WIP

Author: Ye Shaokai

.vscode/launch.json

@@ -250,7 +250,7 @@
                 "--deepspeed", "scripts/zero3.json",
                 "--model_name_or_path", "lmms-lab/llava-onevision-qwen2-0.5b-ov",
                 "--version", "qwen_1_5",
-                "--data_path", "scripts/train/vis_supervision_strong_baseline.yaml",
+                "--data_path", "scripts/train/simple_tim_top5_cut.yaml",
                 "--video_folder", "/data/shaokai/",
                 "--mm_tunable_parts", "mm_vision_tower,mm_mlp_adapter,mm_language_model",
                 "--mm_vision_tower_lr", "2e-6",
@@ -294,11 +294,12 @@
                 "--val_metadata", "/data/shaokai/epic-kitchens-100-annotations/EPIC_100_validation.csv",
                 "--llava_num_frames", "16",
                 "--clip_length", "16",
-                "--action_representation", "GT_random_narration",
+                "--action_representation", "topk_narration_cut_key",
                 "--topk_predictions", "5",
-                "--eval_steps", "10",
+                "--eval_steps", "1",
                 "--vision_supervision", "all_newlines",
-                "--action_types", "97,300,3806"
+                "--action_types", "97,300,3806",
+                "--n_narration", "5"
             ],
             "console": "integratedTerminal",
             "justMyCode": false,

llava/action/ek_eval.py

@@ -199,8 +199,8 @@ def evaluate_on_EK100(eval_args,
 
     crop_size = 336
     labels, mapping_vn2narration, mapping_vn2act, verb_maps, noun_maps = generate_label_map(Path(eval_args.val_metadata).parent,                                                                                            
-                                                                                            eval_args.action_representation,
-                                                                                            cache_file =  Path(eval_args.val_metadata).parent / 'nlp_cache.pkl')
+                                                                                            eval_args.action_representation)
+                                                                                            
 
     if eval_args.action_predictions:
         with open(eval_args.action_predictions, 'r') as f: