChannel(name="intermediate_label_trainer_" + FedConfig.node_id, ids=FedConfig.get_label_trainer() + [FedConfig.node_id]), "gradients_loss": DualChannel(name="gradients_loss_" + FedConfig.node_id, ids=[FedConfig.get_assist_trainer()] + [FedConfig.node_id]) } def predict(self, input_data): for batch_idx, x_batch in enumerate(input_data): # calculate prediction of batch pred_trainer = self.model(x_batch[0]) # send to label_trainer self.dual_channels["intermediate_label_trainer"].send(pred_trainer.numpy().astype(np.float32).flatten(), use_pickle=True) def _sync_config(self): config = self.sync_channel.recv() return config def fit(self): """ train model Model parameters need to be updated before fitting. """ self.check_data() num_cores = -1 encryption_config = self.encryption_config encryption_method = list(self.encryption_config.keys())[0].lower() logger.info("Vertical linear regression training start") # receive encryption key from assist trainer public_context = None if encryption_method == "ckks": logger.info("Receive ckks public key.") public_context = self.broadcast_channel.recv(use_pickle=False) public_context = ts.context_from(public_context) logger.info("Public key received.") elif encryption_method == "paillier": logger.info("Receive paillier public key.") public_context = self.broadcast_channel.recv(use_pickle=False) public_context = Paillier.context_from(public_context) logger.info("Public key received.") elif encryption_method == "plain": pass else: raise ValueError( f"Encryption method {encryption_method} not supported! Valid methods are 'paillier', 'ckks', 'plain'.") rng = secrets.SystemRandom() # train for epoch in range(1, self.global_epoch + 1): for batch_idx, x_batch in enumerate(self.train_dataloader): regular_loss_tmp = 0 regular_gradient_tmp = 0 enc_regular_gradient_tmp = 0 # calculate regular results if self.optimizer_config['p'] == 1: regular_loss_tmp = torch.abs(self.model.linear.weight).sum() * self.optimizer_config['alpha'] regular_gradient_tmp = self.optimizer_config['alpha'] * (torch.abs(self.model.linear.weight) / self.model.linear.weight) elif self.optimizer_config['p'] == 2: regular_loss_tmp = (self.model.linear.weight ** 2).sum() * self.optimizer_config['alpha'] / 2 regular_gradient_tmp = self.optimizer_config['alpha'] * self.model.linear.weight elif self.optimizer_config['p'] == 0: pass # compute theta_trainer * x_trainer and loss of x_trainer pred_trainer = self.model(x_batch[0]) square_tmp = (pred_trainer ** 2).sum() / 2 loss_trainer = square_tmp + regular_loss_tmp # send intermediate results to label trainer. logger.info("Send intermediate result to label trainer.") enc_pred_trainer = None if encryption_method == "ckks": enc_pred_trainer = ts.ckks_vector(public_context, pred_trainer.numpy().astype(np.float32).flatten()) enc_loss_trainer = ts.ckks_vector(public_context, loss_trainer.numpy().astype(np.float32).flatten()) self.dual_channels["intermediate_label_trainer"].send(enc_pred_trainer.serialize(), use_pickle=False) self.dual_channels["intermediate_label_trainer"].send(enc_loss_trainer.serialize(), use_pickle=False) elif encryption_method == "paillier": enc_pred_trainer = Paillier.encrypt(public_context, pred_trainer.numpy().astype(np.float32).flatten(), precision=encryption_config[encryption_method]["precision"], obfuscation=True, num_cores=num_cores) enc_loss_trainer = Paillier.encrypt(public_context, loss_trainer.numpy().astype(np.float32).flatten(), precision=encryption_config[encryption_method]["precision"], obfuscation=True, num_cores=num_cores) self.dual_channels["intermediate_label_trainer"].send(Paillier.serialize(enc_pred_trainer), use_pickle=False) self.dual_channels["intermediate_label_trainer"].send(Paillier.serialize(enc_loss_trainer), use_pickle=False) elif encryption_method == "plain": enc_pred_trainer = pred_trainer.numpy().astype(np.float32).flatten() enc_loss_trainer = loss_trainer.numpy().astype(np.float32).flatten() self.dual_channels["intermediate_label_trainer"].send(enc_pred_trainer, use_pickle=True) self.dual_channels["intermediate_label_trainer"].send(enc_loss_trainer, use_pickle=True) # exchange theta_trainer * x_trainer to calculate loss_between_trainer when encryption is paillier logger.info("Calculate trainer_sum to label trainer when encryption is paillier.") if encryption_method == "paillier" and len(FedConfig.get_trainer()) > 1: trainer_sum = 0 logger.info("Send intermediate result to other trainers when encryption is paillier.") self.broadcast_trainer.broadcast(Paillier.serialize(enc_pred_trainer), use_pickle=False) logger.info("Receive intermediate result from other trainers when encryption is paillier.") trainer_tmp = self.broadcast_trainer.collect(use_pickle=False) for trainer_u in trainer_tmp: trainer_u = Paillier.ciphertext_from(public_context, trainer_u) trainer_sum += np.sum(trainer_u * pred_trainer.numpy().astype(np.float32).flatten()) logger.info("Send trainer_sum to label trainer when encryption is paillier.") self.dual_channels["intermediate_label_trainer"].send(Paillier.serialize(trainer_sum), use_pickle=False) # receive intermediate result d from label_trainer logger.info("Receive intermediate result d from label_trainer.") if encryption_method == "ckks": enc_d = self.dual_channels["intermediate_label_trainer"].recv(use_pickle=False) enc_d = ts.ckks_vector_from(public_context, enc_d) elif encryption_method == "paillier": enc_d = self.dual_channels["intermediate_label_trainer"].recv(use_pickle=False) enc_d = Paillier.ciphertext_from(public_context, enc_d) elif encryption_method == "plain": enc_d = self.dual_channels["intermediate_label_trainer"].recv() # calculate gradient for trainer and send to assist_trainer logger.info("Calculate gradients for trainer.") if encryption_method == "ckks": enc_regular_gradient_tmp = ts.ckks_vector(public_context, regular_gradient_tmp.numpy().astype(np.float32).flatten()) elif encryption_method == "paillier": enc_regular_gradient_tmp = Paillier.encrypt( public_context, regular_gradient_tmp.numpy().astype(np.float32).flatten(), precision=encryption_config[encryption_method]["precision"], obfuscation=True, num_cores=num_cores) elif encryption_method == "plain": enc_regular_gradient_tmp = regular_gradient_tmp.numpy().astype(np.float32).flatten() if encryption_method == "ckks": gradient_trainer_w = enc_d.matmul(x_batch[0].numpy()) + enc_regular_gradient_tmp else: gradient_trainer_w = np.matmul(enc_d.reshape(1, len(enc_d)), x_batch[0].numpy() ) + enc_regular_gradient_tmp # add noise to encrypted gradients and send to assist_trainer if encryption_method == "ckks": logger.info("Calculate noised gradient for trainer.") noise = np.array([rng.randint(1 << 24, 1 << 26) - (1 << 25) for _ in range(x_batch[0].shape[1])], dtype=np.float32) noise /= 100000 noised_gradient_trainer_w = gradient_trainer_w + noise logger.info("Send noised gradient to assist_trainer.") self.dual_channels["gradients_loss"].send(noised_gradient_trainer_w.serialize(), use_pickle=False) # receive decrypted gradient from assist_trainer logger.info("Receive decrypted gradient from assist_trainer.") noised_gradient_trainer_w = self.dual_channels["gradients_loss"].recv() gradient_trainer_w = noised_gradient_trainer_w - noise elif encryption_method == "paillier": logger.info("Calculate noised gradient for trainer.") noise = np.array([rng.randint(1 << 24, 1 << 26) - (1 << 25) for _ in range(x_batch[0].shape[1])], dtype=np.float32) noise /= 100000 noised_gradient_trainer_w = gradient_trainer_w + noise logger.info("Send noised gradient to assist_trainer.") self.dual_channels["gradients_loss"].send(Paillier.serialize(noised_gradient_trainer_w), use_pickle=False) # receive decrypted gradient from assist_trainer logger.info("Receive decrypted gradient from assist_trainer.") noised_gradient_trainer_w = self.dual_channels["gradients_loss"].recv() gradient_trainer_w = noised_gradient_trainer_w - noise # gradient_trainer_w = torch.FloatTensor(gradient_trainer_w).unsqueeze(-1) # update w and b of trainer gradient_trainer_w = gradient_trainer_w / x_batch[0].shape[0] logger.info("Update weights of trainer.") self.model.linear.weight -= (torch.FloatTensor(gradient_trainer_w) * self.optimizer_config["lr"]) # predict train and val for metrics logger.info("Predict train weights of trainer.") self.predict(self.train_dataloader) logger.info("Predict val weights of trainer.") self.predict(self.val_dataloader) # receive flags early_stop_flag, best_model_flag, patient = self.dual_channels["intermediate_label_trainer"].recv( use_pickle=True) # update best model if best_model_flag: self.best_model = copy.deepcopy(self.model) # if need to save results by epoch if self.save_frequency > 0 and epoch % self.save_frequency == 0: # ModelPreserver.save(save_dir=self.save_dir, # model_name=self.save_model_name, # state_dict=self.model.state_dict(), # epoch=epoch) self.save_model(epoch=epoch) # if early stopping, break if early_stop_flag: break # save model for infer # if not early stopping, save model self.save_model(epoch=None) # send w to label trainer self._save_feature_importance(self.dual_channels["intermediate_label_trainer"]) def save_model(self, epoch: Optional[int] = None): if not epoch: save_model_config(stage_model_config=self.export_conf, save_path=Path(self.save_dir)) if self.save_model_name: ModelIO.save_torch_model( state_dict=self.best_model.state_dict(), save_dir=self.save_dir, model_name=self.save_model_name, meta_dict={}, epoch=epoch ) if self.save_onnx_model_name: ModelIO.save_torch_onnx( model=self.best_model, input_dim=(self.data_dim,), save_dir=self.save_dir, model_name=self.save_onnx_model_name, epoch=epoch ) def _save_feature_importance(self, channel): channel.send((FedNode.node_id, self.best_model.state_dict()["linear.weight"][0])) def check_data(self): dim_channel = BroadcastChannel(name="check_data_com", ids=[FedConfig.node_id] + FedConfig.get_trainer()) dim_channel.send(self.data_dim)
16,611
54.373333
120
py
XFL
XFL-master/python/algorithm/framework/vertical/kmeans/base.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import copy import functools from itertools import chain from pathlib import Path import warnings import numpy as np import pandas as pd import pyspark.pandas as ps import torch from common.utils.config_parser import TrainConfigParser from common.utils.logger import logger warnings.filterwarnings("ignore") class VerticalKmeansBase(TrainConfigParser): def __init__(self, train_conf: dict, label: bool = False, *args, **kwargs): """ init Args: train_conf: label: *args: **kwargs: """ super().__init__(train_conf) self.k = 0 self.max_iter = 0 self.tol = 0.0 self.is_converged = False self.init = "random" self.encryption = "plain" self.label = label self.cluster_centers = [] self.cluster_count_list = [] self.train_features, self.train_label, self.train_ids = None, None, None self._init_data() self._init_config() def _init_data(self): logger.info("init data loader.") if not self.input_trainset: return None input_info = self.input_trainset[0] file_path = str(Path(input_info.get("path"), input_info.get("name"))) type_ = input_info.get("type", "None") if input_info.get("has_id", True): index_col = input_info.get("index_col", 'id') else: index_col = None if input_info.get("has_label", True): label_name = input_info.get("label_name", 'y') self.label = True else: label_name = None self.label = False if type_ == "csv": if self.computing_engine == "local": df = pd.read_csv(file_path, index_col=index_col) elif self.computing_engine == "spark": df = ps.read_csv(file_path, index_col=index_col) else: raise NotImplementedError("Computing engine {} is not supported.".format(self.computing_engine)) else: raise NotImplementedError("Dataset type {} is not supported.".format(type_)) if self.label: feature_cols = [_ for _ in df.columns if _ != label_name] self.train_features = df[feature_cols] if label_name: self.train_label = df[label_name] else: self.train_label = None else: self.train_features = df self.train_ids = df.index def _init_config(self): """ Initialize model parameters Returns: """ params = self.train_info.get("train_params", {}) self.k = params.get("k", 5) self.init = params.get("init", "random") self.max_iter = params.get("max_iter", 20) self.tol = params.get("tol", 1e-5) self.random_seed = params.get("extra_config", {}).get("random_seed", 2022) if self.identity != "assist_trainer": self._check() self.encryption = params.get("encryption") def _check(self): """ Check data and parameters Returns: """ if len(self.train_features) <= 0: raise ValueError("error: empty dataset.") if self.k < 2: raise ValueError("k must be an integer value larger than 1.") elif self.k > len(self.train_features): raise ValueError("k is larger than the size of current data.") @staticmethod def euclid_distance(u, center_list): result = [] for i in range(len(center_list)): result.append(sum(np.square(center_list[i] - u))) return result def distance_table(self, centers): """ Args: centers: cluster centroids Returns: (n * k) tensor, whose [i, j] element is square of the distance of sample i to the centroid j. """ if isinstance(centers, ps.DataFrame): centers = centers.to_numpy() elif isinstance(centers, pd.DataFrame): centers = centers.to_numpy() elif isinstance(centers, list): centers = np.array(centers) n = len(self.train_features) if self.train_features.empty: return d = functools.partial(self.euclid_distance, center_list=centers) dt = self.train_features.apply(d, axis=1) return torch.Tensor(list(chain.from_iterable(dt.to_numpy()))).reshape(n, len(centers)) @staticmethod def distance_between_centers(center_list): cluster_dist_list = [] for i in range(0, len(center_list)): for j in range(0, len(center_list)): if j != i: cluster_dist_list.append(np.sum((np.array(center_list[i]) - np.array(center_list[j])) ** 2)) return torch.Tensor(cluster_dist_list) def calc_centers(self, centers, cluster_result): """ Update cluster centers based on clustering results Args: centers: current center slice cluster_result: result of clustering labels Returns: """ feature_sum = {} feature_count = {} for feature, label in zip(self.train_features.values, cluster_result): if label not in feature_sum: feature_sum[label] = copy.deepcopy(feature) else: feature_sum[label] += feature feature_count[label] = feature_count.get(label, 0) + 1 center_list = [] # for k in centroid_feature_sum: for k in range(self.k): if k not in feature_sum: if isinstance(centers, ps.DataFrame): center_list.append(centers.iloc[k]) elif isinstance(centers, pd.DataFrame): center_list.append(centers.iloc[k]) elif isinstance(centers, list): center_list.append(centers[k]) else: raise NotImplementedError else: count = feature_count[k] center_list.append(feature_sum[k] / count) return center_list def calc_cluster_count(self, cluster_result): """ Args: cluster_result: result of clustering labels Returns: """ feature_count = {} for label in cluster_result: feature_count[label] = feature_count.get(label, 0) + 1 cluster_count_list = [] count_all = len(cluster_result) for k in range(self.k): if k not in feature_count: cluster_count_list.append([k, 0, 0]) else: count = feature_count[k] cluster_count_list.append([k, count, count / count_all]) return cluster_count_list @staticmethod def calc_tolerance(centers, centers_new): """ Calculate convergence metrics Returns: """ if isinstance(centers, ps.DataFrame): centers = centers.to_numpy() elif isinstance(centers, pd.DataFrame): centers = centers.to_numpy() elif isinstance(centers, list): centers = np.array(centers) return np.sum(np.sum((centers - np.array(centers_new)) ** 2, axis=1))
