Upload VideoX-Fun/get_prompt_qa3.py

VideoX-Fun/get_prompt_qa3.py

@@ -0,0 +1,90 @@
+import pandas as pd
+import json
+import pdb
+import csv
+import torch
+import os
+import itertools
+# df = pd.read_csv('/nfs/ywang29/video_reward/VideoAlign/datasets/train/ours_reasoning_dataset.csv')
+# df = pd.read_csv('/nfs/ywang29/Reward_finetuning/VideoX-Fun/ours_data_binary.csv', sep='\t')
+# df = df.drop_duplicates(subset=['prompt'])
+def clean_text(x: str) -> str:
+    """去掉多余的换行、制表符、引号"""
+    if not isinstance(x, str):
+        x = str(x)
+    return (
+        x.replace("\n", " ")     # 换行变空格
+         .replace("\r", " ")
+         .replace('"""', '"')
+         .replace("'''", "'")
+         .strip()
+    )
+
+path ='/nfs/ywang29/Reward_finetuning/VideoX-Fun/data/preprocess_general'
+root = '/nfs/ywang29/Reward_finetuning/VideoX-Fun'
+rg_dict = {}
+rg_dict1 = {}
+with open('/nfs/ywang29/vr_datasets.json', 'r') as f:
+    data = json.load(f)
+for sample in data:
+    # pdb.set_trace()
+    rg_dict1[sample['video_id'].split('.mp4')[0]] = sample['prompt']
+    rg_dict[sample['video_id']] = sample['real_video']
+
+
+new_data = []
+top_row = ['video_id', 'real_video_path', 'ref_video', 'prompt', 'questions', 'gt_answers']
+new_data.append(top_row)
+
+# pdb.set_trace()
+for file in os.listdir(path):
+    video_id = file.split('.pt')[0]
+    prompt = rg_dict1[video_id]
+    content = torch.load(f"{path}/{file}")
+
+    if len(content) == 0:
+        continue
+    for question in content.keys():
+        ref_video = content[question]['ref_video']
+        import pdb
+        # pdb.set_trace()
+        # ref_video = f'/s3-code/ywang29/video_reward/WanI2V/WanI2V-480p-{video_id}.mp4'
+        if not os.path.exists(ref_video):
+            continue
+
+        new_line = [video_id, ref_video, ref_video, clean_text(prompt), [question.replace('<video>', '')], [content[question]['response']]]
+        new_data.append(new_line)
+    
+    phy_question = f'''You are a careful video forensics assistant. You are given two videos: a test video (the first 24 frames) which is ai-generated and a real video (the last 12 frames). The test video is generated by the caption of the real video. The caption is "{prompt}"
+Your task is evaluating whether the test video shows physics-related defects. You should compared both videos and use the provided caption as high-level intent. Focus on physical plausibility, not style or aesthetics.
+You need to analyze it in these aspects:
+- Liquid flow irregularity – e.g., non-inertial or discontinuous flow, volume popping, gravity-inconsistent motion, impossible splashes.
+- Abnormal object deformation – e.g., rigid objects bending/stretching without cause, topology changes (parts merging/splitting).
+- Abnormal texture/material change – e.g., surface turns matte→glossy without cause, texture swimming/flicker detached from geometry.
+- Abnormal motion – e.g., inertia/acceleration violations, teleporting, time reversals, jitter not explained by camera motion.
+- Unnatural interpenetration – e.g., objects passing through each other or ground, missing collisions/contacts.
+If the prompt doesn't have related physical aspects, you should answer 'No'.
+'''
+# Output strictly as JSON with this schema (no extra text):
+# {{
+#     "liquid_flow_irregularity": "Yes/No",
+#     "abnormal_deformation":  "Yes/No",
+#     "abnormal_texture_change":    "Yes/No",  
+#     "abnormal_motion":   "Yes/No",  
+#     "unnatural_interpenetration":  "Yes/No",
+# }}
+    # phy_gt = '''```json\n{\n    "liquid_flow_irregularity": "No",\n    "abnormal_deformation": "No",\n    "abnormal_texture_change": "No",\n    "abnormal_motion": "No",\n    "unnatural_interpenetration": "No"\n}\n```'''
+    phy_gt = '{\n    "liquid_flow_irregularity": "No",\n    "abnormal_deformation": "No",\n    "abnormal_texture_change": "No",\n    "abnormal_motion": "No",\n    "unnatural_interpenetration": "No"\n}'
+
+    new_line = [video_id, ref_video, ref_video, clean_text(prompt), [phy_question], [phy_gt]]
+    # new_data.append(new_line)
+
+    # pdb.set_trace() 
+    
+with open(f'{root}/ours_data_large_binary_single_question_with_gt_general.csv', 'w', newline='', encoding='utf-8-sig') as csvfile:
+    # 创建一个 csv writer 对象
+    writer = csv.writer(csvfile, delimiter='\t')
+    # 写入所有数据行
+    writer.writerows(new_data)
+
+# * Fair: Some requirements are satisfied but some not.