Spaces:
Running on Zero
Running on Zero
File size: 11,870 Bytes
9d7cf7f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 | from copy import deepcopy
from dataclasses import dataclass
from lightning.pytorch.utilities.types import EVAL_DATALOADERS, TRAIN_DATALOADERS
from numpy import ndarray
from torch import Tensor
from torch.utils import data
from torch.utils.data import DataLoader, Dataset
from typing import Dict, List, Tuple, Callable, Optional
import os
import lightning.pytorch as pl
import numpy as np
import torch
from .datapath import Datapath, LazyAsset
from .spec import ConfigSpec
from .transform import Transform
from ..model.spec import ModelInput
from ..rig_package.info.asset import Asset
from ..tokenizer.spec import Tokenizer, TokenizeInput
@dataclass
class DatasetConfig(ConfigSpec):
shuffle: bool
batch_size: int
num_workers: int
datapath: Datapath
pin_memory: bool=True
persistent_workers: bool=True
@classmethod
def parse(cls, **kwargs) -> 'DatasetConfig':
cls.check_keys(kwargs)
return DatasetConfig(
shuffle=kwargs.get('shuffle', False),
batch_size=kwargs.get('batch_size', 1),
num_workers=kwargs.get('num_workers', 1),
pin_memory=kwargs.get('pin_memory', True),
persistent_workers=kwargs.get('persistent_workers', True),
datapath=Datapath.parse(**kwargs.get('datapath')),
)
def split_by_cls(self) -> Dict[str|None, 'DatasetConfig']:
res: Dict[str|None, DatasetConfig] = {}
datapath_dict = self.datapath.split_by_cls()
for cls, v in datapath_dict.items():
res[cls] = DatasetConfig(
shuffle=self.shuffle,
batch_size=self.batch_size,
num_workers=self.num_workers,
datapath=v,
pin_memory=self.pin_memory,
persistent_workers=self.persistent_workers,
)
return res
class RigDatasetModule(pl.LightningDataModule):
def __init__(
self,
process_fn: Optional[Callable[[List[ModelInput]], List[Dict]]]=None,
train_dataset_config: Optional[DatasetConfig]=None,
validate_dataset_config: Optional[Dict[str|None, DatasetConfig]]=None,
predict_dataset_config: Optional[Dict[str|None, DatasetConfig]]=None,
train_transform: Optional[Transform]=None,
validate_transform: Optional[Transform]=None,
predict_transform: Optional[Transform]=None,
tokenizer: Optional[Tokenizer]=None,
debug: bool=False,
):
super().__init__()
self.process_fn = process_fn
self.train_dataset_config = train_dataset_config
self.validate_dataset_config = validate_dataset_config
self.predict_dataset_config = predict_dataset_config
self.train_transform = train_transform
self.validate_transform = validate_transform
self.predict_transform = predict_transform
self.tokenizer = tokenizer
self.debug = debug
if debug:
print("\033[31mWARNING: debug mode, dataloader will be extremely slow !!!\033[0m")
# build train datapath
if self.train_dataset_config is not None:
self.train_datapath = self.train_dataset_config.datapath
else:
self.train_datapath = None
# build validate datapath
if self.validate_dataset_config is not None:
self.validate_datapath = {
cls: self.validate_dataset_config[cls].datapath
for cls in self.validate_dataset_config
}
else:
self.validate_datapath = None
# build predict datapath
if self.predict_dataset_config is not None:
self.predict_datapath = {
cls: self.predict_dataset_config[cls].datapath
for cls in self.predict_dataset_config
}
else:
self.predict_datapath = None
self.tokenizer = tokenizer
def prepare_data(self):
pass
def train_dataloader(self) -> TRAIN_DATALOADERS:
if self.train_dataset_config is None:
raise ValueError("do not have train_dataset_config")
if self.train_transform is None:
raise ValueError("do not have train_transform")
if self.train_datapath is not None:
self._train_ds = RigDataset(
process_fn=self.process_fn,
data=self.train_datapath.get_data(),
name="train",
tokenizer=self.tokenizer,
transform=self.train_transform,
debug=self.debug,
)
else:
return None
return self._create_dataloader(
dataset=self._train_ds,
config=self.train_dataset_config,
is_train=True,
drop_last=False,
)
def val_dataloader(self) -> EVAL_DATALOADERS:
if self.validate_dataset_config is None:
raise ValueError("do not have validate_dataset_config")
if self.validate_transform is None:
raise ValueError("do not have validate_transform")
if self.validate_datapath is not None:
self._validation_ds = {}
for cls in self.validate_datapath:
self._validation_ds[cls] = RigDataset(
process_fn=self.process_fn,
data=self.validate_datapath[cls].get_data(),
name=f"validate-{cls}",
tokenizer=self.tokenizer,
transform=self.validate_transform,
debug=self.debug,
)
else:
return None
return self._create_dataloader(
dataset=self._validation_ds,
config=self.validate_dataset_config,
is_train=False,
drop_last=False,
)
def predict_dataloader(self):
if self.predict_dataset_config is None:
raise ValueError("do not have predict_dataset_config")
if self.predict_transform is None:
raise ValueError("do not have predict_transform")
if self.predict_datapath is not None:
self._predict_ds = {}
for cls in self.predict_datapath:
self._predict_ds[cls] = RigDataset(
process_fn=self.process_fn,
data=self.predict_datapath[cls].get_data(),
name=f"predict-{cls}",
tokenizer=self.tokenizer,
transform=self.predict_transform,
debug=self.debug,
)
else:
return None
return self._create_dataloader(
dataset=self._predict_ds,
config=self.predict_dataset_config,
is_train=False,
drop_last=False,
)
def _create_dataloader(
self,
dataset: Dataset|Dict[str, Dataset],
config: DatasetConfig|Dict[str|None, DatasetConfig],
is_train: bool,
**kwargs,
) -> DataLoader|Dict[str, DataLoader]:
def create_single_dataloader(dataset, config: DatasetConfig, **kwargs):
return DataLoader(
dataset,
batch_size=config.batch_size,
shuffle=config.shuffle,
num_workers=config.num_workers,
pin_memory=config.pin_memory,
persistent_workers=config.persistent_workers,
collate_fn=dataset.collate_fn,
**kwargs,
)
if isinstance(dataset, Dict):
assert isinstance(config, dict)
return {k: create_single_dataloader(v, config[k], **kwargs) for k, v in dataset.items()}
else:
assert isinstance(config, DatasetConfig)
return create_single_dataloader(dataset, config, **kwargs)
class RigDataset(Dataset):
def __init__(
self,
data: List[LazyAsset],
transform: Transform,
name: Optional[str]=None,
process_fn: Optional[Callable[[List[ModelInput]], List[Dict]]]=None,
tokenizer: Optional[Tokenizer]=None,
debug: bool=False,
) -> None:
super().__init__()
self.data = data
self.name = name
self.process_fn = process_fn
self.tokenizer = tokenizer
self.transform = transform
self.debug = debug
if not debug:
assert self.process_fn is not None, 'missing data processing function'
def __len__(self) -> int:
return len(self.data)
def __getitem__(self, idx) -> ModelInput:
lazy_asset = self.data[idx]
asset = lazy_asset.load()
self.transform.apply(asset=asset)
if self.tokenizer is not None and asset.parents is not None:
x = TokenizeInput(
joints=asset.joints,
parents=asset.parents,
cls=asset.cls,
joint_names=asset.joint_names,
)
tokens = self.tokenizer.tokenize(input=x)
else:
tokens = None
return ModelInput(asset=asset, tokens=tokens)
def _collate_fn_debug(self, batch):
return batch
def _collate_fn(self, batch):
processed_batch = self.process_fn(batch) # type: ignore
processed_batch: List[Dict]
tensors_stack = {}
tensors_cat = {}
non_tensors = {}
vis = {}
def check(x):
assert x not in vis, f"multiple keys found: {x}"
vis[x] = True
for k, v in processed_batch[0].items():
if k == "cat":
assert isinstance(v, dict)
for k1 in v.keys():
check(k1)
tensors_cat[k1] = []
for i in range(len(processed_batch)):
v1 = processed_batch[i]['cat'][k1]
if isinstance(v1, ndarray):
v1 = torch.from_numpy(v1)
elif isinstance(v1, Tensor):
v1 = v1
else:
raise ValueError(f"cannot concatenate non-tensor type of key {k1}, type: {type(v1)}")
tensors_cat[k1].append(v1)
elif k == "non":
assert isinstance(v, dict)
for k1 in v.keys():
check(k1)
non_tensors[k1] = []
for i in range(len(processed_batch)):
v1 = processed_batch[i]['non'][k1]
if isinstance(v1, ndarray):
v1 = torch.from_numpy(v1)
non_tensors[k1].append(v1)
else:
check(k)
tensors_stack[k] = []
for i in range(len(processed_batch)):
v1 = processed_batch[i][k]
if isinstance(v1, ndarray):
v1 = torch.from_numpy(v1)
elif isinstance(v1, Tensor):
v1 = v1
else:
raise ValueError(f"cannot stack type of key {k}, type: {type(v1)}")
tensors_stack[k].append(v1)
collated_stack = {k: torch.stack(v) for k, v in tensors_stack.items()}
collated_cat = {k: torch.concat(v, dim=1) for k, v in tensors_cat.items()}
collated_batch = {
**collated_stack,
**collated_cat,
**non_tensors,
}
return collated_batch
def collate_fn(self, batch):
if self.debug:
return self._collate_fn_debug(batch)
return self._collate_fn(batch) |