Upload trainer.py

LLAVA-Cherry/trainer.py

@@ -17,6 +17,11 @@
 The Trainer class, to easily train a 🤗 Transformers from scratch or finetune it on a new task.
 """
 
+#########################################################
+from datetime import datetime
+has_run = False
+#########################################################
+
 import contextlib
 import copy
 import functools
@@ -2762,6 +2767,13 @@ class Trainer:
         Return:
             `torch.Tensor`: The tensor with training loss on this batch.
         """
+
+        # #######################################################
+        # # import pdb; pdb.set_trace()
+        # import pprint 
+        # pprint.pprint(inputs)
+        # #######################################################
+
         model.train()
         inputs = self._prepare_inputs(inputs)
 
@@ -2776,16 +2788,31 @@ class Trainer:
         del inputs['dataset_id']
         del inputs['data_info']
         #######################################################
+        
 
         with self.compute_loss_context_manager():
-            loss = self.compute_loss(model, inputs)
+            # loss = self.compute_loss(model, inputs)
+            (loss, outputs) = self.compute_loss(model, inputs,return_outputs=True)
+
+        
+        import pprint 
+        # pprint.pprint(outputs)
+        # import pdb; pdb.set_trace()
+        last_token_logits_yes = outputs.logits[:, -1, :]
+        yes_target_token_id = 4874
+        yes_target_logprob = torch.log_softmax(last_token_logits_yes, dim=-1)[0, yes_target_token_id].item()
 
         #######################################################
         import json
         for i in range(len(data_info_temp)):
-            data_info_temp[i]['loss'] = float(loss[0][i]) 
+            # data_info_temp[i]['loss'] = float(loss[0][i]) 
+            data_info_temp[i]['yes_target_logprob'] = yes_target_logprob
+            data_info_temp[i]['logits_shape'] = outputs.logits.shape
 
-        file_path = '/data/zbz5349/ICLR_2024/ACL_2025/LLaVA_Fliter/inference_demo/cherry_loss_infer_result.jsonl'
+        from datetime import datetime 
+        current_time = datetime.now().strftime('%Y_%m_%d')
+
+        file_path = '/data/zbz5349/ICLR_2024/ACL_2025/LLaVA_Fliter/inference_demo/cherry_AskLLM_infer_result_' + current_time + '.jsonl'
         with open(file_path, 'a', encoding='utf-8') as file:
             # json.dump(data_info_temp[0], file, ensure_ascii=False, indent=4)
             for content in data_info_temp:
@@ -2825,6 +2852,13 @@ class Trainer:
         else:
             labels = None
         outputs = model(**inputs)
+
+        # #######################################################
+        # import pdb; pdb.set_trace()
+        # import pprint
+        # pprint.pprint(outputs)
+        # #######################################################
+
         # Save past state if it exists
         # TODO: this needs to be fixed and made cleaner later.
         if self.args.past_index >= 0: