Update modeling.py

modeling.py

@@ -1,51 +1,47 @@
 import torch
 import torch.nn as nn
-from transformers.modeling_outputs import SequenceClassifierOutput
-from transformers.models.roberta.modeling_roberta import RobertaPreTrainedModel, RobertaModel
-
-
-class RobertaForSequenceClassification(RobertaPreTrainedModel):
-    def __init__(self, config):
-        super().__init__(config)
-
-        self.num_labels = config.num_labels
-        self.roberta = RobertaModel(config)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
-
-        # Load weights
-        self.post_init()
-
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        labels=None,
-        **kwargs
-    ):
-        outputs = self.roberta(
-            input_ids=input_ids,
-            attention_mask=attention_mask,
-            token_type_ids=token_type_ids,
+import torch.nn.functional as F
+from transformers import PreTrainedModel, AutoModel, AutoConfig, ModelOutput
+
+class Encoder(nn.Module):
+    def __init__(self, base_encoder):
+        super().__init__()
+        self.encoder = base_encoder
+
+    def forward(self, inputs):
+        outputs = self.encoder(**inputs, output_hidden_states=True)
+        last_hidden = outputs.hidden_states[-1]
+        mask = inputs["attention_mask"].unsqueeze(-1).float()
+        pooled = (last_hidden * mask).sum(1) / mask.sum(1).clamp(min=1e-9)
+        return F.normalize(pooled, p=2, dim=1)
+
+class Classifier(nn.Module):
+    def __init__(self, input_dim=768, num_classes=28):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            nn.Linear(input_dim, 512),
+            nn.LayerNorm(512),
+            nn.GELU(),
+            nn.Dropout(0.25),
+            nn.Linear(512, num_classes),
         )
 
-        pooled_output = outputs[1]  # CLS token
-        pooled_output = self.dropout(pooled_output)
-        logits = self.classifier(pooled_output)
+    def forward(self, x):
+        return self.mlp(x)
 
-        loss = None
-        if labels is not None:
-            if self.num_labels == 1:
-                loss_fct = nn.MSELoss()
-                loss = loss_fct(logits.squeeze(), labels.squeeze())
-            else:
-                loss_fct = nn.BCEWithLogitsLoss()
-                loss = loss_fct(logits, labels.float())
+class RobertaEmoPillars(PreTrainedModel):
+    config_class = AutoConfig
 
-        return SequenceClassifierOutput(
-            loss=loss,
-            logits=logits,
-            hidden_states=outputs.hidden_states,
-            attentions=outputs.attentions,
-        )
+    def __init__(self, config):
+        super().__init__(config)
+        base_encoder = AutoModel.from_config(config)  # IMPORTANT: use from_config
+        self.encoder = Encoder(base_encoder)
+        self.classifier = Classifier(input_dim=base_encoder.config.hidden_size,
+                                     num_classes=config.num_labels)
+        self.post_init()  # ensures HF weights init
+
+    def forward(self, input_ids=None, attention_mask=None):
+        inputs = {"input_ids": input_ids, "attention_mask": attention_mask}
+        emb = self.encoder(inputs)
+        logits = self.classifier(emb)
+        return ModelOutput(logits=logits)