feat: add support for SequenceClassification

config.json

@@ -3,7 +3,8 @@
     "AutoConfig": "configuration_xlm_roberta.XLMRobertaFlashConfig",
     "AutoModel": "modeling_xlm_roberta.XLMRobertaModel",
     "AutoModelForPreTraining": "modeling_xlm_roberta.XLMRobertaForPreTraining",
-    "AutoModelForMaskedLM": "modeling_xlm_roberta.XLMRobertaForMaskedLM"
+    "AutoModelForMaskedLM": "modeling_xlm_roberta.XLMRobertaForMaskedLM",
+    "AutoModelForSequenceClassification":"modeling_xlm_roberta.XLMRobertaForSequenceClassification"
   },
   "attention_probs_dropout_prob": 0.1,
   "bos_token_id": 0,

modeling_xlm_roberta.py

@@ -18,10 +18,11 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 import torch.utils.checkpoint
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 from einops import rearrange
 from transformers import PretrainedConfig
 from transformers.modeling_utils import PreTrainedModel
-from transformers.modeling_outputs import MaskedLMOutput
+from transformers.modeling_outputs import MaskedLMOutput, SequenceClassifierOutput
 from transformers.models.xlm_roberta.modeling_xlm_roberta import XLMRobertaLMHead
 
 from transformers.models.bert.modeling_bert import (
@@ -429,7 +430,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
             for key, value in kwargs.items():
                 if value is not None:
                     logger.warning(
-                        'Flash attention implementation does not support kwargs: %s',
+                        "Flash attention implementation does not support kwargs: %s",
                         key,
                     )
 
@@ -834,47 +835,47 @@ def inv_remap_state_dict(state_dict, config: PretrainedConfig):
         if not last_layer_subset or d != (config.num_hidden_layers - 1):
             Wqkv_weights = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wqkv.weight")
             Wqkv_biases = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wqkv.bias")
-            state_dict[
-                f"bert.encoder.layers.{d}.attention.self.query.weight"
-            ] = Wqkv_weights[: Wqkv_weights.shape[0] // 3, :]
-            state_dict[
-                f"bert.encoder.layers.{d}.attention.self.key.weight"
-            ] = Wqkv_weights[
-                Wqkv_weights.shape[0] // 3 : 2 * Wqkv_weights.shape[0] // 3, :
-            ]
-            state_dict[
-                f"bert.encoder.layers.{d}.attention.self.value.weight"
-            ] = Wqkv_weights[2 * Wqkv_weights.shape[0] // 3 :, :]
-            state_dict[
-                f"bert.encoder.layers.{d}.attention.self.query.bias"
-            ] = Wqkv_biases[: Wqkv_biases.shape[0] // 3]
-            state_dict[
-                f"bert.encoder.layers.{d}.attention.self.key.bias"
-            ] = Wqkv_biases[Wqkv_biases.shape[0] // 3 : 2 * Wqkv_biases.shape[0] // 3]
-            state_dict[
-                f"bert.encoder.layers.{d}.attention.self.value.bias"
-            ] = Wqkv_biases[2 * Wqkv_biases.shape[0] // 3 :]
+            state_dict[f"bert.encoder.layers.{d}.attention.self.query.weight"] = (
+                Wqkv_weights[: Wqkv_weights.shape[0] // 3, :]
+            )
+            state_dict[f"bert.encoder.layers.{d}.attention.self.key.weight"] = (
+                Wqkv_weights[
+                    Wqkv_weights.shape[0] // 3 : 2 * Wqkv_weights.shape[0] // 3, :
+                ]
+            )
+            state_dict[f"bert.encoder.layers.{d}.attention.self.value.weight"] = (
+                Wqkv_weights[2 * Wqkv_weights.shape[0] // 3 :, :]
+            )
+            state_dict[f"bert.encoder.layers.{d}.attention.self.query.bias"] = (
+                Wqkv_biases[: Wqkv_biases.shape[0] // 3]
+            )
+            state_dict[f"bert.encoder.layers.{d}.attention.self.key.bias"] = (
+                Wqkv_biases[Wqkv_biases.shape[0] // 3 : 2 * Wqkv_biases.shape[0] // 3]
+            )
+            state_dict[f"bert.encoder.layers.{d}.attention.self.value.bias"] = (
+                Wqkv_biases[2 * Wqkv_biases.shape[0] // 3 :]
+            )
         else:
             Wq_weight = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wq.weight")
             Wkv_weights = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wkv.weight")
             Wq_bias = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wq.bias")
             Wkv_biases = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wkv.bias")
-            state_dict[
-                f"bert.encoder.layers.{d}.attention.self.query.weight"
-            ] = Wq_weight
-            state_dict[
-                f"bert.encoder.layers.{d}.attention.self.key.weight"
-            ] = Wkv_weights[: Wkv_weights.shape[0] // 2, :]
-            state_dict[
-                f"bert.encoder.layers.{d}.attention.self.value.weight"
-            ] = Wkv_weights[Wkv_weights.shape[0] // 2 :, :]
+            state_dict[f"bert.encoder.layers.{d}.attention.self.query.weight"] = (
+                Wq_weight
+            )
+            state_dict[f"bert.encoder.layers.{d}.attention.self.key.weight"] = (
+                Wkv_weights[: Wkv_weights.shape[0] // 2, :]
+            )
+            state_dict[f"bert.encoder.layers.{d}.attention.self.value.weight"] = (
+                Wkv_weights[Wkv_weights.shape[0] // 2 :, :]
+            )
             state_dict[f"bert.encoder.layers.{d}.attention.self.query.bias"] = Wq_bias
             state_dict[f"bert.encoder.layers.{d}.attention.self.key.bias"] = Wkv_biases[
                 : Wkv_biases.shape[0] // 2
             ]
-            state_dict[
-                f"bert.encoder.layers.{d}.attention.self.value.bias"
-            ] = Wkv_biases[Wkv_biases.shape[0] // 2 :]
+            state_dict[f"bert.encoder.layers.{d}.attention.self.value.bias"] = (
+                Wkv_biases[Wkv_biases.shape[0] // 2 :]
+            )
 
     def inv_key_mapping_ln(key):
         key = re.sub(r"bert.emb_ln.", "bert.embeddings.LayerNorm.", key)
@@ -946,3 +947,117 @@ def inv_remap_state_dict(state_dict, config: PretrainedConfig):
     )
 
     return state_dict
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaClassificationHead with Roberta->XLMRoberta
+class XLMRobertaClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.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(config.hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = torch.tanh(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaForSequenceClassification with Roberta->XLMRoberta, ROBERTA->XLM_ROBERTA
+class XLMRobertaForSequenceClassification(XLMRobertaPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.roberta = XLMRobertaModel(config, add_pooling_layer=False)
+        self.classifier = XLMRobertaClassificationHead(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,
+            return_dict=return_dict,
+        )
+        sequence_output = outputs[0]
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            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,
+        )