6,602
26.627615
100
py
XFL
XFL-master/python/algorithm/framework/vertical/kmeans/table_agg_otp.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import threading from functools import reduce from itertools import combinations from typing import Dict, OrderedDict, Tuple import numpy as np import pandas as pd from common.communication.gRPC.python.commu import Commu from common.crypto.csprng.drbg import get_drbg_inst from common.crypto.csprng.drbg_base import DRBGBase from common.crypto.key_agreement.diffie_hellman import DiffieHellman from common.crypto.one_time_pad.component import OneTimePadCiphertext, OneTimePadContext, OneTimeKey from common.crypto.one_time_pad.one_time_add import OneTimeAdd from .table_agg_base import TableAggregatorAbstractAssistTrainer, TableAggregatorAbstractTrainer def split_bytes(x: bytes, out_shape: Tuple[int]): if len(out_shape) == 0: return int.from_bytes(x, 'big') elif len(out_shape) == 1: a = len(x) // out_shape[0] return [int.from_bytes(x[a * i: a * (i + 1)], 'big') for i in range(out_shape[0])] else: a = len(x) // out_shape[0] return [split_bytes(x[a * i: a * (i + 1)], out_shape[1:]) for i in range(out_shape[0])] class TableAggregatorOTPTrainer(TableAggregatorAbstractTrainer): def __init__(self, sec_conf: dict, trainer_ids: list, *args, **kwargs) -> None: super().__init__(sec_conf=sec_conf, *args, **kwargs) trainer_pairs = combinations(trainer_ids, 2) # key exchange key_exchange_conf = sec_conf["key_exchange"] df_protocols: Dict[str, DiffieHellman] = {} for _trainer_ids in trainer_pairs: if Commu.node_id in _trainer_ids: df_protocol = DiffieHellman(list(_trainer_ids), key_bitlength=key_exchange_conf['key_bitlength'], optimized=key_exchange_conf["optimized"], channel_name="otp_diffie_hellman") df_protocols[df_protocol.chan.remote_id] = df_protocol entropys: Dict[str, bytes] = {remote_id: None for remote_id in df_protocols} def func(id): entropys[id] = df_protocols[id].exchange(out_bytes=True) thread_list = [] for id in df_protocols: task = threading.Thread(target=func, args=(id,)) thread_list.append(task) for task in thread_list: task.start() for task in thread_list: task.join() csprng_conf = sec_conf["csprng"] self.csprngs: OrderedDict[str, DRBGBase] = OrderedDict() self.is_addition = [] for remote_id in trainer_ids: if remote_id != Commu.node_id: self.csprngs[remote_id] = get_drbg_inst(name=csprng_conf["name"], entropy=entropys[remote_id], method=csprng_conf["method"], nonce=b'', additional_data=b'') self.is_addition.append(Commu.node_id < remote_id) # one-time-pad self.otp_context = OneTimePadContext(modulus_exp=sec_conf["key_bitlength"], data_type=sec_conf["data_type"]) def send(self, table: pd.Series) -> None: """ Args: table: Returns: """ def f(t): return reduce(lambda x, y: x * y, t.shape, 1) * self.otp_context.modulus_exp // 8 if table is None: self.broadcast_chan.send(value=None) return num_bytes_array = list(map(f, table)) csprng_generators = [] for remote_id in self.csprngs: generator = self.csprngs[remote_id].generator(num_bytes=num_bytes_array, additional_data=b'') csprng_generators.append(generator) one_time_key = [] for g in csprng_generators: x = bytearray(next(g)) y = split_bytes(x, table.shape) one_time_key.append(np.array(y)) one_time_key = OneTimeKey(one_time_key, self.otp_context.modulus_exp) encrypted_table = OneTimeAdd.encrypt(context_=self.otp_context, data=table, one_time_key=one_time_key, is_addition=self.is_addition, serialized=False).data self.broadcast_chan.send(value=encrypted_table) class TableAggregatorOTPAssistTrainer(TableAggregatorAbstractAssistTrainer): def __init__(self, sec_conf: dict, *args, **kwargs) -> None: super().__init__(sec_conf=sec_conf, *args, **kwargs) self.otp_context = OneTimePadContext(modulus_exp=sec_conf["key_bitlength"], data_type=sec_conf["data_type"]) def aggregate(self) -> pd.Series: message = self.broadcast_chan.collect() ciphertext = None for table in message: if table is None: continue if ciphertext is None: ciphertext = OneTimePadCiphertext(data=table, context_=self.otp_context) else: ciphertext += OneTimePadCiphertext(data=table, context_=self.otp_context) ret = ciphertext.decode() return ret
6,067
39.18543
100
py
XFL
XFL-master/python/algorithm/framework/vertical/kmeans/table_agg_plain.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import copy import pandas as pd from .table_agg_base import TableAggregatorAbstractAssistTrainer from .table_agg_base import TableAggregatorAbstractTrainer class TableAggregatorPlainTrainer(TableAggregatorAbstractTrainer): def __init__(self, sec_conf: dict, *args, **kwargs) -> None: super().__init__(sec_conf=sec_conf, *args, **kwargs) def send(self, table: pd.Series) -> None: """ Args: table: Returns: """ self.broadcast_chan.send(value=table) class TableAggregatorPlainAssistTrainer(TableAggregatorAbstractAssistTrainer): def __init__(self, sec_conf: dict, *args, **kwargs) -> None: super().__init__(sec_conf=sec_conf, *args, **kwargs) def aggregate(self) -> pd.Series: message = self.broadcast_chan.collect() ret = None for table in message: if table is None: continue if ret is None: ret = copy.deepcopy(table) else: ret += table return ret
1,684
29.636364
78
py
XFL
XFL-master/python/algorithm/framework/vertical/kmeans/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from typing import Dict import numpy as np from service.fed_config import FedConfig from service.fed_node import FedNode from common.communication.gRPC.python.channel import DualChannel from .trainer import VerticalKmeansTrainer class VerticalKmeansLabelTrainer(VerticalKmeansTrainer): def __init__(self, train_conf: dict, *args, **kwargs): """ Args: train_conf: training parameters *args: **kwargs: """ super().__init__(train_conf, *args, **kwargs) def init_centers(self): init_center_chan: Dict[str, DualChannel] = {} for party_id in FedConfig.get_trainer(): init_center_chan[party_id] = DualChannel( name="init_center_" + party_id, ids=[FedNode.node_id, party_id] ) self.channels["init_center"] = init_center_chan if self.init == "random": center_ids = list(np.random.choice(len(self.train_features), self.k, replace=False)) for party_id in FedConfig.get_trainer(): self.channels["init_center"][party_id].send(center_ids) return center_ids elif self.init == "kmeans++": self.channels["init_prob"] = DualChannel(name="init_prob", ids=[FedConfig.get_assist_trainer(), FedNode.node_id]) center_ids = [] while len(center_ids) < self.k: m = len(self.train_features) if len(center_ids) < 1: center_ids.append(np.random.choice(m)) else: dist_table = self.distance_table(self.train_features.iloc[center_ids]) self.table_agg_executor.send(dist_table) p = self.channels["init_prob"].recv() p[center_ids] = 0 p = np.array(p / p.sum()) center_ids.append(np.random.choice(m, p=p)) for party_id in FedConfig.get_trainer(): self.channels["init_center"][party_id].send(center_ids) return center_ids
2,731
39.776119
107
py
XFL
XFL-master/python/algorithm/framework/vertical/kmeans/assist_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from typing import Dict import numpy as np from common.communication.gRPC.python.channel import DualChannel from common.evaluation.metrics import ClusteringMetric from common.utils.logger import logger from service.fed_config import FedConfig from service.fed_node import FedNode from service.fed_control import ProgressCalculator from .api import get_table_agg_scheduler_inst from .base import VerticalKmeansBase class VerticalKmeansAssistTrainer(VerticalKmeansBase): def __init__(self, train_conf: dict, *args, **kwargs): """ Args: train_conf: *args: **kwargs: """ self.channels = {} self.party_id_list = FedConfig.get_label_trainer() + FedConfig.get_trainer() sync_chann: Dict[str, DualChannel] = {} for party_id in self.party_id_list: sync_chann[party_id] = DualChannel( name="sync_" + party_id, ids=[FedNode.node_id, party_id]) self.channels["sync"] = sync_chann conf = self._sync_config() train_conf.update(conf) super().__init__(train_conf, label=True, *args, **kwargs) self.progress_calculator = ProgressCalculator(self.max_iter) self.DBI = None self.dist_sum = None cluster_res_chan: Dict[str, DualChannel] = {} tolerance_chan: Dict[str, DualChannel] = {} converged_flag_chan: Dict[str, DualChannel] = {} check_data_com: Dict[str, DualChannel] = {} for party_id in self.party_id_list: cluster_res_chan[party_id] = DualChannel( name="cluster_res_" + party_id, ids=[FedNode.node_id, party_id]) tolerance_chan[party_id] = DualChannel( name="tolerance_" + party_id, ids=[FedNode.node_id, party_id]) converged_flag_chan[party_id] = DualChannel( name="converged_flag_" + party_id, ids=[FedNode.node_id, party_id] ) check_data_com[party_id] = DualChannel( name="check_data_com_" + party_id, ids=[FedNode.node_id, party_id] ) self.channels["cluster_result"] = cluster_res_chan self.channels["tolerance"] = tolerance_chan self.channels["converged_flag"] = converged_flag_chan self.channels["check_data_com"] = check_data_com self.dist_table_agg_executor = get_table_agg_scheduler_inst( sec_conf=self.encryption, trainer_ids=FedConfig.get_label_trainer() + FedConfig.get_trainer() ) def _sync_config(self): def count_key(conf): if isinstance(conf, dict): num = len(conf.keys()) for k, v in conf.items(): num += count_key(v) return num else: return 0 conf_to_update = {} max_key_num = 0 for party_id in self.party_id_list: conf = self.channels["sync"][party_id].recv() num = count_key(conf) if num >= max_key_num: conf_to_update = conf max_key_num = num for party_id in self.party_id_list: self.channels["sync"][party_id].send(conf_to_update) return conf_to_update def init_centers(self): if self.init == "kmeans++": self.channels["init_prob"] = DualChannel(name="init_prob", ids=[FedNode.node_id] + FedConfig.get_label_trainer()) for i in range(1, self.k): dist_sum = self.dist_table_agg_executor.aggregate() p = (dist_sum ** 0.5).sum(axis=1) self.channels["init_prob"].send(p) @staticmethod def get_cluster(dist_sum): """ Assign clustering results Args: dist_sum: Returns: """ labels = np.argmin(dist_sum, axis=1) return np.array(labels) def check_data(self): lens, n_features = None, 0 for party_id in self.party_id_list: recv = self.channels["check_data_com"][party_id].recv() m, n = recv if lens is None: lens = m elif lens != m: raise ValueError("Lengths of the train set mismatched.") n_features += n if n_features <= 0: raise ValueError("Number of the feature is zero. Stop training.") def fit(self): logger.info("vertical K-means scheduler training start.") self.check_data() self.init_centers() logger.info("{}::initialized centers.".format(self.identity)) for iter_ in range(self.max_iter): if iter_ <= 0: self.dist_sum = self.dist_table_agg_executor.aggregate() cluster_result = self.get_cluster(self.dist_sum) for party_id in self.party_id_list: self.channels["cluster_result"][party_id].send(cluster_result) self.cluster_count_list = self.calc_cluster_count(cluster_result) tol_list = [] for party_id in self.party_id_list: tol_list.append(self.channels["tolerance"][party_id].recv()) tol_sum = sum(tol_list) logger.info("iter: {}, tol: {}.".format(iter_, tol_sum)) self.is_converged = True if tol_sum < self.tol else False for party_id in self.party_id_list: self.channels["converged_flag"][party_id].send( self.is_converged) self.dist_sum = self.dist_table_agg_executor.aggregate() self._calc_metrics(self.dist_sum, cluster_result, iter_) # calculate and update the progress of the training self.progress_calculator.cal_custom_progress(iter_ + 1) if self.is_converged: if iter_ + 1 < self.max_iter: # update the progress of 100 to show the training is finished ProgressCalculator.finish_progress() break def _calc_metrics(self, dist_sum, cluster_result, epoch): self.calc_dbi(dist_sum, cluster_result, epoch) def calc_dbi(self, dist_sum, cluster_result, epoch): dist_table = self.calc_ave_dist(dist_sum, cluster_result) if len(dist_table) == 1: raise ValueError( "DBI calculation error: All data are clustered into one group.") center_dist = self.dist_table_agg_executor.aggregate() cluster_avg_intra_dist = [] for i in range(len(dist_table)): cluster_avg_intra_dist.append(dist_table[i][2]) self.DBI = ClusteringMetric.calc_dbi( cluster_avg_intra_dist, center_dist) logger.info("epoch %d: dbi score=%.6g" % (epoch, self.DBI)) def calc_ave_dist(self, dist_sum, cluster_result): distances_centers = dict() for vec, label in zip(dist_sum, cluster_result): if label not in distances_centers: distances_centers[label] = np.sqrt(vec[label]) else: distances_centers[label] += np.sqrt(vec[label]) calc_ave_dist_list = [] for label in range(len(self.cluster_count_list)): count = self.cluster_count_list[label][1] if label not in distances_centers: calc_ave_dist_list.append([label, count, np.nan]) else: calc_ave_dist_list.append( [label, count, distances_centers[label] / count]) return calc_ave_dist_list
8,199
36.788018
107
py
XFL
XFL-master/python/algorithm/framework/vertical/kmeans/api.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from typing import Union from .table_agg_otp import TableAggregatorOTPAssistTrainer, TableAggregatorOTPTrainer from .table_agg_plain import TableAggregatorPlainAssistTrainer, TableAggregatorPlainTrainer def _get_table_agg_inst(role: str, sec_conf: dict, *args, **kwargs) -> Union[TableAggregatorPlainTrainer, TableAggregatorPlainAssistTrainer]: """ get a proper TableAggregator instance. role: "trainer" or "scheduler" """ if sec_conf is None or not sec_conf: # sec_conf may be None or {} method = "plain" sec_conf = { "plain": {} } else: method = list(sec_conf.keys())[0] opt = { "otp": { "trainer": TableAggregatorOTPTrainer, "scheduler": TableAggregatorOTPAssistTrainer }, "plain": { "trainer": TableAggregatorPlainTrainer, "scheduler": TableAggregatorPlainAssistTrainer } } try: return opt[method][role](sec_conf[method], *args, **kwargs) except KeyError as e: raise KeyError("Combination of method {} and role {} is not supported " "for creating TableAggregator instance".format(method, role)) from e except Exception as e: raise e def get_table_agg_scheduler_inst(sec_conf: dict, *args, **kwargs) -> Union[TableAggregatorPlainAssistTrainer, TableAggregatorOTPAssistTrainer]: return _get_table_agg_inst('scheduler', sec_conf, *args, **kwargs) def get_table_agg_trainer_inst(sec_conf: dict, *args, **kwargs) -> Union[TableAggregatorPlainTrainer, TableAggregatorOTPTrainer]: return _get_table_agg_inst('trainer', sec_conf, *args, **kwargs)
