add roberta.py and gpt2.py files for custom heads and forwards

gpt2.py

@@ -0,0 +1,200 @@
+import logging
+from typing import Optional, Tuple, Union
+
+import torch
+from torch import nn
+from torch.nn import MSELoss, CrossEntropyLoss, BCEWithLogitsLoss
+from transformers import GPT2Model, GPT2ForSequenceClassification
+from transformers.modeling_outputs import SequenceClassifierOutputWithPast
+
+
+logger = logging.getLogger(__name__)
+
+
+# Simple version #
+
+class GPT2ClassificationHeadCustomSimple(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        hidden_size = config.n_embd
+        self.dense_1 = nn.Linear(hidden_size, 2 * hidden_size)
+        self.dense_2 = nn.Linear(2 * hidden_size, hidden_size)
+        self.dropout = nn.Dropout(config.resid_pdrop)
+        self.out_proj = nn.Linear(hidden_size, config.num_labels, bias=False)
+
+    def forward(self, x):
+        x = self.dense_1(x)
+        x = torch.relu(x)
+        x = self.dropout(x)
+
+        x = self.dense_2(x)
+        x = torch.relu(x)
+        x = self.dropout(x)
+
+        x = self.out_proj(x)
+        return x
+
+
+class GPT2ForSequenceClassificationCustomSimple(GPT2ForSequenceClassification):
+    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.transformer = GPT2Model(config)
+        self.score = GPT2ClassificationHeadCustomSimple(config)
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+
+# Version with custom forward 1 #
+
+class GPT2ClassificationHeadCustom(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        hidden_size = config.n_embd
+        self.dense_1_input = nn.Linear(hidden_size, 2 * hidden_size)
+        self.dense_1_hidden = nn.Linear(hidden_size, 2 * hidden_size)
+        self.dense_2 = nn.Linear(4 * hidden_size, hidden_size)
+        self.dropout = nn.Dropout(config.resid_pdrop)
+        self.out_proj = nn.Linear(hidden_size, config.num_labels, bias=False)
+
+    def forward(self, x, **kwargs):
+        if 'hidden_states' in kwargs and kwargs['hidden_states'] is not None:
+            # Get hidden states from second from the end
+            hidden = kwargs['hidden_states'][-2]
+        else:
+            hidden = torch.zeros(x.size(), dtype=x.dtype, device=x.device)
+
+        x = self.dense_1_input(x)
+        x = torch.relu(x)
+        x = self.dropout(x)
+
+        hidden = self.dense_1_hidden(hidden)
+        hidden = torch.relu(hidden)
+        hidden = self.dropout(hidden)
+
+        x = torch.cat((x, hidden), dim=2)
+        x = self.dense_2(x)
+        x = torch.relu(x)
+        x = self.dropout(x)
+
+        x = self.out_proj(x)
+        return x
+
+
+class GPT2ForSequenceClassificationCustom(GPT2ForSequenceClassification):
+    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.transformer = GPT2Model(config)
+        self.score = GPT2ClassificationHeadCustom(config)
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states or self.config.use_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states, hidden_states=transformer_outputs.hidden_states)
+
+        if input_ids is not None:
+            batch_size, sequence_length = input_ids.shape[:2]
+        else:
+            batch_size, sequence_length = inputs_embeds.shape[:2]
+
+        assert (
+            self.config.pad_token_id is not None or batch_size == 1
+        ), "Cannot handle batch sizes > 1 if no padding token is defined."
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                sequence_lengths = torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1
+            else:
+                sequence_lengths = -1
+                logger.warning(
+                    f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                    "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+                )
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )

roberta.py

@@ -0,0 +1,322 @@
+from typing import Optional, Union, Tuple
+
+import torch
+from torch import nn
+from torch.nn import MSELoss, CrossEntropyLoss, BCEWithLogitsLoss
+from transformers import RobertaForSequenceClassification, RobertaModel
+from transformers.modeling_outputs import SequenceClassifierOutput
+
+
+# Simple version #
+
+class RobertaClassificationHeadCustomSimple(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config):
+        super().__init__()
+        hidden_size = config.hidden_size
+        self.dense_1 = nn.Linear(hidden_size, 2 * hidden_size)
+        self.dense_2 = nn.Linear(2 * hidden_size, hidden_size)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.out_proj = nn.Linear(hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+
+        x = self.dense_1(x)
+        x = torch.relu(x)
+        x = self.dropout(x)
+
+        x = self.dense_2(x)
+        x = torch.relu(x)
+        x = self.dropout(x)
+
+        x = self.out_proj(x)
+        return x
+
+
+class RobertaForSequenceClassificationCustomSimple(RobertaForSequenceClassification):
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        self.classifier = RobertaClassificationHeadCustomSimple(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+
+# Version with custom forward 1 #
+
+class RobertaClassificationHeadCustom(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.use_hidden_states = config.use_hidden_states
+        hidden_size = config.hidden_size
+        if self.use_hidden_states:
+            hidden_size *= 2
+
+        self.dense_1 = nn.Linear(hidden_size, 2 * hidden_size)
+        self.dense_2 = nn.Linear(2 * hidden_size, hidden_size)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.out_proj = nn.Linear(hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        if 'hidden_states' in kwargs and kwargs['hidden_states'] is not None:
+            if self.use_hidden_states:
+                x = torch.cat(
+                    (
+                        features[:, 0, :],
+                        # take <s> token (equiv. to [CLS]) from hidden states from second from the end
+                        kwargs['hidden_states'][-2][:, 0, :]
+                    ),
+                    dim=1
+                )
+            else:
+                x = features[:, 0, :] + kwargs['hidden_states'][-2][:, 0, :]
+            del kwargs['hidden_states']
+        else:
+            x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+            if self.use_hidden_states:
+                x = torch.cat(
+                    (
+                        features[:, 0, :],
+                        torch.zeros(x.size(), dtype=x.dtype, device=x.device)
+                    ),
+                    dim=1
+                )
+
+        x = self.dense_1(x)
+        x = torch.relu(x)
+        x = self.dropout(x)
+
+        x = self.dense_2(x)
+        x = torch.relu(x)
+        x = self.dropout(x)
+
+        x = self.out_proj(x)
+        return x
+
+
+class RobertaForSequenceClassificationCustom(RobertaForSequenceClassification):
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        self.classifier = RobertaClassificationHeadCustom(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.roberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states or self.config.use_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = outputs[0]
+        logits = self.classifier(sequence_output, hidden_states=outputs.hidden_states)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+# Version with custom forward 2 #
+
+class RobertaClassificationHeadCustomAlternative(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config):
+        super().__init__()
+        hidden_size = config.hidden_size
+
+        self.dense_1_input = nn.Linear(hidden_size, 2 * hidden_size)
+        self.dense_1_hidden = nn.Linear(hidden_size, 2 * hidden_size)
+        self.dense_2 = nn.Linear(4 * hidden_size, hidden_size)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.out_proj = nn.Linear(hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        if 'hidden_states' in kwargs and kwargs['hidden_states'] is not None:
+            # take <s> token (equiv. to [CLS]) from hidden states from second from the end
+            hidden = kwargs['hidden_states'][-2][:, 0, :]
+            del kwargs['hidden_states']
+        else:
+            hidden = torch.zeros(x.size(), dtype=x.dtype, device=x.device)
+
+        x = self.dense_1_input(x)
+        x = torch.relu(x)
+        x = self.dropout(x)
+
+        hidden = self.dense_1_hidden(hidden)
+        hidden = torch.relu(hidden)
+        hidden = self.dropout(hidden)
+
+        x = torch.cat((x, hidden), dim=1)
+        x = self.dense_2(x)
+        x = torch.relu(x)
+        x = self.dropout(x)
+
+        x = self.out_proj(x)
+        return x
+
+
+class RobertaForSequenceClassificationCustomAlternative(RobertaForSequenceClassification):
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        self.classifier = RobertaClassificationHeadCustomAlternative(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.roberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states or self.config.use_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = outputs[0]
+        logits = self.classifier(sequence_output, hidden_states=outputs.hidden_states)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )