Update modeling_gpt2.py

modeling_gpt2.py

@@ -34,7 +34,7 @@ from transformers.modeling_outputs import (
     SequenceClassifierOutputWithPast,
     TokenClassifierOutput,
 )
-from transformers.modeling_utils import PreTrainedModel, SequenceSummary
+from transformers.modeling_utils import PreTrainedModel
 from transformers.pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
 from transformers.utils import (
     ModelOutput,
@@ -1106,6 +1106,105 @@ class GPT2CustomLMHeadModel(GPT2PreTrainedModel, GenerationMixin):
         )
 
 
+class GPT2SequenceSummary(nn.Module):
+    r"""
+    Compute a single vector summary of a sequence hidden states.
+
+    Args:
+        config ([`GPT2Config`]):
+            The config used by the model. Relevant arguments in the config class of the model are (refer to the actual
+            config class of your model for the default values it uses):
+
+            - **summary_type** (`str`) -- The method to use to make this summary. Accepted values are:
+
+                - `"last"` -- Take the last token hidden state (like XLNet)
+                - `"first"` -- Take the first token hidden state (like Bert)
+                - `"mean"` -- Take the mean of all tokens hidden states
+                - `"cls_index"` -- Supply a Tensor of classification token position (GPT/GPT-2)
+                - `"attn"` -- Not implemented now, use multi-head attention
+
+            - **summary_use_proj** (`bool`) -- Add a projection after the vector extraction.
+            - **summary_proj_to_labels** (`bool`) -- If `True`, the projection outputs to `config.num_labels` classes
+              (otherwise to `config.hidden_size`).
+            - **summary_activation** (`Optional[str]`) -- Set to `"tanh"` to add a tanh activation to the output,
+              another string or `None` will add no activation.
+            - **summary_first_dropout** (`float`) -- Optional dropout probability before the projection and activation.
+            - **summary_last_dropout** (`float`)-- Optional dropout probability after the projection and activation.
+    """
+
+    def __init__(self, config: GPT2Config):
+        super().__init__()
+
+        self.summary_type = getattr(config, "summary_type", "last")
+        if self.summary_type == "attn":
+            # We should use a standard multi-head attention module with absolute positional embedding for that.
+            # Cf. https://github.com/zihangdai/xlnet/blob/master/modeling.py#L253-L276
+            # We can probably just use the multi-head attention module of PyTorch >=1.1.0
+            raise NotImplementedError
+
+        self.summary = nn.Identity()
+        if hasattr(config, "summary_use_proj") and config.summary_use_proj:
+            if hasattr(config, "summary_proj_to_labels") and config.summary_proj_to_labels and config.num_labels > 0:
+                num_classes = config.num_labels
+            else:
+                num_classes = config.hidden_size
+            self.summary = nn.Linear(config.hidden_size, num_classes)
+
+        activation_string = getattr(config, "summary_activation", None)
+        self.activation: Callable = get_activation(activation_string) if activation_string else nn.Identity()
+
+        self.first_dropout = nn.Identity()
+        if hasattr(config, "summary_first_dropout") and config.summary_first_dropout > 0:
+            self.first_dropout = nn.Dropout(config.summary_first_dropout)
+
+        self.last_dropout = nn.Identity()
+        if hasattr(config, "summary_last_dropout") and config.summary_last_dropout > 0:
+            self.last_dropout = nn.Dropout(config.summary_last_dropout)
+
+    def forward(
+        self, hidden_states: torch.FloatTensor, cls_index: Optional[torch.LongTensor] = None
+    ) -> torch.FloatTensor:
+        """
+        Compute a single vector summary of a sequence hidden states.
+
+        Args:
+            hidden_states (`torch.FloatTensor` of shape `[batch_size, seq_len, hidden_size]`):
+                The hidden states of the last layer.
+            cls_index (`torch.LongTensor` of shape `[batch_size]` or `[batch_size, ...]` where ... are optional leading dimensions of `hidden_states`, *optional*):
+                Used if `summary_type == "cls_index"` and takes the last token of the sequence as classification token.
+
+        Returns:
+            `torch.FloatTensor`: The summary of the sequence hidden states.
+        """
+        if self.summary_type == "last":
+            output = hidden_states[:, -1]
+        elif self.summary_type == "first":
+            output = hidden_states[:, 0]
+        elif self.summary_type == "mean":
+            output = hidden_states.mean(dim=1)
+        elif self.summary_type == "cls_index":
+            if cls_index is None:
+                cls_index = torch.full_like(
+                    hidden_states[..., :1, :],
+                    hidden_states.shape[-2] - 1,
+                    dtype=torch.long,
+                )
+            else:
+                cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
+                cls_index = cls_index.expand((-1,) * (cls_index.dim() - 1) + (hidden_states.size(-1),))
+            # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
+            output = hidden_states.gather(-2, cls_index).squeeze(-2)  # shape (bsz, XX, hidden_size)
+        elif self.summary_type == "attn":
+            raise NotImplementedError
+
+        output = self.first_dropout(output)
+        output = self.summary(output)
+        output = self.activation(output)
+        output = self.last_dropout(output)
+
+        return output
+
+
 @add_start_docstrings(
     """
 The GPT2 Model transformer with a language modeling and a multiple-choice classification head on top e.g. for
@@ -1123,7 +1222,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
         config.num_labels = 1
         self.transformer = GPT2Model(config)
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
-        self.multiple_choice_head = SequenceSummary(config)
+        self.multiple_choice_head = GPT2SequenceSummary(config)
 
         # Model parallel
         self.model_parallel = False