2,500
38.698413
112
py
XFL
XFL-master/python/algorithm/framework/vertical/kmeans/table_agg_base.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import abc import pandas as pd from common.communication.gRPC.python.channel import BroadcastChannel from common.communication.gRPC.python.commu import Commu from service.fed_config import FedConfig class TableAggregatorAbstractTrainer(object): __metaclass__ = abc.ABCMeta def __init__(self, sec_conf: dict = None, *args, **kwargs) -> None: self.sec_conf = sec_conf self.broadcast_chan = BroadcastChannel(name='table', ids=Commu.trainer_ids, root_id=FedConfig.get_assist_trainer(), auto_offset=True) def send(self, table: pd.Series) -> None: pass class TableAggregatorAbstractAssistTrainer(object): __metaclass__ = abc.ABCMeta def __init__(self, sec_conf: dict = None, *args, **kwargs) -> None: self.sec_conf = sec_conf self.broadcast_chan = BroadcastChannel(name='table', ids=Commu.trainer_ids, root_id=FedConfig.get_assist_trainer(), auto_offset=True) def aggregate(self) -> pd.Series: pass
1,867
35.627451
86
py
XFL
XFL-master/python/algorithm/framework/vertical/kmeans/__init__.py
0
0
0
py
XFL
XFL-master/python/algorithm/framework/vertical/kmeans/trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os from pathlib import Path import numpy as np import pandas as pd from common.checker.matcher import get_matched_config from common.checker.x_types import All from common.communication.gRPC.python.channel import DualChannel from common.utils.logger import logger from common.utils.utils import update_dict from common.utils.model_io import ModelIO from service.fed_config import FedConfig from service.fed_node import FedNode from .api import get_table_agg_trainer_inst from .base import VerticalKmeansBase class VerticalKmeansTrainer(VerticalKmeansBase): def __init__(self, train_conf: dict, *args, **kwargs): """ Args: train_conf: *args: **kwargs: """ self.channels = {} self.channels["sync"] = DualChannel( name="sync_" + FedNode.node_id, ids=[FedConfig.get_assist_trainer(), FedNode.node_id] ) conf = self._sync_config(train_conf) update_dict(train_conf, conf) super().__init__(train_conf, label=False, *args, **kwargs) self.dist_table = None self.cluster_result = None self.local_tol = 0.0 self.channels["cluster_result"] = DualChannel( name="cluster_res_" + FedNode.node_id, ids=[FedConfig.get_assist_trainer(), FedNode.node_id] ) self.channels["tolerance"] = DualChannel( name="tolerance_" + FedNode.node_id, ids=[FedConfig.get_assist_trainer(), FedNode.node_id] ) self.channels["converged_flag"] = DualChannel( name="converged_flag_" + FedNode.node_id, ids=[FedConfig.get_assist_trainer(), FedNode.node_id] ) self.channels["check_data_com"] = DualChannel( name="check_data_com_" + FedNode.node_id, ids=[FedConfig.get_assist_trainer(), FedNode.node_id] ) self.table_agg_executor = get_table_agg_trainer_inst( sec_conf=self.encryption, trainer_ids=FedConfig.get_label_trainer() + FedConfig.get_trainer() ) def _sync_config(self, config): sync_rule = { "train_info": All() } config_to_sync = get_matched_config(config, sync_rule) self.channels["sync"].send(config_to_sync) conf = self.channels["sync"].recv() return conf def init_centers(self): """ Initialize cluster centers Returns: """ self.channels["init_center"] = DualChannel( name="init_center_" + FedNode.node_id, ids=[FedNode.node_id] + FedConfig.get_label_trainer() ) if self.init == "random": center_ids = self.channels["init_center"].recv() return center_ids elif self.init == "kmeans++": center_ids = [] while len(center_ids) < self.k: if len(center_ids) >= 1: dist_table = self.distance_table(self.train_features.iloc[center_ids]) self.table_agg_executor.send(dist_table) center_ids = self.channels["init_center"].recv() return center_ids def check_data(self): m, n = len(self.train_ids), len(self.train_features.columns) self.channels["check_data_com"].send((m, n)) def fit(self): logger.info("vertical K-means trainer start.") self.check_data() np.random.seed(self.random_seed) center_ids = self.init_centers() logger.info("{}::initialized centers.".format(self.identity)) self.cluster_centers = self.train_features.iloc[center_ids] iter_ = 0 for iter_ in range(self.max_iter): if iter_ <= 0: self.dist_table = self.distance_table(self.cluster_centers) self.table_agg_executor.send(self.dist_table) self.cluster_result = self.channels["cluster_result"].recv() centers = self.calc_centers( self.cluster_centers, self.cluster_result) self.local_tol = self.calc_tolerance(self.cluster_centers, centers) self.channels["tolerance"].send(self.local_tol) self.is_converged = self.channels["converged_flag"].recv() self.cluster_centers = centers self.dist_table = self.distance_table(self.cluster_centers) self.table_agg_executor.send(self.dist_table) center_dist = self.distance_between_centers(self.cluster_centers) self.table_agg_executor.send(center_dist) if self.is_converged: break self.save(epoch=iter_, final=True) def save(self, epoch: int = None, final: bool = False): """ Args: epoch: final: Returns: """ save_dir = str(Path(self.output.get("path"))) if not os.path.exists(save_dir): os.makedirs(save_dir) model_name = self.output.get("model", {}).get("name", "") cluster_centers = [list(i) for i in self.cluster_centers] kmeans_output = { "k": self.k, "iter": epoch, "is_converged": self.is_converged, "tol": self.tol, "cluster_centers": cluster_centers, } ModelIO.save_json_model( model_dict=kmeans_output, save_dir=save_dir, model_name=model_name, ) result_dataframe = pd.DataFrame( { "id": self.train_ids.to_numpy(), "cluster_result": self.cluster_result } ) result_name = self.output.get("result", {}).get("name", "") result_path = Path(save_dir, result_name) result_dataframe.to_csv(result_path, header=True, index=False) logger.info("result saved as: {}.".format(result_path)) summary_df = result_dataframe.groupby( "cluster_result").size().to_frame("count") summary_df = summary_df.reset_index() summary_name = self.output.get("summary", {}).get("name", "") summary_path = Path(save_dir, summary_name) summary_df.to_csv(summary_path, header=True, index=False) logger.info("summary info saved to: {}.".format(summary_path))
5,876
30.260638
98
py
XFL
XFL-master/python/algorithm/framework/vertical/xgboost/base.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os from typing import Any import numpy as np import pandas as pd from algorithm.core.data_io import CsvReader, NdarrayIterator from algorithm.core.encryption_param import get_encryption_param from algorithm.core.tree.cat_param_parser import parse_category_param from algorithm.core.tree.tree_param import XGBTreeParam from algorithm.framework.vertical.vertical_model_base import VerticalModelBase from common.utils.logger import logger class VerticalXgboostBase(VerticalModelBase): def __init__(self, train_conf: dict, is_label_trainer: bool = False, *args, **kwargs): super().__init__(train_conf) self.train_conf = train_conf self.train_features, self.train_label, self.train_ids = None, None, None self.val_features, self.val_label, self.val_ids = None, None, None self.test_features, self.test_label, self.test_ids = None, None, None self.xgb_config = None self.is_label_trainer = is_label_trainer self.feature_importances_ = {} self.__init_xgb_config() self.__init_data() self.__convert_to_binned_data() def __init_data(self) -> None: """ Init data, include features and label. Returns: None """ self.bs = self.train_params.get("batch_size_val") if self.input_trainset: _ = self.__load_data(self.input_trainset) self.train_features, self.train_label, self.train_ids, self.train_names = _ self.train_dataset = NdarrayIterator(self.train_features.to_numpy(), self.bs) else: self.train_dataset = None if self.input_valset: _ = self.__load_data(self.input_valset) self.val_features, self.val_label, self.val_ids, self.val_names = _ self.val_dataset = NdarrayIterator(self.val_features.to_numpy(), self.bs) else: self.val_dataset = None if self.input_testset: _ = self.__load_data(self.input_testset) self.test_features, self.test_label, self.test_ids, self.test_names = _ self.test_dataset = NdarrayIterator(self.test_features.to_numpy(), self.bs) else: self.test_dataset = None def __convert_to_binned_data(self): ''' Note self.train_features will be converted to binned feature ''' cat_columns = parse_category_param(self.train_features, col_index=self.xgb_config.cat_col_index, col_names=self.xgb_config.cat_col_names, col_index_type=self.xgb_config.cat_col_index_type, col_names_type=self.xgb_config.cat_col_names_type, max_num_value=self.xgb_config.cat_max_num_value, max_num_value_type=self.xgb_config.cat_max_num_value_type) self.cat_columns = cat_columns self.cat_feature_names = [] if len(cat_columns) > 0: self.cat_feature_names = self.train_features.columns[cat_columns].to_list() self.train_features[self.cat_feature_names] = self.train_features[self.cat_feature_names].astype('category') def f(x): if self.train_features[x].dtypes == "category": value_counts = self.train_features[x].value_counts() # descending order if value_counts.shape[0] > self.xgb_config.num_bins: values = value_counts.index.to_list() list_unique = values[:self.xgb_config.num_bins - 1] list_group = values[self.xgb_config.num_bins - 1:] uniques = np.array(list_unique + [list_group], dtype=object) value_map = {v: i for i, v in enumerate(list_unique)} value_map.update({v: len(list_unique) for v in list_group}) codes = self.train_features[x].map(value_map) else: codes, uniques = pd.factorize(self.train_features[x], na_sentinel=0) # na_sentinel will not be activated actually uniques = uniques.to_numpy() # uniques: array of values that belongs to the same category # codes: binned values return pd.Series(codes, name=x), uniques.tolist() else: binned_values, split_points = pd.cut(self.train_features[x], bins=self.xgb_config.num_bins, retbins=True, labels=range(self.xgb_config.num_bins)) return binned_values, split_points if self.input_trainset: out = pd.Series(self.train_features.columns).apply(f) if self.xgb_config.num_bins <= 256: dtype = np.uint8 elif self.xgb_config.num_bins <= 2 ** 16: dtype = np.uint16 else: dtype = np.uint32 self.train_features = pd.DataFrame([out[i][0] for i in range(len(out))], dtype=dtype).T # For continuous features, self.split_points stores the split points between bins, for example, 15 split points for 16 bins. # For categorial features, self.split_points stores original values correspond to the bin values, for example, 16 values for 16 bins. self.split_points = [out[i][1][1:-1] if i not in self.cat_columns else out[i][1][:] for i in range(len(out))] def __load_data(self, config): """ Load data from dataset config. Args: config: Dataset config. Returns: [CsvReader, ...] """ if len(config) > 1: logger.warning("More than one dataset is not supported.") if not config: return None, None, None, None config = config[0] if config["type"] == "csv": path = os.path.join(config["path"], config["name"]) if not path: return None, None, None data_reader = CsvReader(path, has_id=config["has_id"], has_label=config["has_label"]) features = data_reader.features(type="pandas.dataframe") features.replace({np.nan: 0, self.xgb_config.missing_value: 0}, inplace=True) ids = data_reader.ids names = data_reader.feature_names() if self.is_label_trainer: labels = data_reader.label() else: labels = None else: raise NotImplementedError("Dataset type {} is not supported.".format(config["type"])) return features, labels, ids, names def col_sample(self) -> tuple[Any, dict]: col_size = self.train_features.shape[1] if 0 < self.xgb_config.subsample_feature_rate <= 1: sample_num = int(col_size * self.xgb_config.subsample_feature_rate) else: sample_num = col_size sampled_idx = np.sort(np.random.choice(col_size, sample_num, replace=False)) feature_id_mapping = {a: b for a, b in enumerate(sampled_idx)} sampled_features = self.train_features.iloc[:, sampled_idx] return sampled_features, feature_id_mapping def __init_xgb_config(self) -> None: """ Init xgboost config. Returns: None """ default_config = self.train_info.get("train_params") cat_params = default_config.get("category", {}).get("cat_features", {}) encryption_methods = list(default_config.get("encryption", {}).keys()) if len(encryption_methods) > 0: encryption_method = encryption_methods[0] else: encryption_method = "plain" encryption_params = default_config.get("encryption", {"plain": {}})[encryption_method] downsampling_params = default_config.get("downsampling", {}) self.xgb_config = XGBTreeParam(loss_param=default_config.get("lossfunc"), # ("BCEWithLogitsLoss"), num_trees=default_config.get("num_trees"), learning_rate=default_config.get("learning_rate"), gamma=default_config.get("gamma"), lambda_=default_config.get("lambda_"), max_depth=default_config.get("max_depth"), num_bins=default_config.get("num_bins", 16), min_split_gain=default_config.get("min_split_gain"), min_sample_split=default_config.get("min_sample_split"), min_leaf_node=default_config.get("min_leaf_node"), feature_importance_type=default_config.get("feature_importance_type"), run_goss=downsampling_params.get("row", {}).get("run_goss", False), top_rate=downsampling_params.get("row", {}).get("top_rate"), other_rate=downsampling_params.get("row", {}).get("other_rate"), metrics=default_config.get("metric"), early_stopping_param=default_config.get("early_stopping", {"patience": -1, "key": "ks", "delta": 0.001}), encryption_param=get_encryption_param(encryption_method, encryption_params), subsample_feature_rate=downsampling_params.get("column", {}).get("rate", 1.0), missing_value=float('inf'), max_num_cores=default_config.get("max_num_cores", 999), col_batch=default_config.get("advanced", {}).get("col_batch", 64), row_batch=default_config.get("advanced", {}).get("row_batch", 40000), cat_col_index=cat_params.get("col_index", ""), cat_col_names=cat_params.get("col_names", []), cat_max_num_value=cat_params.get("max_num_value", 0), cat_col_index_type=cat_params.get("col_index_type", "inclusive"), cat_col_names_type=cat_params.get("col_names_type", "inclusive"), cat_max_num_value_type=cat_params.get("max_num_value_type", "union"), cat_smooth=default_config.get("category", {}).get("cat_smooth", 1.0))
11,631
51.633484
161
py
XFL
