#!/usr/bin/env python3 # -*- coding: utf-8 -*- import os import glob import string import torch import librosa import soundfile as sf import jiwer from tqdm.auto import tqdm from collections import defaultdict from transformers import Qwen2_5OmniForConditionalGeneration, Qwen2_5OmniProcessor from datasets import load_dataset from qwen_omni_utils import process_mm_info import datasets import csv # ---- 文本预处理 & 评分 ---- def preprocess_text(text: str) -> str: return text.lower().translate(str.maketrans("", "", string.punctuation)) def validate_prediction(pred: str, gt_text: str, bm: str) -> float: bm = bm.lower() if bm in ["llama-questions", "voicebench"]: return 1.0 if preprocess_text(gt_text) in preprocess_text(pred) else 0.0 if bm in ["mmar", "mmau"]: return 1.0 if gt_text.strip() in pred.strip() else 0.0 elif bm in ["speech-triavia-qa", "speech-web-questions"]: # 把预测结果左右空白去掉 pred_stripped = pred.strip() # 尝试把 gt_text 解析成列表 try: if gt_text.startswith("[") and gt_text.endswith("]"): gt_list = eval(gt_text) # 或者用 json.loads 更安全 else: gt_list = [gt_text.strip()] except: gt_list = [gt_text.strip()] # 改成：只要列表中任意一个 gt 子串出现在预测里，就算正确 score = 1.0 if any(gt.strip() in pred_stripped for gt in gt_list) else 0.0 if bm == "tedlium": wer = jiwer.wer(gt_text, pred) return max(0.0, 1.0 - wer) return 0.0 # ---- 对单个 checkpoint 评估，返回 (overall, {bm_modality: acc}) ---- def evaluate_checkpoint(ckpt_path: str, eval_dataset) -> tuple[float, dict]: print(f"\n>>> Loading model from {ckpt_path}") model = Qwen2_5OmniForConditionalGeneration.from_pretrained( ckpt_path, torch_dtype="auto", device_map="cuda:1" ) processor = Qwen2_5OmniProcessor.from_pretrained(ckpt_path) # 音频重采样缓存目录 resampled_dir = "/home/chenyifu/new_rl/dissrc/evaluategemini" os.makedirs(resampled_dir, exist_ok=True) # 准备错误记录文件 err_path = os.path.join( resampled_dir, f"errors_{os.path.basename(ckpt_path)}.csv" ) err_f = open(err_path, "w", newline="", encoding="utf-8") err_writer = csv.writer(err_f) err_writer.writerow([ "sample_id", "benchmark_type", "modality", "ground_truth", "prediction" ]) scores_by_bm_modality = defaultdict(list) for sample in tqdm(eval_dataset, desc=" Samples", leave=False): bm = sample["benchmark_type"].lower() if bm in ["uro-bench", "wavbench","speech-triavia-qa", "speech-web-questions"]: continue gt_text = sample["answer"] sample_id = sample["id"] system_prompt = ( "You are Qwen, a virtual human developed by the Qwen Team, " "Alibaba Group, capable of perceiving auditory and visual inputs, " "as well as generating text and speech." ) # 选 system_content if bm == "tedlium": system_content = ( "You are an expert in automatic speech recognition, " "and you are asked to recognize and output the content of this speech. " "Note! Just output the content corresponding to the speech, and don't have any other words" ) else: system_content = ( "Please answer the questions with clear and concise answers, " "do not output any other explanation" ) # —— Audio 模态 —— audio_path = sample["audio"].replace( "/root/autodl-tmp/", "/home/chenyifu/new_rl/" ) try: if os.path.exists(audio_path): info = sf.info(audio_path) if info.samplerate != 16000: data, sr = librosa.load(audio_path, sr=None) data_resampled = librosa.resample( data, orig_sr=sr, target_sr=16000 ) audio_path_for_model = os.path.join( resampled_dir, f"{sample_id}_resampled.wav" ) sf.write(audio_path_for_model, data_resampled, 16000) else: audio_path_for_model = audio_path # 如果有 transcription，一并加到 conv_audio if sample["transcription"]: conv_audio = [ {"role":"system", "content":[{"type":"text","text":system_prompt}]}, {"role":"user", "content":[ {"type":"audio","audio":audio_path_for_model}, {"type":"text", "text":sample["transcription"]+system_content} ]} ] else: conv_audio = [ {"role":"system", "content":[{"type":"text","text":system_prompt}]}, {"role":"user", "content":[{"type":"text","text":system_content},{"type":"audio","audio":audio_path_for_model}]} ] text_prompt = processor.apply_chat_template( conv_audio, add_generation_prompt=True, tokenize=False ) audios, _, _ = process_mm_info( conv_audio, use_audio_in_video=False ) inputs = processor( text=text_prompt, audio=audios, return_tensors="pt" ).to(model.device) text_ids = model.generate( **inputs, return_audio=False, do_sample=False, use_audio_in_video=False ) pred = processor.batch_decode( text_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False )[0].split("assistant\n")[-1].strip() score = validate_prediction(pred, gt_text, bm) scores_by_bm_modality[(bm, "audio")].append(score) # **新增**：预测错误则写入 CSV if score < 1.0: err_writer.writerow([ sample_id, bm, "audio", gt_text, pred ]) except Exception as e: print(f"{e} in index {sample_id}") # 可以视需要，也把异常记录到 err_writer continue # —— Text 模态 —— question = sample.get("question", "") transcription = sample.get("transcription", "") if question: try: conv_text = [ {"role":"system", "content":[{"type":"text","text":system_prompt}]}, {"role":"user", "content":[{"type":"text","text":question+system_content}]} ] text_prompt = processor.apply_chat_template( conv_text, add_generation_prompt=True, tokenize=False ) inputs = processor( text=text_prompt, return_tensors="pt" ).to(model.device) text_ids = model.generate( **inputs, return_audio=False, do_sample=False, use_audio_in_video=False ) pred = processor.batch_decode( text_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False )[0].split("assistant\n")[-1].strip() score = validate_prediction(pred, gt_text, bm) scores_by_bm_modality[(bm, "text")].append(score) # **新增**：预测错误则写入 CSV if score < 1.0: err_writer.writerow([ sample_id, bm, "text", gt_text, pred ]) except Exception as e: print(f"{e} in index {sample_id}") continue # 关闭错误记录文件，释放模型和显存 err_f.close() del model torch.cuda.empty_cache() # 计算总体 & 各分组准确率 all_scores = [ s for scores in scores_by_bm_modality.values() for s in scores ] overall = sum(all_scores) / len(all_scores) if all_scores else 0.0 accuracies = { f"{bm}_{modality}": sum(scores) / len(scores) for (bm, modality), scores in scores_by_bm_modality.items() } print(f" → 保存错误示例到 {err_path}") return overall, accuracies # ---- 主函数 ---- def main(): # 1) 加载数据集 dataset_path = "/home/chenyifu/new_rl/dataset/validation" if not os.path.exists(dataset_path): raise FileNotFoundError(f"{dataset_path} not found") eval_dataset = datasets.load_from_disk(dataset_path) print(f"Loaded dataset with {len(eval_dataset)} samples.") # 2) 收集所有 checkpoint 并排序 ckpt_glob = "/home/chenyifu/new_rl/dissrc/qwenomnithinker_rl/exp/gemini_model_s0.8a0.2/checkpoint-*" ckpts = sorted(glob.glob(ckpt_glob), key=lambda x: int(x.rsplit("-",1)[1])) print(f"Found {len(ckpts)} checkpoints under {ckpt_glob}") # # 3) 逐 checkpoint 评估并打印结果 for ckpt in tqdm(ckpts, desc="Checkpoints"): overall, accs = evaluate_checkpoint(ckpt, eval_dataset) name = os.path.basename(ckpt) print(f"\n=== Results for {name} ===") print(f"总体平均准确率: {overall:.4f}") print("-" * 50) for key in sorted(accs): print(f"{key}_accuracy: {accs[key]:.4f}") if __name__ == "__main__": main()