fixed known issues

Author: Haozhe Qi

llava/action/ek_eval.py

@@ -341,7 +341,9 @@ def collate_fn(batch):
     meta_data = None
     if eval_args.learn_neighbor_actions:
         from llava.action.generate_interval_pred import  get_lookup_dict
-        lookup_table = get_lookup_dict(eval_args.val_metadata, test_type = eval_args.test_type, pseudo_folder = eval_args.pseudo_folder)
+        if eval_args.test_type.startswith('temporal_cot'):
+
+            lookup_table = get_lookup_dict(eval_args.val_metadata, test_type = eval_args.test_type, pseudo_folder = eval_args.pseudo_folder)
 
 
     for idx, (frames, mc_data, time_meta, global_index) in tqdm(enumerate(val_dataloader)):                  

llava/action/generate_interval_pred.py

@@ -131,6 +131,7 @@ def get_lookup_dict(ann_file, test_type = 'base', delta = 3, pseudo_folder = Non
 
     pseudo_dict = None
     if test_type == 'temporal_cot':
+        assert os.path.exists(pseudo_folder), f"Folder {pseudo_folder} does not exist"
         pseudo_dict = get_pseudo_dict(pseudo_folder)
 
     for vid, intervals in vid_to_intervals.items():

llava/action/llava_inference.py

@@ -55,6 +55,8 @@ def llava_inference(
             question_type = "caption"
         elif test_type == 'temporal_cot':
             question_type = 'temporal_cot'
+        elif test_type == 'temporal_cot_oracle':
+            question_type = 'temporal_cot_oracle'
 
         if  test_type == 'caption_then_answer':        
             caption_answer = llava_inference([video_frames], 

llava/action/utils.py

@@ -253,7 +253,7 @@ def format_task_related_prompt(question, question_type, meta_data = None, perspe
             prefix = f"{perspective_prefix}\n"
             assert isinstance(question, list)
             suffix = ", ".join(question)           
-            suffix = f"{prev2_offset} seconds ago, you started an action {prev2_narration}. {prev1_offset} seconds ago, you started an action {prev1_narration}. What action are you currently performing? Explain your thoughts and then answer the question. Here are the options of actions you can select:\n" + suffix 
+            suffix = f"{prev2_offset} seconds ago, you started an action {prev2_narration}. {prev1_offset} seconds ago, you started an action {prev1_narration}. What action are you currently performing? Here are the options of actions you can select:\n" + suffix 
             ret = prefix + suffix         
         else:
             action_rep_suffix = "Given multiple choices, format your answer briefly such as 'A. move knife'. "