XFL-master/python/algorithm/framework/vertical/xgboost/decision_tree_label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import gc from typing import Dict, List, Optional, Union import numpy as np import pandas as pd from pathos.pools import ThreadPool from algorithm.core.encryption_param import PaillierParam, PlainParam from algorithm.core.paillier_acceleration import embed, umbed from algorithm.core.tree.big_feature import Feature from algorithm.core.tree.feature_importance import FeatureImportance from algorithm.core.tree.gain_calc import BestSplitInfo, cal_cat_rank, cal_gain, cal_weight from algorithm.core.tree.goss import Goss from algorithm.core.tree.tree_param import XGBTreeParam from algorithm.core.tree.tree_structure import Node, SplitInfo, Tree from algorithm.core.tree.xgboost_loss import get_xgb_loss_inst from common.communication.gRPC.python.channel import BroadcastChannel, DualChannel from common.crypto.paillier.paillier import Paillier, PaillierContext from common.crypto.paillier.utils import get_core_num from common.utils.constants import PAILLIER, PLAIN from common.utils.logger import logger from service.fed_config import FedConfig from service.fed_control import ProgressCalculator from .debug_params import EMBEDING from service.fed_node import FedNode class VerticalDecisionTreeLabelTrainer(object): def __init__(self, tree_param: XGBTreeParam, y: np.ndarray, y_pred: np.ndarray, features: pd.DataFrame, cat_columns: list, split_points: np.ndarray, channels: Dict[str, Union[BroadcastChannel, DualChannel]], encryption_context: Optional[PaillierContext] = None, feature_id_mapping: Optional[Dict[int, int]] = None, tree_index: Optional[int] = None): logger.info( f"Label trainer decision tree {tree_index} initialize start.") if tree_param.encryption_param.method not in [PAILLIER, PLAIN]: raise ValueError( f"Encryption method {tree_param.encryption_param.method} not supported.") self.tree_param = tree_param self.progress_calculator = ProgressCalculator(self.tree_param.num_trees, self.tree_param.max_depth) self.y = y self.y_pred = y_pred self.cat_columns = cat_columns self.split_points = split_points self.party_id = FedConfig.node_id self.max_num_cores = get_core_num(tree_param.max_num_cores) self.tree_index = tree_index loss_inst = get_xgb_loss_inst(list(self.tree_param.loss_param.keys())[0]) self.grad = loss_inst.cal_grad( self.y, self.y_pred, after_prediction=True) if tree_param.run_goss: goss = Goss(tree_param.top_rate, tree_param.other_rate) self.goss_selected_idx = goss.sampling(self.grad) hess = loss_inst.cal_hess( self.y[self.goss_selected_idx], self.y_pred[self.goss_selected_idx], after_prediction=True) self.hess = np.zeros_like(self.grad) self.hess[self.goss_selected_idx] = hess goss.update_gradients(self.grad, self.hess) else: self.hess = loss_inst.cal_hess(self.y, self.y_pred, after_prediction=True) self.goss_selected_idx = range(self.y.shape[0]) sample_index = self.goss_selected_idx self.individual_grad_hess: BroadcastChannel = channels["individual_grad_hess"] self.tree_node_chann: BroadcastChannel = channels["tree_node"] self.summed_grad_hess_channs: Dict[str, DualChannel] = channels["summed_grad_hess"] self.min_split_info_channs: Dict[str, DualChannel] = channels["min_split_info"] self.sample_index_after_split_channs: Dict[str, DualChannel] = channels["sample_index_after_split"] encryption_param = self.tree_param.encryption_param self.pri_context = encryption_context self.feature_importance = {} self.feature_importance_type = tree_param.feature_importance_type self.feature_id_mapping = feature_id_mapping if isinstance(encryption_param, PlainParam): self.individual_grad_hess.broadcast( [self.grad[sample_index], self.hess[sample_index]], use_pickle=True) elif isinstance(encryption_param, PaillierParam): num_cores = self.max_num_cores if encryption_param.parallelize_on else 1 if EMBEDING: grad_hess = embed([self.grad[sample_index], self.hess[sample_index]], interval=(1 << 128), precision=64) enc_grad_hess = Paillier.encrypt(context=self.pri_context, data=grad_hess, precision=0, # must be 0 obfuscation=True, num_cores=num_cores) self.individual_grad_hess.broadcast(Paillier.serialize(enc_grad_hess, compression=False), use_pickle=True) else: enc_grad = Paillier.encrypt(context=self.pri_context, data=self.grad[sample_index], precision=encryption_param.precision, obfuscation=True, num_cores=num_cores) enc_hess = Paillier.encrypt(context=self.pri_context, data=self.hess[sample_index], precision=encryption_param.precision, obfuscation=True, num_cores=num_cores) self.individual_grad_hess.broadcast( [Paillier.serialize(enc_grad, compression=False), Paillier.serialize(enc_hess, compression=False)], use_pickle=True) else: raise ValueError("Encryption param not supported.") if features.shape[1] == 0: self.big_feature = None else: self.big_feature = Feature.create(values=features.iloc[sample_index, :], sample_index=sample_index, grad=self.grad[sample_index], hess=self.hess[sample_index]) logger.info( f"Label trainer decision tree {tree_index} initialize finished.") def _cal_local_best_split(self, node: Node): best_split_info = BestSplitInfo(feature_ower=self.party_id) if node.sample_index is None or len(node.sample_index) == self.big_feature.data.shape[0]: big_feature = self.big_feature else: big_feature = self.big_feature.slice_by_sample_index(node.sample_index) res_hist_list = [] for col_name in big_feature.feature_columns: res_hist_list.append(big_feature.data.groupby([col_name])[['xfl_grad', 'xfl_hess']].agg({'sum'})) # ({'count', 'sum'}) # for categorial features, resort # cat column is count from the first col of cat feature for feature_idx in self.cat_columns: cat_rank = cal_cat_rank(res_hist_list[feature_idx][('xfl_grad', 'sum')], res_hist_list[feature_idx][('xfl_hess', 'sum')], self.tree_param.cat_smooth) cat_rank.sort_values(inplace=True) # index is saved in the Series's index res_hist_list[feature_idx] = res_hist_list[feature_idx].loc[cat_rank.index.to_list()] for feature_idx in range(len(res_hist_list)): res_hist_list[feature_idx] = res_hist_list[feature_idx].cumsum(axis=0) res_hist_list[feature_idx].rename(columns={"sum": "cum_sum"}, inplace=True) cum_grad = res_hist_list[feature_idx][('xfl_grad', 'cum_sum')].to_numpy() cum_hess = res_hist_list[feature_idx][('xfl_hess', 'cum_sum')].to_numpy() gains = cal_gain(cum_grad, cum_hess, self.tree_param.lambda_) if len(gains) == 1 and gains[0] == -np.inf: continue max_gain_index = np.argmax(gains) max_gain = gains[max_gain_index].item() if max_gain > best_split_info.gain: best_split_info.gain = max_gain best_split_info.feature_owner = self.party_id best_split_info.feature_idx = self.feature_id_mapping[feature_idx].item() # For categorial feature, split_point stores categories in left child branch if feature_idx in self.cat_columns: # It is not much precise if some categorial values are not be sampled left_cat = res_hist_list[feature_idx].index.to_list()[:max_gain_index + 1] best_split_info.left_cat = [] for cat in left_cat: ori_cat = self.split_points[feature_idx][cat] if isinstance(ori_cat, list): best_split_info.left_cat += ori_cat else: best_split_info.left_cat.append(ori_cat) best_split_info.split_point = None best_split_info.is_category = True filter = big_feature.data.iloc[:, feature_idx + 3].isin(left_cat) else: # Because of sampling max_split_index = int(res_hist_list[feature_idx][('xfl_grad', 'cum_sum')].index[max_gain_index]) max_split_index = min(max_split_index, len(self.split_points[feature_idx]) - 1) best_split_info.split_point = self.split_points[feature_idx][max_split_index] best_split_info.left_cat = None best_split_info.is_category = False filter = big_feature.data.iloc[:, feature_idx + 3] <= max_split_index best_split_info.left_sample_index = big_feature.data[filter]['xfl_id'].tolist() best_split_info.right_sample_index = big_feature.data[~filter]['xfl_id'].tolist() left_weight = cal_weight(cum_grad[max_gain_index], cum_hess[max_gain_index], self.tree_param.lambda_).item() right_weight = cal_weight(cum_grad[-1] - cum_grad[max_gain_index], cum_hess[-1] - cum_hess[max_gain_index], self.tree_param.lambda_).item() best_split_info.left_bin_weight = left_weight best_split_info.right_bin_weight = right_weight best_split_info.num_left_bin = len(best_split_info.left_sample_index) best_split_info.num_right_bin = len(best_split_info.right_sample_index) best_split_info.max_gain_index = max_gain_index # only valid for continuous feature return best_split_info def _cal_remote_best_split(self) -> Dict[str, BestSplitInfo]: best_split_info_dict: Dict[str, BestSplitInfo] = { party_id: BestSplitInfo(feature_ower=party_id) for party_id in self.summed_grad_hess_channs } gain_infos: Dict[str, list] = { party_id: [] for party_id in self.summed_grad_hess_channs } is_continue_flags = np.array([True for party_id in self.summed_grad_hess_channs], dtype=bool) def decrypt_hist(hist_list: List[np.ndarray], num_cores: int, out_origin: bool = True) -> list: len_list = [len(item) for item in hist_list] cum_len = np.cumsum([0] + len_list) hist = np.concatenate(hist_list) hist = Paillier.decrypt(self.pri_context, hist, num_cores=num_cores, out_origin=out_origin) res = [] for i in range(len(cum_len) - 1): res.append(hist[cum_len[i]: cum_len[i + 1]]) return res while True: for i, party_id in enumerate(self.summed_grad_hess_channs): if not is_continue_flags[i]: continue data = self.summed_grad_hess_channs[party_id].recv(use_pickle=True, wait=False) if data is None: # Data has not been send, try it next round. continue is_continue, grad_hess_hist_list, remote_cat_index = data if self.tree_param.encryption_param.method == PAILLIER: if EMBEDING: grad_hess_hist = [] count_hist_list = [] for item in grad_hess_hist_list: grad_hess_hist.append(item[0]) count_hist_list.append(item[1]) grad_hess_hist = decrypt_hist(grad_hess_hist, num_cores=self.max_num_cores, out_origin=True) grad_hist_list = [] hess_hist_list = [] for hist in grad_hess_hist: a, b = umbed(hist, num=2, interval=(1 << 128), precison=64) grad_hist_list.append(a) hess_hist_list.append(b) else: grad_hist_list = [] hess_hist_list = [] count_hist_list = [] for item in grad_hess_hist_list: grad_hist_list.append(item[0]) hess_hist_list.append(item[1]) count_hist_list.append(item[2]) grad_hist_list = decrypt_hist(grad_hist_list, num_cores=self.max_num_cores, out_origin=False) hess_hist_list = decrypt_hist(hess_hist_list, num_cores=self.max_num_cores, out_origin=False) else: grad_hist_list = [] hess_hist_list = [] count_hist_list = [] for item in grad_hess_hist_list: grad_hist_list.append(item[0]) hess_hist_list.append(item[1]) count_hist_list.append(item[2]) for idx in range(len(grad_hess_hist_list)): grad_hist, hess_hist, count_hist = \ np.array(grad_hist_list[idx], dtype=np.float32), np.array(hess_hist_list[idx], dtype=np.float32), np.array(count_hist_list[idx]) # for categorial feature, resort if idx in remote_cat_index: cat_rank = cal_cat_rank(grad_hist, hess_hist, self.tree_param.cat_smooth) cat_rank = np.argsort(cat_rank).tolist() grad_hist = grad_hist[cat_rank] hess_hist = hess_hist[cat_rank] count_hist = count_hist[cat_rank] else: cat_rank = [] cum_grad_hist = np.cumsum(grad_hist) cum_hess_hist = np.cumsum(hess_hist) gains = cal_gain(cum_grad_hist, cum_hess_hist, self.tree_param.lambda_) max_gain_index = np.argmax(gains) max_gain = gains[max_gain_index].item() num_left_sample = np.sum(count_hist[:max_gain_index + 1]) num_right_sample = np.sum(count_hist[max_gain_index + 1:]) info = { 'max_gain': max_gain, 'cum_grad': cum_grad_hist, 'cum_hess': cum_hess_hist, 'max_gain_index': max_gain_index, "is_category": idx in remote_cat_index, 'cat_rank': cat_rank, 'num_left_sample': num_left_sample, 'num_right_sample': num_right_sample } gain_infos[party_id].append(info) if not is_continue: is_continue_flags[i] = is_continue # No data will be send later, cal best_split_info best_split_info: BestSplitInfo = best_split_info_dict[party_id] for feature_idx, gain_info in enumerate(gain_infos[party_id]): max_gain = gain_info["max_gain"] cum_grad = gain_info["cum_grad"] cum_hess = gain_info["cum_hess"] max_gain_index = gain_info["max_gain_index"] is_category = gain_info["is_category"] cat_rank = gain_info["cat_rank"] if max_gain > best_split_info.gain: if len(cum_grad) == 1: max_gain_split_index = 0 else: max_gain_split_index = max_gain_index if max_gain > best_split_info.gain: best_split_info.gain = max_gain best_split_info.feature_owner = party_id best_split_info.feature_idx = feature_idx best_split_info.split_point = None # should not know best_split_info.missing_value_on_left = None # need not know best_split_info.left_sample_index = None # get it later best_split_info.right_sample_index = None # get it later best_split_info.num_left_bin = gain_info["num_left_sample"] best_split_info.num_right_bin = gain_info["num_right_sample"] left_weight = cal_weight(cum_grad[max_gain_split_index], cum_hess[max_gain_split_index], self.tree_param.lambda_).item() right_weight = cal_weight(cum_grad[-1] - cum_grad[max_gain_split_index], cum_hess[-1] - cum_hess[max_gain_split_index], self.tree_param.lambda_).item() best_split_info.left_bin_weight = left_weight best_split_info.right_bin_weight = right_weight best_split_info.max_gain_index = max_gain_index # note this is not the final result of the left category best_split_info.left_cat = [] if not cat_rank else cat_rank[:max_gain_index + 1] best_split_info.is_category = is_category flag = np.any(is_continue_flags) if not flag: break gc.collect() return best_split_info_dict def get_feature_importance(self): return self.feature_importance def update_feature_importance(self, split_info: SplitInfo): inc_split, inc_gain = 1, split_info.gain owner_id = split_info.owner_id fid = split_info.feature_idx owner_name = owner_id for node_id in FedNode.config["trainer"]: if owner_id == node_id: owner_name = FedNode.config["trainer"][owner_id]["name"] break if (owner_name, fid) not in self.feature_importance: self.feature_importance[(owner_name, fid)] = FeatureImportance( 0, 0, self.feature_importance_type) self.feature_importance[(owner_name, fid)].add_split(inc_split) if inc_gain is not None: self.feature_importance[(owner_name, fid)].add_gain(inc_gain) def fit(self) -> Tree: tree = Tree(self.party_id, self.tree_index) thread_pool = ThreadPool(2) logger.info(f"Decision tree {self.tree_index} training start..") if self.tree_param.run_goss: tree.root_node.sample_index = self.goss_selected_idx for depth in range(self.tree_param.max_depth): logger.info(f"Decision tree depth {depth} training start..") this_depth_nodes = tree.search_nodes(depth) for node in this_depth_nodes: logger.info(f"Depth {depth} - node {node.id} start training.") self.tree_node_chann.broadcast(node, use_pickle=True) best_split_info_dict: Dict[str, BestSplitInfo] = {} logger.info("Calculating local best split..") if self.big_feature is not None: res1 = thread_pool.apipe(self._cal_local_best_split, node) logger.info("Calculating remote best split..") res2 = thread_pool.apipe(self._cal_remote_best_split) if self.big_feature is not None: best_split_info_dict[self.party_id] = res1.get() logger.info("Calculating local best