custom_models.py

import torch
import torch.nn as nn
from transformers import AutoModel, AutoConfig
from transformers.modeling_utils import PreTrainedModel
from transformers import PretrainedConfig


class CustomConfig(PretrainedConfig):
    model_type = "roberta"

    def __init__(
        self,
        num_classes: int = 2,
        **kwargs,
    ):
        self.num_classes = num_classes
        super().__init__(**kwargs)


# ====================================================
# Model
# ====================================================
# class MeanPooling(nn.Module):
class MeanPooling(PreTrainedModel):
    def __init__(
        self,
        config
        # **kwargs,
    ):
        super(MeanPooling, self).__init__(config)

    def forward(self, last_hidden_state, attention_mask):
        input_mask_expanded = (
            attention_mask.unsqueeze(-1).expand(last_hidden_state.size()).float()
        )
        sum_embeddings = torch.sum(last_hidden_state * input_mask_expanded, 1)
        sum_mask = input_mask_expanded.sum(1)
        sum_mask = torch.clamp(sum_mask, min=1e-9)
        mean_embeddings = sum_embeddings / sum_mask
        return mean_embeddings


# class CustomModel(nn.Module):
class CustomModel(PreTrainedModel):
    config_class = CustomConfig

    def __init__(
        self,
        cfg,
        num_labels=2,
        config_path=None,
        pretrained=True,
        binary_classification=True,
        **kwargs,
    ):
        # super().__init__()
        self.cfg = cfg
        self.num_labels = num_labels
        if config_path is None:
            self.config = AutoConfig.from_pretrained(
                self.cfg.model_name, output_hidden_states=True
            )

        else:
            self.config = torch.load(config_path)

        super().__init__(self.config)

        if pretrained:
            self.model = AutoModel.from_pretrained(
                self.cfg.model_name, config=self.config
            )
        else:
            self.model = AutoModel(self.config)

        if self.cfg.gradient_checkpointing:
            self.model.gradient_checkpointing_enable()

        self.pool = MeanPooling(config=self.config)

        self.binary_classification = binary_classification

        if self.binary_classification:
            # for binary classification we only want to output a single value
            self.fc = nn.Linear(self.config.hidden_size, self.num_labels - 1)
        else:
            self.fc = nn.Linear(self.config.hidden_size, self.num_labels)

        self._init_weights(self.fc)

        self.sigmoid_fn = nn.Sigmoid()

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)

    def feature(self, input_ids, attention_mask, token_type_ids):
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
        )
        last_hidden_states = outputs[0]
        feature = self.pool(last_hidden_states, attention_mask)
        return feature

    def forward(self, input_ids, attention_mask, token_type_ids):
        feature = self.feature(input_ids, attention_mask, token_type_ids)
        output = self.fc(feature)
        if self.binary_classification:
            # for binary classification we have to use Sigmoid Function
            # https://towardsdatascience.com/sigmoid-and-softmax-functions-in-5-minutes-f516c80ea1f9
            # https://towardsdatascience.com/bert-to-the-rescue-17671379687f
            output = self.sigmoid_fn(output)

        return output