import os import json import random import warnings import requests import cv2 import numpy as np import torch import torch.multiprocessing as mp from tqdm import trange from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor from qwen_vl_utils import process_vision_info from PIL import Image import re warnings.filterwarnings("ignore") # ================== 可根据需要修改的路径与配置 ================== # MODEL_PATH = 'CodeGoat24/UnifiedReward-qwen-7b' MODEL_PATH = "qunwang13/vr-thinker" metadata = './VideoAlign/videogen_metadata.json' model_rwft = 'Oct31_1' steps = 2000 model_base = 'pretrain' video_root_rwft = f'/nfs/ywang29/Reward_finetuning/VideoX-Fun/validation_samples/samples_videogen_eval/wan-videos-t2v-512*288-30-seed-1/output_{model_rwft}/steps_{steps}' video_root_base = f'/nfs/ywang29/Reward_finetuning/VideoX-Fun/validation_samples/samples_videogen_eval1/wan-videos-t2v-512*288-30-seed42/output_{model_base}' output_json_path = f'preference_eval/results1/{model_base}_{model_rwft}_unireward.json' os.makedirs(os.path.dirname(output_json_path), exist_ok=True) # ============================================================ def load_metadata(meta_path): if meta_path.endswith('.json'): with open(meta_path, 'r', encoding='utf-8') as f: data = json.load(f) # 期望是 { "0001": {...}, "0002": {...} } 或 { "0001": "prompt", ...} items = sorted(data.items(), key=lambda kv: int(kv[0])) return items elif meta_path.endswith('.txt'): with open(meta_path, 'r', encoding='utf-8') as f: lines = [line.strip() for line in f.readlines()] # 若是 txt，就用行号当 prefix，整行当 prompt return [(str(i), line) for i, line in enumerate(lines)] else: raise ValueError(f"Unsupported metadata file: {meta_path}") def extract_prompt(value): """兼容两种格式：value 是字符串 / 或 dict 里含 'prompt' 字段。""" if isinstance(value, str): return value if isinstance(value, dict): # 优先常见字段名 for k in ['prompt', 'caption', 'text']: if k in value: return value[k] # 兜底：把可序列化内容转成字符串 return json.dumps(value, ensure_ascii=False) return str(value) def get_results(model, processor, video_path_1, video_path_2, prompt): prompt_for_videos = prompt dim_name_1, dim_explain_1 = "Temporal Alignment (TA)", "How well the video adheres to the temporal aspects of the prompt." dim_name_2, dim_explain_2 = "Video Quality (VQ)", "The visual and aesthetic quality of the video." dim_name_3, dim_explain_3 = "Motion Quality (MQ)", "The smoothness and realism of the motion in the video." N = 49 prompt_text = \ f"""Task Description: Your task is to compare two videos generated based on the same prompt by analyzing their frames in detail and provide an overall judgment along with a judgment for each dimension. This involves: - Iterative reasoning, - Zooming in on details, - Dynamically selecting frames for further analysis. The provided frames are downsampled from these videos: - Video 1: First four input frames. - Video 2: Next four input frames. The prompt is: {prompt_for_videos} Evaluation Dimensions: 1. **{dim_name_1}**: {dim_explain_1} 2. **{dim_name_2}**: {dim_explain_2} 3. **{dim_name_3}**: {dim_explain_3} Frames and Analysis Rules: - 8 sampled frames are provided, evenly downsampled from {N} frames. - First 4 input frames sampled from {N/2} actual frames of Video 1, next 4 input frames sampled from {N/2} actual frames of Video - Insufficient frames? Request more using the tool. Format Requirement: 1. Snapshot: Every time you receive new visual information, summarize any information that might be useful for your final judgment within tags. 2. Think:\nPlace all reasoning content within tags.\n\n3. Answer:\nIf the final answer can be determined, output the answer within tags. If the answer is still uncertain, output the recommended answer and confidence level within tags. - For TA, MQ, VQ, and OA: 1 represents Video 1 is better, 2 represents Video 2 is better, and 0 represents a Tie. - For CF (Confidence level): 1 (low), 2 (medium), 3 (high), 4 (very high), 5 (confirmed). Examples:\nTA=0, VQ=1, MQ=0, OA=1, CF=2 TA=1, VQ=1, MQ=0, OA=1.""" sys_prompt = \ """You are a helpful assistant.\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within XML tags: \n{\"type\": \"function\", \"function\": {\"name\": \"select_frames\", \"description\": \"Select frames from a video.\", \"parameters\": {\"type\": \"object\", \"properties\": {\"target_frames\": {\"type\": \"array\", \"description\": \"List of frame indices to select from the video (no more than 8 frames in total).\", \"items\": {\"type\": \"integer\", \"description\": \"Frame index from 1 to N.\"}}}, \"required\": [\"target_frames\"]}}}\n\n\nFor each function call, return a json object with function name and arguments within XML tags:\n\n{\"name\": , \"arguments\": }""" content_list = [{"type": "video", "video": video_path, "nframes": 4 } for video_path in [video_path_1, video_path_2]] content_list.append({"type": "text", "text": prompt_text}) messages = [ { "role": "system", "content": sys_prompt, }, { "role": "user", "content": content_list, } ] text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) image_inputs, video_inputs, video_kwargs = process_vision_info(messages, return_video_kwargs=True) inputs = processor(text=[text], images=image_inputs, videos=video_inputs, return_tensors="pt", **video_kwargs).to(model.device) with torch.inference_mode(): generated_ids = model.generate(**inputs, max_new_tokens=1024) generated_ids_trimmed = [out[len(inp):] for inp, out in zip(inputs.input_ids, generated_ids)] output_text = processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True)[0] import pdb # pdb.set_trace() m = re.search(r"\s*(.*?)\s*", output_text, flags=re.S) block = m.group(1) if m else "" # 提取形如 KEY=VALUE 的对，并转成 int result = {k: int(v) for k, v in re.findall(r"\b([A-Z]{2})\s*=\s*([+-]?\d+)", block)} # print(result) return result def worker(rank, world_size, items, output_json_path): device = f'cuda:{rank}' torch.cuda.set_device(device) # 每个进程独立加载一份模型到自己的 GPU（避免跨卡拷贝） model = Qwen2_5_VLForConditionalGeneration.from_pretrained( MODEL_PATH, torch_dtype="auto", device_map={"": device} ) processor = AutoProcessor.from_pretrained(MODEL_PATH) A_win = 0 B_win = 0 tie = 0 results = [] # 每个进程处理 i % world_size == rank 的样本子集 local_indices = [i for i in range(len(items)) if i % world_size == rank] # tqdm 在多进程里容易抢 stdout，这里给每个进程单独的条，或直接关闭 for idx in local_indices: prefix, meta_val = items[idx] prompt = extract_prompt(meta_val) base_path = f'{video_root_base}/{prefix}.mp4' rwft_path = f'{video_root_rwft}/{prefix}.mp4' # 简单的存在性检查，缺文件就跳过，避免中断 if not (os.path.isfile(base_path) and os.path.isfile(rwft_path)): # 你也可以把缺失样本记录下来 continue try: output = get_results(model, processor, rwft_path, base_path, prompt) results.append(output) except Exception as e: # 个别样本异常时跳过，保证整体不崩 print(f"[Rank {rank}] Error on {prefix}: {e}") continue # 写入局部统计 part_path = f"{output_json_path}.part{rank}" with open(part_path, 'w') as f: json.dump(results, f, indent=2) def aggregate(world_size, output_json_path): all_results = [] for r in range(world_size): part_path = f"{output_json_path}.part{r}" if not os.path.isfile(part_path): continue with open(part_path, 'r') as f: d = json.load(f) all_results += d with open(output_json_path, 'w') as f: json.dump(all_results, f, indent=4) # 清理所有分片文件 import glob for p in glob.glob(f"{output_json_path}.part*"): try: os.remove(p) except FileNotFoundError: pass def main(): # 读入任务列表 items = load_metadata(metadata) # 自动检测 GPU 数量；也可用环境变量限制 GPU 可见性 world_size = torch.cuda.device_count() if world_size <= 0: raise RuntimeError("No CUDA device is available.") print(f"Using {world_size} GPUs.") # 先清理旧的 part for r in range(world_size): p = f"{output_json_path}.part{r}" if os.path.isfile(p): os.remove(p) mp.set_start_method("spawn", force=True) ctx = mp.get_context("spawn") procs = [] for rank in range(world_size): p = ctx.Process(target=worker, args=(rank, world_size, items, output_json_path), daemon=False) p.start() procs.append(p) for p in procs: p.join() # 汇总 aggregate(world_size, output_json_path) if __name__ == "__main__": main()