split done.") best_split_info_dict_remote = res2.get() logger.info("Calculating remote best split done.") best_split_info_dict.update(best_split_info_dict_remote) party_ids = list(best_split_info_dict.keys()) best_split_party_id = party_ids[ np.argmax( [best_split_info_dict[party_id].gain for party_id in party_ids]) ] best_split_info = best_split_info_dict[best_split_party_id] if best_split_info.gain < self.tree_param.min_split_gain or \ min(best_split_info.num_left_bin, best_split_info.num_right_bin) < self.tree_param.min_sample_split: for party_id in self.min_split_info_channs: self.min_split_info_channs[party_id].send([-1, -1, -1], use_pickle=True) continue if best_split_info.feature_owner == self.party_id: for party_id in self.min_split_info_channs: self.min_split_info_channs[party_id].send([-1, -1, -1], use_pickle=True) split_info = SplitInfo(owner_id=best_split_info.feature_owner, feature_idx=best_split_info.feature_idx, is_category=best_split_info.is_category, split_point=best_split_info.split_point, left_cat=best_split_info.left_cat, gain=best_split_info.gain) else: for party_id in self.min_split_info_channs: if best_split_info.feature_owner == party_id: self.min_split_info_channs[party_id].send( [best_split_info.feature_idx, best_split_info.max_gain_index, best_split_info.left_cat], use_pickle=True ) else: self.min_split_info_channs[party_id].send([-1, -1, -1], use_pickle=True) split_info = SplitInfo(owner_id=best_split_info.feature_owner, feature_idx=best_split_info.feature_idx, gain=best_split_info.gain) best_split_info.left_sample_index, best_split_info.right_sample_index = \ self.sample_index_after_split_channs[best_split_info.feature_owner].recv( use_pickle=True) left_node_id, right_node_id = tree.split(node_id=node.id, split_info=split_info, left_sample_index=best_split_info.left_sample_index, right_sample_index=best_split_info.right_sample_index, left_sample_weight=best_split_info.left_bin_weight, right_sample_weight=best_split_info.right_bin_weight) node.update_as_non_leaf(split_info=split_info, left_node_id=left_node_id, right_node_id=right_node_id) self.update_feature_importance(split_info) logger.info(f"Depth {depth} - node {node.id} finish training.") gc.collect() # calculate and update the progress of the training self.progress_calculator.cal_custom_progress(self.tree_index, depth+1) self.tree_node_chann.broadcast(None, use_pickle=True) logger.info(f"Decision tree {self.tree_index} training finished") return tree
26,246
50.974257
152
py
XFL
XFL-master/python/algorithm/framework/vertical/xgboost/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import copy import os from pathlib import Path from typing import Dict, Optional import numpy as np import pandas as pd from pathos.pools import ThreadPool from algorithm.core.data_io import NdarrayIterator from algorithm.core.encryption_param import PaillierParam, PlainParam from algorithm.core.tree.tree_structure import BoostingTree, Tree from algorithm.core.tree.xgboost_loss import get_xgb_loss_inst from common.checker.matcher import get_matched_config from common.checker.x_types import All from common.communication.gRPC.python.channel import BroadcastChannel, DualChannel from common.crypto.paillier.paillier import Paillier from common.utils.algo_utils import earlyStopping from common.utils.logger import logger from common.utils.utils import save_model_config from service.fed_config import FedConfig from service.fed_node import FedNode from .base import VerticalXgboostBase from .decision_tree_label_trainer import VerticalDecisionTreeLabelTrainer from service.fed_control import ProgressCalculator from common.utils.model_io import ModelIO class VerticalXgboostLabelTrainer(VerticalXgboostBase): def __init__(self, train_conf: dict, *args, **kwargs): self.channels = dict() self.channels["sync"] = BroadcastChannel(name="sync") self._sync_config(train_conf) super().__init__(train_conf, is_label_trainer=True, *args, **kwargs) self.party_id = FedConfig.node_id self.channels["encryption_context"] = BroadcastChannel( name="encryption_context") self.channels["individual_grad_hess"] = BroadcastChannel( name="individual_grad_hess") self.channels["tree_node"] = BroadcastChannel(name="tree_node") self.channels["check_dataset_com"] = BroadcastChannel( name="check_dataset_com") summed_grad_hess_channs: Dict[str, DualChannel] = {} min_split_info_channs: Dict[str, DualChannel] = {} sample_index_after_split_channs: Dict[str, DualChannel] = {} val_com: Dict[str, DualChannel] = {} restart_com: Dict[str, DualChannel] = {} early_stop_com: Dict[str, DualChannel] = {} for party_id in FedConfig.get_trainer(): summed_grad_hess_channs[party_id] = \ DualChannel(name="summed_grad_hess_" + party_id, ids=[FedConfig.node_id, party_id]) min_split_info_channs[party_id] = \ DualChannel(name="min_split_info_" + party_id, ids=[FedConfig.node_id, party_id]) sample_index_after_split_channs[party_id] = \ DualChannel(name="sample_index_after_split_" + party_id, ids=[FedConfig.node_id, party_id]) val_com[party_id] = \ DualChannel(name="val_com_" + party_id, ids=[FedConfig.node_id, party_id]) restart_com[party_id] = \ DualChannel(name="restart_com_" + party_id, ids=[FedConfig.node_id, party_id]) early_stop_com[party_id] = \ DualChannel(name="early_stop_com_" + party_id, ids=[FedConfig.node_id, party_id]) self.channels["summed_grad_hess"] = summed_grad_hess_channs self.channels["min_split_info"] = min_split_info_channs self.channels["sample_index_after_split"] = sample_index_after_split_channs self.channels["val_com"] = val_com self.channels["restart_com"] = restart_com self.channels["early_stop_com"] = early_stop_com if isinstance(self.xgb_config.encryption_param, (PlainParam, type(None))): self.private_context = None elif isinstance(self.xgb_config.encryption_param, PaillierParam): self.private_context = Paillier.context(self.xgb_config.encryption_param.key_bit_size, self.xgb_config.encryption_param.djn_on) self.public_context = self.private_context.to_public() self.channels["encryption_context"].broadcast( self.public_context.serialize(), use_pickle=False) else: raise TypeError( f"Encryption param type {type(self.xgb_config.encryption_param)} not valid.") self.es = earlyStopping(key=self.xgb_config.early_stopping_param["key"], patience=self.xgb_config.early_stopping_param["patience"], delta=self.xgb_config.early_stopping_param["delta"]) self.best_round = -1 self.best_prediction_val = None self.best_prediction_train = None if self.train_features is not None: input_schema = ','.join([_ for _ in self.train_features.columns if _ not in set(["y", "id"])]) else: input_schema = "" self.export_conf = [{ "class_name": "VerticalXGBooster", "identity": self.identity, "filename": self.output.get("proto_model", {}).get("name", ''), # "filename": self.output.get("proto_model", {"name": "vertical_xgboost_guest.pmodel"})["name"], "input_schema": input_schema, "version": '1.4.0' }] def _sync_config(self, config): sync_rule = { "train_info": { "interaction_params": All(), "train_params": { "lossfunc": All(), "num_trees": All(), "num_bins": All(), "batch_size_val": All(), "downsampling": { "row": { "run_goss": All() } }, "encryption": All() } } } config_to_sync = get_matched_config(config, sync_rule) self.channels["sync"].broadcast(config_to_sync) def fit(self): f_names = self.channels["sync"].collect() self.remote_f_names = {} for name_dict in f_names: self.remote_f_names.update(name_dict) self.check_dataset() boosting_tree = BoostingTree() # train_y_pred_primitive, tree_list = np.zeros_like(self.train_label), [] train_y_pred_primitive = np.zeros_like(self.train_label) val_y_pred_primitive = np.zeros_like(self.val_label) loss_inst = get_xgb_loss_inst( list(self.xgb_config.loss_param.keys())[0]) train_y_pred, val_y_pred = loss_inst.predict( train_y_pred_primitive), loss_inst.predict(val_y_pred_primitive) for tree_idx in range(1, self.xgb_config.num_trees+1): logger.info("Tree {} start training.".format(tree_idx)) # 0: no need to restart, 1: restart, 2: max number of try reached restart_status = 1 while True: # train section sampled_features, feature_id_mapping = self.col_sample() cat_columns_after_sampling = list(filter( lambda x: feature_id_mapping[x] in self.cat_columns, list(feature_id_mapping.keys()))) split_points_after_sampling = [ self.split_points[feature_id_mapping[k]] for k in feature_id_mapping.keys()] trainer = VerticalDecisionTreeLabelTrainer(tree_param=self.xgb_config, y=self.train_label, y_pred=train_y_pred, features=sampled_features, cat_columns=cat_columns_after_sampling, split_points=split_points_after_sampling, channels=self.channels, encryption_context=self.private_context, feature_id_mapping=feature_id_mapping, tree_index=tree_idx) tree = trainer.fit() if not tree.root_node.is_leaf: restart_status = 0 else: if self.xgb_config.early_stopping_param["patience"] <= 0: # if not set patience, terminate immediately restart_status = 2 else: self.es.counter += 1 if self.es.counter == self.xgb_config.early_stopping_param["patience"]: restart_status = 2 for party_id in FedConfig.get_trainer(): self.channels["restart_com"][party_id].send(restart_status) if restart_status != 1: break logger.info(f"label trainer tree {tree_idx} training restart.") if restart_status == 2: logger.info("label trainer early stopped because a tree's root is leaf, best round: {}.".format( self.best_round)) break self.update_feature_importance(trainer.get_feature_importance()) if self.xgb_config.run_goss: train_y_pred_primitive += self.predict_on_tree( tree, self.train_dataset) * self.xgb_config.learning_rate else: for _, node in tree.nodes.items(): if node.is_leaf: train_y_pred_primitive[node.sample_index] += node.weight * \ self.xgb_config.learning_rate train_y_pred = loss_inst.predict(train_y_pred_primitive) if self.interaction_params.get("echo_training_metrics"): train_loss = loss_inst.cal_loss( self.train_label, train_y_pred_primitive, after_prediction=False) self._calc_metrics(self.train_label, train_y_pred, tree_idx, stage="train", loss={ loss_inst.name: train_loss}) tree.clear_training_info() boosting_tree.append(tree=tree, lr=self.xgb_config.learning_rate, max_depth=self.xgb_config.max_depth) logger.info("Tree {} training done.".format(tree_idx)) # validation section logger.info("Validation on tree {} start.".format(tree_idx)) val_y_pred_primitive += self.predict_on_tree( tree, self.val_dataset) * self.xgb_config.learning_rate val_y_pred = loss_inst.predict(val_y_pred_primitive) val_loss = loss_inst.cal_loss( self.val_label, val_y_pred_primitive, after_prediction=False) metric = self._calc_metrics(self.val_label, val_y_pred, tree_idx, stage="val", loss={loss_inst.name: val_loss}) logger.info("Validation on tree {} done.".format(tree_idx)) if self.xgb_config.early_stopping_param["patience"] > 0: early_stop_flag, save_flag = self.es(metric) else: early_stop_flag, save_flag = False, True if save_flag: # self.best_round = tree_idx + 1 self.best_round = tree_idx self.best_prediction_train = copy.deepcopy(train_y_pred) self.best_prediction_val = copy.deepcopy(val_y_pred) for party_id in FedConfig.get_trainer(): self.channels["early_stop_com"][party_id].send(early_stop_flag) if early_stop_flag: logger.info( "label trainer early stopped. best round: {}.".format(self.best_round)) break # if self.interaction_params.get("save_frequency") > 0 and (tree_idx + 1) % self.interaction_params.get("save_frequency") == 0: if self.interaction_params.get("save_frequency") > 0 and tree_idx % self.interaction_params.get("save_frequency") == 0: # self.save(boosting_tree, epoch=tree_idx+1) self.save(boosting_tree, epoch=tree_idx) self._write_prediction( # self.train_label, train_y_pred, self.train_ids, epoch=tree_idx + 1) self.train_label, train_y_pred, self.train_ids, epoch=tree_idx) self._write_prediction( # self.val_label, val_y_pred, self.val_ids, epoch=tree_idx + 1, stage='val') self.val_label, val_y_pred, self.val_ids, epoch=tree_idx, stage='val') # add metrics during training for plot self._write_loss(train_loss, val_loss, tree_idx) # update the progress of 100 to show the training is finished ProgressCalculator.finish_progress() # model preserve if self.xgb_config.early_stopping_param["patience"] <= 0: self.best_round = len(boosting_tree) self.best_prediction_train = copy.deepcopy(train_y_pred) self.best_prediction_val = copy.deepcopy(val_y_pred) else: logger.info("num trees: %d, best: %d" % (len(boosting_tree), self.best_round)) if boosting_tree.trees: logger.info("save") # self.save(boosting_tree, final=True) if self.best_round <= 0: self.best_round = len(boosting_tree) self.save(boosting_tree[:self.best_round], final=True) logger.info('_write_prediction train') self._write_prediction( self.train_label, train_y_pred, self.train_ids, final=True) logger.info('_write_prediction val') self._write_prediction( self.val_label, val_y_pred, self.val_ids, final=True, stage='val') logger.info("Writing roc data...") self._write_roc_data( self.train_label, train_y_pred, self.val_label, val_y_pred) logger.info("Writing ks data...") self._write_ks_data(self.train_label, train_y_pred, self.val_label, val_y_pred) logger.info("Writing lift and gain data...") self._write_lift_gain_data( self.train_label, train_y_pred, self.val_label, val_y_pred ) logger.info("Writing pr curve data...") self._write_pr_data( self.train_label, train_y_pred, self.val_label, val_y_pred) self._write_feature_importance() else: logger.error("Model is none, ture off run_goss (false) and downsampling (1) please.") raise SystemError( "Model is none, ture off run_goss (false) and downsampling (1) please.") def save(self, boosting_tree: BoostingTree, epoch: Optional[int] = None, final: bool = False): if final: save_model_config(stage_model_config=self.export_conf, save_path=self.output.get("path")) save_dir = self.output.get("path") if not os.path.exists(save_dir): os.makedirs(save_dir) # dump out ks plot suggest_threshold = 0.5 # if "ks" in self.xgb_config.metrics or "auc_ks" in self.xgb_config.metrics: # # tc = ThresholdCutter(os.path.join(save_dir, "ks_plot_valid.csv")) # tc = ThresholdCutter(os.path.join( # save_dir, self.output.get("ks_plot_val")["name"])) # # tc.cut_by_value(self.val_label, self.best_prediction_val) # # suggest_threshold = float(tc.bst_threshold) # # tc.save() # if final: # self.val_ks_metrics = tc.metrics # self.val_ks_bst_threshold = tc.bst_threshold # self.val_ks_bst_score = tc.bst_score # if self.interaction_params.get("echo_training_metrics"): # tc = ThresholdCutter(os.path.join( # save_dir, self.output.get("ks_plot_train")["name"])) # # tc.cut_by_value(self.train_label, self.best_prediction_train) # # tc.save() # if final: # self.train_ks_metrics = tc.metrics # self.train_ks_bst_threshold = tc.bst_threshold # self.train_ks_bst_score = tc.bst_score model_name = self.output.get("model", {}).get("name") proto_name = self.output.get("proto_model", {}).get("name") if model_name: # model_dict = boosting_tree[:self.best_round].to_dict( # suggest_threshold, compute_group=True) model_dict = boosting_tree.to_dict(suggest_threshold, compute_group=True) ModelIO.save_json_model(model_dict, save_dir, model_name, epoch=epoch, version='1.4.0') if proto_name: # TODO: temp model_name_list = self.output.get("proto_model")["name"].split(".") name_prefix, name_postfix = ".".join( model_name_list[:-1]), model_name_list[-1] if not final and epoch: new_model_name = name_prefix + \ "_epoch_{}".format(epoch) + "." + name_postfix else: new_model_name = name_prefix + "." + name_postfix model_path = os.path.join(save_dir, new_model_name) # xgb_output = boosting_tree[:self.best_round].to_proto( # suggest_threshold, compute_group=True) xgb_output = boosting_tree.to_proto(suggest_threshold, compute_group=True) with open(model_path, 'wb') as f: f.write(xgb_output) logger.info("model saved as: {}.".format(model_path)) self.make_readable_feature_importance( os.path.join(save_dir, self.output.get("feature_importance")["name"])) def make_readable_feature_importance(self, file_name): with open(file_name, "w") as f: f.write("owner_id,fid,importance\n") normalizer = np.sum([_.get() for _ in self.feature_importances_.values()]) for k, v in sorted(self.feature_importances_.items(), key=lambda d: d[1], reverse=True): f.write("%s,%s,%.6g\n" % (k[0], k[1], v.get() / normalizer)) def _make_indicator_for_prediction(self, tree: Tree, feature: np.ndarray): indicator = {} for node_id, node in tree.nodes.items(): if not node.is_leaf and node.split_info.owner_id == self.party_id: feature_idx = node.split_info.feature_idx data = feature[:, feature_idx] if node.split_info.is_category: indicator[node_id] = np.isin( data, node.split_info.left_cat) else: indicator[node_id] = (data <= node.split_info.split_point) return indicator def _gen_prediction(self, tree: Tree, indicator: Dict[str, np.ndarray], feature: np.ndarray): prediction = np.zeros((len(feature),), dtype=np.float32) depth = 0 sample_in_node = {} while True: node_list = tree.search_nodes(depth) if not node_list: break for node in node_list: if node.is_leaf: prediction[sample_in_node[node.id]] = node.weight else: if depth == 0: sample_in_node[node.left_node_id] = np.where( indicator[node.id] == 1)[0] sample_in_node[node.right_node_id] = np.where( indicator[node.id] == 0)[0] else: sample_in_node[node.left_node_id] = np.intersect1d( sample_in_node[node.id], np.where(indicator[node.id] == 1)[0]) sample_in_node[node.right_node_id] = np.intersect1d( sample_in_node[node.id], np.where(indicator[node.id] == 0)[0]) depth += 1 return prediction def predict_on_tree(self, tree: Tree, data_iterator: NdarrayIterator) -> np.ndarray: prediction = np.zeros((len(data_iterator.data),), dtype=np.float32) indicator = {} def _update_local(tree, data): res = self._make_indicator_for_prediction(tree, data) return res def _update_remote(channel, len_data): remote_indicator = channel.recv() res = {k: np.unpackbits(v)[:len_data] for k, v in remote_indicator.items()} return res thread_pool = ThreadPool(len(self.channels["val_com"]) + 1) for i, data in enumerate(data_iterator): indicator = {} threads = [] threads.append(thread_pool.apipe(_update_local, tree, data)) for party_id in self.channels["val_com"]: threads.append(thread_pool.apipe( _update_remote, self.channels["val_com"][party_id], len(data))) for t in threads: indicator.update(t.get()) prediction[i * data_iterator.bs: ( i + 1) * data_iterator.bs] = self._gen_prediction(tree, indicator, data) return prediction # Non-parallelized version # def predict_on_tree(self, tree: Tree, data_iterator: NdarrayIterator) -> np.ndarray: # prediction = np.zeros((len(data_iterator.data),), dtype=np.float32) # for i, data in enumerate(data_iterator): # indicator = {} # indicator.update(self._make_indicator_for_prediction(tree, data)) # for party_id in self.channels["val_com"]: # remote_indicator = self.channels["val_com"][party_id].recv() # indicator.update({k: np.unpackbits(v)[:len(data)] for k, v in remote_indicator.items()}) # prediction[i * data_iterator.bs: (i + 1) * data_iterator.bs] = self._gen_prediction(tree, indicator, data) # return prediction def predict_on_boosting_tree(self, boosting_tree: BoostingTree, data_iterator: NdarrayIterator) -> np.ndarray: prediction = np.zeros((len(data_iterator.data),), dtype=np.float32) def _update_local(trees, data): res = {} for tree in trees: res.update(self._make_indicator_for_prediction(tree, data)) return res def _update_remote(channel, len_data): remote_indicator = channel.recv() res = {k: np.unpackbits(v)[:len_data] for k, v in remote_indicator.items()} return res thread_pool = ThreadPool(len(self.channels["val_com"]) + 1) for i, data in enumerate(data_iterator): indicator = {} threads = [] threads.append(thread_pool.apipe( _update_local, boosting_tree.trees, data)) for party_id in self.channels["val_com"]: threads.append(thread_pool.apipe( _update_remote, self.channels["val_com"][party_id], len(data))) for t in threads: indicator.update(t.get()) for j, tree in enumerate(boosting_tree.trees): prediction[i * data_iterator.bs: (i + 1) * data_iterator.bs] += \ self._gen_prediction( tree, indicator, data) * boosting_tree.lr[j] return prediction # Non-parallelized version # def predict_on_boosting_tree(self, boosting_tree: BoostingTree, data_iterator: NdarrayIterator) -> np.ndarray: # prediction = np.zeros((len(data_iterator.data),), dtype=np.float32) # for i, data in enumerate(data_iterator): # indicator = {} # for tree in boosting_tree.trees: # indicator.update(self._make_indicator_for_prediction(tree, data)) # for party_id in self.channels["val_com"]: # remote_indicator = self.channels["val_com"][party_id].recv() # indicator.update({k: np.unpackbits(v)[:len(data)] for k, v in remote_indicator.items()}) # for j, tree in enumerate(boosting_tree.trees): # prediction[i * data_iterator.bs: (i + 1) * data_iterator.bs] += \ # self._gen_prediction(tree, indicator, data) * boosting_tree.lr[j] # return prediction def update_feature_importance(self, tree_feature_importance): for (owner_name, fid) in tree_feature_importance: if owner_name == FedConfig.node_name: f_name = self.train_names[fid] else: f_name = self.remote_f_names[owner_name][fid] if (owner_name, f_name) not in self.feature_importances_: self.feature_importances_[ (owner_name, f_name)] = tree_feature_importance[(owner_name, fid)] else: self.feature_importances_[ (owner_name, f_name)] += tree_feature_importance[(owner_name, fid)] # if (owner_id, fid) not in self.feature_importances_: # self.feature_importances_[ # (owner_id, fid)] = tree_feature_importance[(owner_id, fid)] # else: # self.feature_importances_[ # (owner_id, fid)] += tree_feature_importance[(owner_id, fid)] logger.debug("cur feature importance {}".format( self.feature_importances_)) def load_model(self): pretrain_path = self.input.get("pretrained_model", {}).get("path", '') model_name = self.input.get("pretrained_model", {}).get("name", '') model_path = Path( pretrain_path, model_name ) suffix = model_name.split(".")[-1] if suffix != "pmodel": model_dict = ModelIO.load_json_model(model_path) boosting_tree = BoostingTree.from_dict(model_dict) else: with open(model_path, 'rb') as f: byte_str = f.read() boosting_tree = BoostingTree.from_proto(byte_str) return boosting_tree def check_dataset(self): shapes = self.channels["check_dataset_com"].collect() if self.train_dataset is not None: m = len(self.train_ids) n = len(self.train_features.columns) for d in shapes: if d["train"][0] != m: raise ValueError( "Lengths of the train set mismatched: %d, %d." % (d["train"][0], m)) n += d["train"][1] if n <= 0: raise ValueError( "Number of the feature is zero. Stop training.") if self.val_dataset is not None: m = len(self.val_ids) n = len(self.val_features.columns) for d in shapes: if d["valid"][0] != m: raise ValueError( "Lengths of the valid set mismatched: %d, %d." % (d["valid"][0], m)) n += d["valid"][1] if n <= 0: raise ValueError( "Number of the feature is zero. Stop training.") if self.test_dataset is not None: m = len(self.test_ids) n = len(self.test_features.columns) for d in shapes: if d["test"][0] != m: raise ValueError( "Lengths of the test set mismatched: %d, %d." % (d["test"][0], m)) n += d["test"][1] if n <= 0: raise ValueError( "Number of the feature is zero. Stop predicting.") else: if len(shapes) > 0 and "test" in shapes[0]: m = shapes[0]["test"][0] n = 0 for d in shapes: if d["test"][0] != m: raise ValueError("Lengths of the test set mismatched.") n += d["test"][1] if n <= 0: raise ValueError( "Number of the feature is zero. Stop predicting.") else: self.test_dataset = NdarrayIterator( np.zeros((m, 0)), self.bs) self.test_ids = np.arange(m) def predict(self): out_dict_list = self.channels["sync"].collect() self.check_dataset() boosting_tree = self.load_model() test_y_pred_primitive = self.predict_on_boosting_tree(boosting_tree=boosting_tree, data_iterator=self.test_dataset) loss_inst = get_xgb_loss_inst(boosting_tree.loss_method) test_y_pred = loss_inst.predict(test_y_pred_primitive) output = { "testset": test_y_pred } for out_keys_dict in out_dict_list: for key in out_keys_dict: if key in output: out_keys_dict[key] = output["testset"] self.channels["sync"].scatter(out_dict_list) save_path = self.output.get("path", '') if save_path: if not os.path.exists(save_path): os.makedirs(save_path) file_path = Path(save_path, self.output.get("testset", {}).get("name", '')) if file_path: logger.info("predicted results saved at {}".format(file_path)) pd.DataFrame({"id": self.test_ids, "pred": test_y_pred}).to_csv( file_path, float_format="%.6g", index=False, header=True )
30,579
44.57377
139
py
XFL
XFL-master/python/algorithm/framework/vertical/xgboost/__init__.py
0
0
0
py
XFL
XFL-master/python/algorithm/framework/vertical/xgboost/decision_tree_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import gc import math from typing import Dict, Optional, Union import numpy as np import pandas as pd import ray from algorithm.core.encryption_param import PaillierParam, PlainParam from algorithm.core.tree.big_feature import Feature from algorithm.core.tree.tree_param import XGBTreeParam from algorithm.core.tree.tree_structure import Node, SplitInfo from common.communication.gRPC.python.channel import BroadcastChannel, DualChannel from common.crypto.paillier.paillier import Paillier, PaillierContext from common.crypto.paillier.utils import get_core_num from common.utils.constants import PAILLIER, PLAIN from common.utils.logger import logger try: from ray.internal.internal_api import free except Exception: from ray._private.internal_api import free from service.fed_config import FedConfig from .debug_params import EMBEDING class VerticalDecisionTreeTrainer(object): def __init__(self, tree_param: XGBTreeParam, features: pd.DataFrame, cat_columns: list, split_points: np.ndarray, channels: Dict[str, Union[BroadcastChannel, DualChannel]], encryption_context: Optional[PaillierContext] = None, feature_id_mapping: Optional[Dict[int, int]] = None, tree_index: Optional[int] = None): logger.info(f"Trainer decision tree {tree_index} initialize start.") if tree_param.encryption_param.method not in [PAILLIER, PLAIN]: raise ValueError(f"Encryption method {tree_param.encryption_param.method} not supported.") self.tree_param = tree_param self.features = features self.cat_columns = cat_columns self.split_points = split_points self.party_id = FedConfig.node_id self.max_num_cores = get_core_num(tree_param.max_num_cores) self.tree_index = tree_index self.individual_grad_hess: BroadcastChannel = channels["individual_grad_hess"] self.tree_node_chann: BroadcastChannel = channels["tree_node"] self.summed_grad_hess_chann: DualChannel = channels["summed_grad_hess"] self.min_split_info_chann: DualChannel = channels["min_split_info"] self.sample_index_after_split_chann: DualChannel = channels["sample_index_after_split"] self.encryption_param = self.tree_param.encryption_param self.pub_context = encryption_context self.feature_id_mapping = feature_id_mapping if isinstance(self.encryption_param, PlainParam): self.grad, self.hess = self.individual_grad_hess.recv(use_pickle=True) elif isinstance(self.encryption_param, PaillierParam): if EMBEDING: self.grad_hess = self.individual_grad_hess.recv(use_pickle=True) self.grad_hess = Paillier.ciphertext_from(self.pub_context, self.grad_hess, compression=False) else: self.grad, self.hess = self.individual_grad_hess.recv(use_pickle=True) self.grad = Paillier.ciphertext_from(self.pub_context, self.grad, compression=False) # ciphertext self.hess = Paillier.ciphertext_from(self.pub_context, self.hess, compression=False) else: raise ValueError("Encryption param not supported.") logger.info(f"Trainer decision tree {tree_index} initialize finished.") def fit(self) -> Dict[str, Node]: logger.info(f"Decision tree {self.tree_index} training start..") nodes = {} count = 0 while True: node: Node = self.tree_node_chann.recv(use_pickle=True) if node is None: break logger.info(f"Node {node.id} training start..") if count == 0: if node.sample_index is None: node.sample_index = range(self.features.shape[0]) self.sample_index = node.sample_index if isinstance(self.encryption_param, PlainParam): self.big_feature = Feature.create(values=self.features.iloc[node.sample_index, :], sample_index=node.sample_index, grad=self.grad, hess=self.hess) else: if EMBEDING: self.big_feature = Feature.create(values=self.features.iloc[node.sample_index, :], sample_index=node.sample_index, grad_hess=self.grad_hess) else: self.big_feature = Feature.create(values=self.features.iloc[node.sample_index, :], sample_index=node.sample_index, grad=self.grad, hess=self.hess) big_feature = self.big_feature else: big_feature = self.big_feature.slice_by_sample_index(node.sample_index) gc.collect() count += 1 logger.info(f"Node {node.id} finish preparing data.") def cal_grad_hess_hist_apart(col_name: str): res = big_feature.data.groupby([col_name])[['xfl_grad', 'xfl_hess']].agg({'count', 'sum'}) return res logger.info(f"Node {node.id} calculating grad hess hist start..") send_times = math.ceil(len(big_feature.feature_columns) / self.tree_param.col_batch) res_hist_list = [] if isinstance(self.encryption_param, PaillierParam): if EMBEDING: for i in range(send_times): # split by features cat_index = list( set(self.cat_columns).intersection(set(range(i * self.tree_param.col_batch, (i + 1) * self.tree_param.col_batch))) ) num = int(math.ceil(big_feature.data.shape[0] / self.tree_param.row_batch)) for j in range(num): data_id = ray.put(big_feature.data.iloc[self.tree_param.row_batch * j: self.tree_param.row_batch * (j + 1), :]) @ray.remote def cal_grad_hess_hist_embeding(col_name: str): res = ray.get(data_id).groupby([col_name])['xfl_grad_hess'].agg({'count', 