Create model.py

model.py

@@ -0,0 +1,81 @@
+from abc import ABCMeta
+import torch
+from transformers.pytorch_utils import nn
+import torch.nn.functional as F
+from transformers import RobertaModel, RobertaForSequenceClassification, PreTrainedModel
+from transformers.modeling_outputs import SequenceClassifierOutput
+from transformers import RobertaConfig
+from transformers import PretrainedConfig
+
+class RobertaABSAConfig(PretrainedConfig):
+    model_type = "robertaCNNForSequenceClassification"
+
+    def __init__(self,
+                 num_classes=2,
+                 embed_dim=768,
+                 conv_out_channels=256,  # New parameter for Conv1d
+                 conv_kernel_size=3,
+                 fc_hidden=128,           # New parameter for FC layer
+                 dropout_rate=0.1,
+                 num_layers=12,
+                 **kwargs):
+        super().__init__(**kwargs)
+        self.num_classes = num_classes
+        self.embed_dim = embed_dim
+        self.conv_out_channels = conv_out_channels  # Assign Conv1d output channels
+        self.conv_kernel_size = conv_kernel_size    # Assign Conv1d kernel size
+        self.fc_hidden = fc_hidden                  # Assign FC layer hidden units
+        self.dropout_rate = dropout_rate
+        self.num_layers = num_layers
+        self.id2label = {
+            0: "fake",
+            1: "true",
+        }
+        self.label2id = {
+            "fake": 0,
+            "true": 1,
+        }
+
+                     
+class RobertaCNNForSequenceClassification(PreTrainedModel, metaclass=ABCMeta):
+    config_class = RobertaABSAConfig
+
+    def __init__(self, config):
+        super(RobertaCNNForSequenceClassification, self).__init__(config)
+        self.num_classes = config.num_classes
+        self.embed_dim = config.embed_dim
+        self.num_layers = config.num_layers
+        self.conv_out_channels = config.conv_out_channels
+        self.conv_kernel_size = config.conv_kernel_size
+        self.dropout = nn.Dropout(config.dropout_rate)
+        self.roberta = RobertaModel.from_pretrained('roberta-base',
+                                                  output_hidden_states=True,
+                                                  output_attentions=False)
+        print("RoBERTa Model Loaded")
+        self.conv1d = nn.Conv1d(in_channels=self.embed_dim, out_channels=self.conv_out_channels, kernel_size=self.conv_kernel_size)
+        self.fc = nn.Linear(self.conv_out_channels, self.num_classes)
+
+    def forward(self, input_ids, attention_mask, labels=None):
+        roberta_output = self.roberta(input_ids=input_ids, attention_mask=attention_mask)
+        hidden_states = roberta_output["hidden_states"]
+
+        hidden_states = torch.stack([hidden_states[layer_i][:, 0].squeeze()
+                                     for layer_i in range(0, self.num_layers)], dim=-1)  # noqa
+        hidden_states = hidden_states.view(-1, self.num_layers, self.embed_dim)
+        hidden_states = hidden_states.permute(0, 2, 1)  # Permute to match Conv1d input shape
+        conv_output = self.conv1d(hidden_states)
+        conv_output = F.relu(conv_output)
+        conv_output = F.max_pool1d(conv_output, kernel_size=conv_output.size(2))  # Global Max Pooling
+        conv_output = conv_output.squeeze(-1)
+        conv_output = self.dropout(conv_output)
+        logits = self.fc(conv_output)
+        loss = None
+        if labels is not None:
+            loss = F.cross_entropy(logits, labels)
+        out = SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=roberta_output.hidden_states,
+            attentions=roberta_output.attentions,
+        )
+        return out