| 108 | return (predictions == labels).float().mean() |
| 109 | |
| 110 | class OptILMClassifier(nn.Module): |
| 111 | def __init__(self, base_model, num_labels): |
| 112 | super().__init__() |
| 113 | self.base_model = base_model |
| 114 | self.effort_encoder = nn.Sequential( |
| 115 | nn.Linear(1, 64), |
| 116 | nn.ReLU(), |
| 117 | nn.Linear(64, 64), |
| 118 | nn.ReLU() |
| 119 | ) |
| 120 | self.classifier = nn.Linear(base_model.config.hidden_size + 64, num_labels) |
| 121 | |
| 122 | def forward(self, input_ids, attention_mask, effort): |
| 123 | outputs = self.base_model(input_ids=input_ids, attention_mask=attention_mask) |
| 124 | pooled_output = outputs.last_hidden_state[:, 0] # Shape: (batch_size, hidden_size) |
| 125 | effort_encoded = self.effort_encoder(effort.unsqueeze(1)) # Shape: (batch_size, 64) |
| 126 | combined_input = torch.cat((pooled_output, effort_encoded), dim=1) |
| 127 | logits = self.classifier(combined_input) |
| 128 | return logits |
| 129 | |
| 130 | def train(model, train_dataloader, val_dataloader, optimizer, scheduler, num_epochs, patience, clip_value): |
| 131 | best_val_accuracy = 0.0 |