'sum'}) return res ray_tasks = [] for col_name in big_feature.feature_columns[i * self.tree_param.col_batch: (i + 1) * self.tree_param.col_batch]: ray_tasks.append(cal_grad_hess_hist_embeding.remote(col_name)) b = ray.get(ray_tasks) free(data_id) free(ray_tasks) if j == 0: res = b else: b_id = ray.put(b) res_id = ray.put(res) @ray.remote def merge_embeding(k): res = ray.get(res_id) b = ray.get(b_id) r = pd.merge(res[k], b[k], how='outer', left_index=True, right_index=True).fillna(0) r = pd.Series(b[k].columns).apply(lambda x: r[x+'_x'] + r[x+'_y']).T r.columns = list(b[k].columns) return r ray_tasks = [merge_embeding.remote(k) for k in range(len(b))] res = ray.get(ray_tasks) free(b_id) free(res_id) free(ray_tasks) gc.collect() res_hist_partial_list = res hist_list = [(res_hist['sum'].to_numpy(), res_hist['count'].to_numpy()) for res_hist in res_hist_partial_list] if (i + 1) == send_times: # 【stop_flag, hist_list, index of category feature(in binary form)] self.summed_grad_hess_chann.send([False, hist_list, cat_index], use_pickle=True) else: self.summed_grad_hess_chann.send([True, hist_list, cat_index], use_pickle=True) res_hist_list += res_hist_partial_list gc.collect() else: for i in range(send_times): cat_index = list( set(self.cat_columns).intersection(set(range(i * self.tree_param.col_batch, (i + 1) * self.tree_param.col_batch))) ) num = int(math.ceil(big_feature.data.shape[0] / self.tree_param.row_batch)) for j in range(num): data_id = ray.put(big_feature.data.iloc[self.tree_param.row_batch * j: self.tree_param.row_batch * (j + 1), :]) @ray.remote def cal_grad_hess_hist(col_name: str): res = ray.get(data_id).groupby([col_name])[['xfl_grad', 'xfl_hess']].agg({'count', 'sum'}) return res ray_tasks = [] for col_name in big_feature.feature_columns[i * self.tree_param.col_batch: (i + 1) * self.tree_param.col_batch]: ray_tasks.append(cal_grad_hess_hist.remote(col_name)) b = ray.get(ray_tasks) free(data_id) free(ray_tasks) if j == 0: res = b else: b_id = ray.put(b) res_id = ray.put(res) @ray.remote def merge(k): res = ray.get(res_id) b = ray.get(b_id) r = pd.merge(res[k], b[k], how='outer', left_index=True, right_index=True).fillna(0) r = pd.Series(list(b[k].columns)).apply(lambda x: r[(x[0]+'_x', x[1])] + r[(x[0]+'_y', x[1])]).T r.columns = pd.MultiIndex.from_tuples(b[k].columns) return r ray_tasks = [merge.remote(k) for k in range(len(b))] res = ray.get(ray_tasks) free(b_id) free(res_id) free(ray_tasks) gc.collect() res_hist_partial_list = res hist_list = [(res_hist[('xfl_grad', 'sum')].to_numpy(), res_hist[('xfl_hess', 'sum')].to_numpy(), res_hist[('xfl_grad', 'count')].to_numpy()) for res_hist in res_hist_partial_list] if (i + 1) == send_times: self.summed_grad_hess_chann.send([False, hist_list, cat_index], use_pickle=True) else: self.summed_grad_hess_chann.send([True, hist_list, cat_index], use_pickle=True) res_hist_list += res_hist_partial_list gc.collect() else: cat_index = self.cat_columns res_hist_list = pd.Series(big_feature.feature_columns).apply(cal_grad_hess_hist_apart) hist_list = [(res_hist[('xfl_grad', 'sum')].to_numpy(), res_hist[('xfl_hess', 'sum')].to_numpy(), res_hist[('xfl_grad', 'count')].to_numpy()) for res_hist in res_hist_list] self.summed_grad_hess_chann.send([False, hist_list, cat_index], use_pickle=True) gc.collect() logger.info(f"Node {node.id} calculating grad hess hist finished") feature_idx, max_gain_index, left_cat = self.min_split_info_chann.recv(use_pickle=True) if feature_idx == -1: continue if feature_idx in self.cat_columns: if isinstance(self.encryption_param, PaillierParam) and EMBEDING: left_cat_index = res_hist_list[feature_idx]['sum'].index[left_cat] else: left_cat_index = res_hist_list[feature_idx][('xfl_grad', 'sum')].index[left_cat] left_cat_values_ori = [self.split_points[feature_idx][index] for index in left_cat_index] left_cat_values = [] for cat in left_cat_values_ori: if isinstance(cat, list): left_cat_values += cat else: left_cat_values.append(cat) split_info = SplitInfo(owner_id=self.party_id, feature_idx=self.feature_id_mapping[feature_idx].item(), is_category=True, split_point=None, left_cat=left_cat_values) if isinstance(self.encryption_param, PaillierParam) and EMBEDING: left_sample_index = big_feature.data[big_feature.data.iloc[:, feature_idx + 2].isin(left_cat)]['xfl_id'].tolist() right_sample_index = big_feature.data[big_feature.data.iloc[:, feature_idx + 2].isin(left_cat)]['xfl_id'].tolist() else: left_sample_index = big_feature.data[big_feature.data.iloc[:, feature_idx + 3].isin(left_cat)]['xfl_id'].tolist() right_sample_index = big_feature.data[big_feature.data.iloc[:, feature_idx + 3].isin(left_cat)]['xfl_id'].tolist() else: if isinstance(self.encryption_param, PaillierParam) and EMBEDING: # works when goss is true split_point_index = int(res_hist_list[feature_idx]['sum'].index[max_gain_index]) else: split_point_index = int(res_hist_list[feature_idx][('xfl_grad', 'sum')].index[max_gain_index]) # may not be necessary, just for safe split_point_index = min(split_point_index, len(self.split_points[feature_idx]) - 1) split_point = self.split_points[feature_idx][split_point_index] split_info = SplitInfo(owner_id=self.party_id, feature_idx=self.feature_id_mapping[feature_idx].item(), is_category=False, split_point=split_point, left_cat=None) if isinstance(self.encryption_param, PaillierParam) and EMBEDING: left_sample_index = big_feature.data[big_feature.data.iloc[:, feature_idx + 2] <= split_point_index]['xfl_id'].tolist() right_sample_index = big_feature.data[big_feature.data.iloc[:, feature_idx + 2] > split_point_index]['xfl_id'].tolist() else: left_sample_index = big_feature.data[big_feature.data.iloc[:, feature_idx + 3] <= split_point_index]['xfl_id'].tolist() right_sample_index = big_feature.data[big_feature.data.iloc[:, feature_idx + 3] > split_point_index]['xfl_id'].tolist() nodes[node.id] = Node(id=node.id, split_info=split_info, is_leaf=False) self.sample_index_after_split_chann.send([left_sample_index, right_sample_index], use_pickle=True) logger.info(f"Node {node.id} training finished.") return nodes
17,678
53.733746
142
py
XFL
XFL-master/python/algorithm/framework/vertical/xgboost/debug_params.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. EMBEDING = True
622
35.647059
74
py
XFL
XFL-master/python/algorithm/framework/vertical/xgboost/trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import json import os from pathlib import Path from typing import Dict import pandas as pd import numpy as np import ray from algorithm.core.data_io import NdarrayIterator from algorithm.core.encryption_param import PaillierParam, PlainParam from algorithm.core.tree.tree_structure import Node, NodeDict from common.communication.gRPC.python.channel import BroadcastChannel, DualChannel from common.crypto.paillier.paillier import Paillier from common.crypto.paillier.utils import get_core_num from common.utils.logger import logger from common.utils.utils import save_model_config, update_dict from service.fed_config import FedConfig from .base import VerticalXgboostBase from .decision_tree_trainer import VerticalDecisionTreeTrainer from service.fed_job import FedJob from service.fed_node import FedNode from common.utils.model_io import ModelIO class VerticalXgboostTrainer(VerticalXgboostBase): def __init__(self, train_conf: dict, *args, **kwargs): self.channels = dict() self.channels["sync"] = BroadcastChannel(name="sync") conf = self._sync_config() update_dict(train_conf, conf) super().__init__(train_conf, is_label_trainer=False, *args, **kwargs) self.channels["encryption_context"] = BroadcastChannel( name="encryption_context") self.channels["individual_grad_hess"] = BroadcastChannel( name="individual_grad_hess") self.channels["tree_node"] = BroadcastChannel(name="tree_node") self.channels["check_dataset_com"] = BroadcastChannel( name="check_dataset_com") self.channels["summed_grad_hess"] = DualChannel(name="summed_grad_hess_" + FedConfig.node_id, ids=FedConfig.get_label_trainer() + [FedConfig.node_id]) self.channels["min_split_info"] = DualChannel(name="min_split_info_" + FedConfig.node_id, ids=FedConfig.get_label_trainer() + [FedConfig.node_id]) self.channels["sample_index_after_split"] = DualChannel(name="sample_index_after_split_" + FedConfig.node_id, ids=FedConfig.get_label_trainer() + [FedConfig.node_id]) self.channels["val_com"] = DualChannel(name="val_com_" + FedConfig.node_id, ids=FedConfig.get_label_trainer() + [FedConfig.node_id]) self.channels["restart_com"] = DualChannel(name="restart_com_" + FedConfig.node_id, ids=FedConfig.get_label_trainer() + [FedConfig.node_id]) self.channels["early_stop_com"] = DualChannel(name="early_stop_com_" + FedConfig.node_id, ids=FedConfig.get_label_trainer() + [FedConfig.node_id]) if isinstance(self.xgb_config.encryption_param, (PlainParam, type(None))): self.public_context = None elif isinstance(self.xgb_config.encryption_param, PaillierParam): self.public_context = self.channels["encryption_context"].recv( use_pickle=False) self.public_context = Paillier.context_from(self.public_context) else: raise TypeError( f"Encryption param type {type(self.xgb_config.encryption_param)} not valid.") if self.train_features is not None: input_schema = ','.join([_ for _ in self.train_features.columns if _ not in set(["y", "id"])]) else: input_schema = "" self.export_conf = [{ "class_name": "VerticalXGBooster", "identity": self.identity, "filename": self.output.get("proto_model", {}).get("name", ''), "input_schema": input_schema, }] ray.init(num_cpus=get_core_num(self.xgb_config.max_num_cores), ignore_reinit_error=True) def _sync_config(self): config = self.channels["sync"].recv() return config def fit(self): self.channels["sync"].send({FedNode.node_name: self.train_names}) self.check_dataset() # nodes_dict = {} node_dict = NodeDict() for tree_idx in range(1, self.xgb_config.num_trees+1): # training section logger.info("Tree {} start training.".format(tree_idx)) restart_status = 0 while True: sampled_features, feature_id_mapping = self.col_sample() # col index in feature cat_columns_after_sampling = list( filter(lambda x: feature_id_mapping[x] in self.cat_columns, list(feature_id_mapping.keys()))) split_points_after_sampling = [self.split_points[feature_id_mapping[k]] for k in feature_id_mapping.keys()] trainer = VerticalDecisionTreeTrainer(tree_param=self.xgb_config, features=sampled_features, cat_columns=cat_columns_after_sampling, split_points=split_points_after_sampling, channels=self.channels, encryption_context=self.public_context, feature_id_mapping=feature_id_mapping, tree_index=tree_idx) nodes = trainer.fit() restart_status = self.channels["restart_com"].recv() if restart_status != 1: break logger.info(f"trainer tree {tree_idx} training restart.") logger.info("Tree {} training done.".format(tree_idx)) if restart_status == 2: logger.info( "trainer early stopped. because a tree's root is leaf.") break node_dict.update(nodes) if self.xgb_config.run_goss: self.predict_on_tree(nodes, self.train_dataset) # valid section logger.info( "trainer: Validation on tree {} start.".format(tree_idx)) self.predict_on_tree(nodes, self.val_dataset) if self.channels["early_stop_com"].recv(): logger.info("trainer early stopped.") break logger.info("Validation on tree {} done.".format(tree_idx)) # if self.interaction_params.get("save_frequency") > 0 and (tree_idx + 1) % self.interaction_params.get( # "save_frequency") == 0: # self.save(node_dict, epoch=tree_idx + 1) if self.interaction_params.get("save_frequency") > 0 and tree_idx % self.interaction_params.get( "save_frequency") == 0: self.save(node_dict, epoch=tree_idx) # model preserve self.save(node_dict, final=True) ray.shutdown() def _make_indicator_for_prediction(self, nodes: Dict[str, Node], feature: np.ndarray): indicator = {} for node_id, node in nodes.items(): feature_idx = node.split_info.feature_idx data = feature[:, feature_idx] if node.split_info.is_category: indicator[node_id] = np.isin(data, node.split_info.left_cat) else: indicator[node_id] = (data <= node.split_info.split_point) return indicator def predict_on_tree(self, nodes: Dict[str, Node], data_iterator: NdarrayIterator): for data in data_iterator: indicator = self._make_indicator_for_prediction(nodes, data) indicator = {k: np.packbits(v) for k, v in indicator.items()} self.channels["val_com"].send(indicator) def predict_on_boosting_tree(self, nodes: Dict[str, Node], data_iterator: NdarrayIterator): self.predict_on_tree(nodes, data_iterator) def save(self, node_dict: NodeDict, epoch: int = None, final: bool = False): if final: save_model_config(stage_model_config=self.export_conf, save_path=self.output.get("path")) save_dir = self.output.get("path") if not os.path.exists(save_dir): os.makedirs(save_dir) model_name = self.output.get("model", {}).get("name") proto_name = self.output.get("proto_model", {}).get("name") if model_name: out_dict = node_dict.to_dict() model_dict = {"nodes": out_dict} ModelIO.save_json_model(model_dict, save_dir, model_name, epoch=epoch, version='1.4.0') if proto_name: model_name_list = self.output.get("proto_model")["name"].split(".") name_prefix, name_postfix = ".".join( model_name_list[:-1]), model_name_list[-1] if not final and epoch: model_name = name_prefix + "_epoch_{}".format(epoch) + "." + name_postfix else: model_name = name_prefix + "." + name_postfix model_path = os.path.join(save_dir, model_name) xgb_output = node_dict.to_proto() with open(model_path, 'wb') as f: f.write(xgb_output) logger.info("model saved as: {}.".format(model_path)) def load_model(self): pretrain_path = self.input.get("pretrained_model", {}).get("path", '') model_name = self.input.get("pretrained_model", {}).get("name", '') model_path = Path( pretrain_path, model_name ) suffix = model_name.split(".")[-1] if suffix != "pmodel": model_dict = ModelIO.load_json_model(model_path) node_dict = NodeDict.from_dict(model_dict["nodes"]) else: with open(model_path, 'rb') as f: byte_str = f.read() node_dict = NodeDict.from_proto(byte_str) return node_dict def check_dataset(self): d = dict() if self.train_dataset is not None: d["train"] = len(self.train_ids), len(self.train_features.columns) if self.val_dataset is not None: d["valid"] = len(self.val_ids), len(self.val_features.columns) if self.test_dataset is not None: d["test"] = len(self.test_ids), len(self.test_features.columns) self.channels["check_dataset_com"].send(d) def predict(self): out_dict = {key: None for key, value in self.train_conf.get("output", {}).items() if key != "path" and value.get("name")} self.channels["sync"].send(out_dict) self.check_dataset() node_dict = self.load_model() self.predict_on_boosting_tree(nodes=node_dict.nodes, data_iterator=self.test_dataset) out_dict = self.channels["sync"].recv() save_path = self.output.get("path", '') if save_path: if not os.path.exists(save_path): os.makedirs(save_path) for key in out_dict: file_path = Path(save_path, self.output[key]["name"]) if key == "testset" and file_path: logger.info("predicted results saved at {}".format(file_path)) pd.DataFrame({"id": self.test_ids, "pred": out_dict[key]}).to_csv( file_path, float_format="%.6g", index=False, header=True )
12,254
43.563636
129
py
XFL
XFL-master/python/algorithm/framework/vertical/binning_woe_iv/base.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import json import os import time from pathlib import Path import numpy as np import pandas as pd from common.utils.config_parser import TrainConfigParser from common.utils.logger import logger from sklearn.preprocessing import LabelEncoder def fintech_cut(ser, bins, nan_l, method): # all labels begins with 1 def rebin_freq_2(threshold, num): if num <= threshold[0]: return 1 elif threshold[0] < num <= threshold[1]: return 2 else: return 3 def ret_bin(ser_, cut_v): l_tmp = [float("%.6f" % i) for i in list(cut_v[1])] if len(cut_v[1]) > 2: cut_bins = [-np.inf] + l_tmp[1:-1] + [np.inf] if method == "equal_frequency": cut_ser = pd.DataFrame(LabelEncoder().fit_transform( cut_v[0])).set_index(cut_v[0].index)[0] cut_ser = cut_ser + 1 elif method == "equal_width": cut_ser = cut_v[0] cut_ser = cut_ser.cat.remove_unused_categories() else: # never equal_width cut_bins = [-np.inf] + [float("%.6f" % i) for i in l_tmp] + [np.inf] cut_ser = ser_.apply(lambda x: rebin_freq_2(cut_v[1], x)) return cut_ser, cut_bins def range_map(nums, _tmp, i): if i > 0: left = _tmp[nums[i - 1] - 1][1] right = _tmp[nums[i] - 1][1] else: left = _tmp[num[i] - 1][0] right = _tmp[num[i] - 1][1] return f"({left}, {right}]", [left, right] # split nan and norm col_dict = {} ser_name = ser.name bins_ser = [] ser_cut = ser.loc[ser[~ser.isin(nan_l)].index] nan_ind = ser[ser.isin(nan_l)].index ser_nan = ser.loc[nan_ind] if len(ser_nan) > 0: nan_value = float("%.6f" % ser_nan.loc[np.min(nan_ind)]) bins_ser.append(nan_value) ser_nan = pd.DataFrame(np.zeros(len(ser_nan)) + nan_value).set_index(ser_nan.index)[0] col_dict[nan_value] = nan_value else: pass if len(ser_cut) > 0: # first cut cut_value = None if len(set(ser_cut)) > 1: if method == "equal_width": cut_value = pd.cut(ser_cut, bins, retbins=True, labels=[ i + 1 for i in range(bins)]) elif method == "equal_frequency": cut_value = pd.qcut( ser_cut, bins, retbins=True, duplicates='drop') elif len(set(ser_cut)) == 1: cut_value = (pd.Series(len(ser_cut)*[1]), list(set(ser_cut))) cut_ser_, cut_bin = ret_bin(ser_cut, cut_value) # label-range map tmp_ = [[cut_bin[i], cut_bin[i + 1]] for i in range(len(cut_bin) - 1)] num = sorted(set(cut_ser_)) tt = [] col_dict.update(dict(zip([num[i] for i in range(len(num))], [ range_map(num, tmp_, i)[0] for i in range(len(num))]))) for i in range(len(num)): tt += range_map(num, tmp_, i)[1] cut_bin_final = sorted(set(tt)) # concat nan and cut ser = pd.concat([ser_nan, cut_ser_]).loc[ser.index] bins_ser = bins_ser + cut_bin_final # rename the rightest value to -inf/inf if bins_ser[-1] != np.inf: bins_ser = bins_ser[:-1] + [np.inf] col_dict[list(col_dict.keys( ))[-1]] = f"{col_dict[list(col_dict.keys())[-1]].split(' ')[0]} {np.inf}]" if bins_ser[0] != -np.inf and bins_ser[0] not in nan_l: bins_ser = [-np.inf] + bins_ser[1:] col_dict[list(col_dict.keys())[ 0]] = f"({-np.inf}, {col_dict[list(col_dict.keys())[0]].split(' ')[1]}]" else: ser = ser_nan ser.name = ser_name return ser, bins_ser, {ser.name: col_dict} class VerticalBinningWoeIvBase(TrainConfigParser): def __init__(self, train_conf: dict, label: bool = False, *args, **kwargs): """[summary] Args: train_conf (dict): [description] """ super().__init__(train_conf) self.train_conf = train_conf self.df = None self.val = None self.label = label self.woe_map = {} self.binning_split = {} if self.interaction_params: self.save_model = self.interaction_params.get("save_model", False) else: self.save_model = False self.save_dir = Path(self.output.get("path")) self.transform_switch = False self._init_data() if self.save_model: self.save_model_name = self.output.get("model").get("name") self.export_conf = [{ "class_name": "VerticalBinningWoeIv", "filename": self.save_model_name, "bins": self.train_params["binning"]["bins"], "input_schema": ','.join([_ for _ in self.df.columns if _ not in set(["y", "id"])]), }] self.feature_binning() def _init_data(self) -> None: """Load data: input data with id Returns: """ logger.info("Start reading data.") if self.input_trainset[0].get("has_id", True): # index_col = self.input_trainset[0].get("index_col", 'id') index_col = 0 else: index_col = None if self.input_trainset[0]["type"] == "csv": file_path = str( Path(self.input_trainset[0]["path"], self.input_trainset[0]["name"])) if self.computing_engine == "local": self.df = pd.read_csv(file_path, index_col=index_col) logger.info("Reading data successfully.") if self.input_trainset[0].get("has_label", False): # self.y = self.df[label_name] # self.df = self.df.drop(label_name, axis=1) self.df.columns = ["y"] + list(self.df.columns[1:]) self.y = self.df.iloc[:, 0] self.df = self.df.iloc[:, 1:] else: self.y = None # read val for transform if len(self.input_valset) > 0: self.transform_switch = True self.val = pd.read_csv( str(Path(self.input_valset[0]["path"], self.input_valset[0]["name"])), index_col=index_col) def feature_binning(self) -> None: """Parse and execute feature binning. Returns: None """ logger.info("Start binning") nan_l = self.input_trainset[0]["nan_list"] method = self.train_params["binning"]['method'] bin_num = self.train_params["binning"]["bins"] time_start = time.time() tmp = pd.Series(self.df.columns).apply( lambda x: fintech_cut(self.df[x], bin_num, nan_l, method)) def get_cut_result(result): self.df[result[0].name] = result[0] self.binning_split.update({result[0].name: result[1]}) self.woe_map.update(result[2]) tmp.apply(lambda x: get_cut_result(x)) time_end = time.time() logger.info("Cost time of binning is: {}s".format( time_end - time_start)) if not os.path.exists(self.save_dir): os.makedirs(self.save_dir) file_path = f'{self.save_dir}/{self.output["split_points"]["name"]}' with open(file_path, "w") as wf: json.dump(self.binning_split, wf) logger.info("Binning split points saved as {}.".format(file_path)) if isinstance(self.y, pd.Series): self.df = pd.concat([self.y, self.df], axis=1)
8,207
35.642857
107
py
XFL
XFL-master/python/algorithm/framework/vertical/binning_woe_iv/label_trainer.py
# Copyright 2022 The XFL Authors. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import json import math import time import numpy as np import pandas as pd from service.fed_control import ProgressCalculator from common.checker.matcher import get_matched_config from common.checker.x_types import All from common.communication.gRPC.python.channel import BroadcastChannel from common.crypto.paillier.paillier import Paillier, PaillierCiphertext from common.utils.logger import logger from .base import VerticalBinningWoeIvBase class VerticalBinningWoeIvLabelTrainer(VerticalBinningWoeIvBase): def __init__(self, train_conf: dict, *args, **kwargs): """ Args: train_conf: training parameters *args: **kwargs: """ self.sync_channel = BroadcastChannel(name="sync") self._sync_config(train_conf) super().__init__(train_conf, label=True, *args, **kwargs) self.neg_total_count, self.pos_total_count = 0, 0 logger.info("node-1:successfully binning.") self.woe_dict_total = {} self.iv_dict_total = {} self.neg_bin_count = {} self.pos_bin_count = {} self.neg_bin_ratio = {} self.pos_bin_ratio = {} self.broadcast_channel = BroadcastChannel(name="vertical_binning_woe_iv_channel") def _sync_config(self, config): sync_rule = { "train_info": All() } config_to_sync = get_matched_config(config, sync_rule) self.sync_channel.broadcast(config_to_sync) def fit(self): # broadcast_channel = BroadcastChannel(name="vertical_binning_woe_iv_channel") encryption_config_pre = self.train_params["encryption"] encryption_method = list(encryption_config_pre.keys())[0].lower() encryption_config = encryption_config_pre[encryption_method] # if encryption_method == "paillier": # pri_context = Paillier.context(encryption_config["key_bit_size"], djn_on=encryption_config["djn_on"]) # self.broadcast_channel.broadcast(pri_context.to_public().serialize(), use_pickle=False) # elif encryption_method == "plain": # pass # else: # raise ValueError( # f"Encryption method {encryption_method} not supported! Valid methods are 'paillier', 'plain'.") # logger.info("Start calculate host IV with WOE values.") self.label_trainer_woe_iv() if encryption_method == "paillier": pri_context = Paillier.context(encryption_config["key_bit_size"], djn_on=encryption_config["djn_on"]) self.broadcast_channel.broadcast(pri_context.to_public().serialize(), use_pickle=False) num_cores = -1 if encryption_config["parallelize_on"] else 1 label = self.df[["y"]].to_numpy().flatten().astype(np.int32) logger.info(f"Encrypting label using {encryption_method} method.") en_label = Paillier.encrypt(pri_context, label, precision=encryption_config["precision"], obfuscation=True, num_cores=num_cores) logger.info("Encryption complete.") self.broadcast_channel.broadcast(Paillier.serialize(en_label), use_pickle=False) elif encryption_method == "plain": id_label_pair = self.df[["y"]] self.broadcast_channel.broadcast(id_label_pair, use_pickle=True) feedback_list = self.broadcast_channel.collect() assert len(self.broadcast_channel.remote_ids) == len(feedback_list) for uid, feedback in zip(self.broadcast_channel.remote_ids, feedback_list): client_woe_dict, client_iv_dict = {}, {} if encryption_method == "paillier": logger.info(f"Decrypting woe_feedback using {encryption_method} method.") for _id, feature in feedback["woe_feedback_list"].items(): c = feature.apply(lambda x: PaillierCiphertext.deserialize_from(pri_context, x)) feedback["woe_feedback_list"][_id] = c.apply(lambda x: Paillier.decrypt(pri_context, x, dtype='float', num_cores=num_cores)) logger.info("Decryption Complete.") woe_feedback_list, bins_count = feedback["woe_feedback_list"], feedback["bins_count"] logger.info("Start calculate woe for trainer") time_s = time.time() for k, v in woe_feedback_list.items(): # featName = "{}_{}".format(uid, k) client_woe_dict[k], client_iv_dict[k] = {}, 0 neg_ = bins_count[k] - v pos_prob, neg_prob = (v / self.pos_total_count), (neg_ / self.neg_total_count) pos_prob = pos_prob.apply(lambda x: 1e-7 if x == 0 else x) neg_prob = neg_prob.apply(lambda x: 1e-7 if x == 0 else x) woe_pre = pos_prob / neg_prob woe = woe_pre.apply(lambda x: float("%.6f" % math.log(x))) # woe.index = pd.Series(woe.index).apply(lambda x: int(x)) # v.index = pd.Series(v.index).apply(lambda x: int(x)) # neg_.index = pd.Series(neg_.index).apply(lambda x: int(x)) self.pos_bin_count[k] = v.to_dict() self.neg_bin_count[k] = neg_.to_dict() pos_prob = pos_prob.apply(lambda x: float("%.6f" % x)) neg_prob = neg_prob.apply(lambda x: float("%.6f" % x)) self.pos_bin_ratio[k] = pos_prob.to_dict() self.neg_bin_ratio[k] = neg_prob.to_dict() client_woe_dict[k] = woe.to_dict() client_iv_dict[k] += float("%.6f" % np.sum((pos_prob - neg_prob) * woe)) logger.info("Trainer woe cost:" + str(time.time() - time_s)) logger.info("Calculate host IV with WOE values completed.") # logger.info("Host WOE dictionary: {}".format(client_woe_dict)) # logger.info("Host IV dictionary: {}".format(client_iv_dict)) # Save host dicts self.woe_dict_total.update(client_woe_dict) self.iv_dict_total.update(client_iv_dict) guest_file_path = f'{self.save_dir}/{self.output["iv"]["name"]}' with open(guest_file_path, "w") as wf: json.dump({"woe": self.woe_dict_total, "iv": self.iv_dict_total, "count_neg": self.neg_bin_count, "count_pos": self.pos_bin_count, "ratio_pos": self.pos_bin_ratio, "ratio_neg": self.neg_bin_ratio}, wf) logger.info("Host {} WOE & IV values saved as {}.".format(uid, guest_file_path)) ProgressCalculator.finish_progress() def label_trainer_woe_iv(self): logger.info("Start calculate Guest IV with WOE values.") woe_dict, iv_dict = {}, {} # count_neg_dict, count_pos_dict, bins_total, percentage, bad_rate = {}, {}, {}, {}, {} # good_percentage, bad_percentage = {}, {} # # count total label = 0, 1 total_count = self.df.groupby("y")["y"].count() self.neg_total_count, self.pos_total_count = total_count[0], total_count[1] feat_woe = set(self.df.columns).difference(set("y")) logger.info("Start calculate woe for label trainer") time_s = time.time() for feature in feat_woe: woe_dict[feature], iv_dict[feature] = {}, 0 feature_df = self.df[[feature, "y"]] tmp_count = feature_df.groupby([feature])['y'].agg({'count', 'sum'}) neg_bin_count = tmp_count['count'] - tmp_count['sum'] pos_prob, neg_prob = (tmp_count['sum'] / self.pos_total_count), (neg_bin_count / self.neg_total_count) pos_prob = pos_prob.apply(lambda x: 1e-7 if x == 0 else x) neg_prob = neg_prob.apply(lambda x: 1e-7 if x == 0 else x) woe_pre = pos_prob / neg_prob woe = woe_pre.apply(lambda x: float("%.6f" % math.log(x))) iv_dict[feature] = float("%.6f" % np.sum((pos_prob - neg_prob) * woe)) woe.index = pd.Series(woe.index).apply(lambda x: self.woe_map[feature][x]) tmp_count['sum'].index = pd.Series(tmp_count['sum'].index).apply(lambda x: self.woe_map[feature][x]) neg_bin_count.index = pd.Series(neg_bin_count.index).apply(lambda x: self.woe_map[feature][x]) self.pos_bin_count[feature] = tmp_count['sum'].to_dict() self.neg_bin_count[feature] = neg_bin_count.to_dict() pos_prob = pos_prob.apply(lambda x: float("%.6f" % x)) neg_prob = neg_prob.apply(lambda x: float("%.6f" % x)) pos_prob.index = pd.Series(pos_prob.index).apply(lambda x: self.woe_map[feature][x]) neg_prob.index = pd.Series(neg_prob.index).apply(lambda x: self.woe_map[feature][x]) self.pos_bin_ratio[feature] = pos_prob.to_dict() self.neg_bin_ratio[feature] = neg_prob.to_dict() woe_dict[feature] = woe.to_dict() logger.info("label trainer cost:" + str(time.time() - time_s)) logger.info("Calculate Guest IV with WOE values completed.") # # logger.info("Guest WOE dictionary: {}".format(woe_dict)) # # logger.info("Guest IV dictionary: {}".format(iv_dict)) # # Save guest dicts # save_dir = self.output["dataset"]["path"] # if not os.path.exists(save_dir): os.makedirs(save_dir) # guest_file_path = f'{save_dir}/{self.output["dataset"]["name"]}.json' # with open(guest_file_path, "a") as wf: # json.dump({"woe": woe_dict, "iv": iv_dict, "count_neg": count_neg_dict, "count_pos": count_pos_dict, # "bins_total": bins_total, "percentage": percentage, "bad_rate": bad_rate, # "good_percentage": good_percentage, "bad_percentage": bad_percentage}, wf) self.woe_dict_total.update(woe_dict) self.iv_dict_total.update(iv_dict) # logger.info("Guest WOE & IV values saved as {}.".format(guest_file_path)) logger.info("Guest WOE & IV values saved")
10,999
49.925926
115
py
XFL
XFL-master/python/algorithm/framework/vertical/binning_woe_iv/__init__.py
0
0
0
py