sh/qwen_test.py

import os
import json
import soundfile as sf
from transformers import Qwen2_5OmniForConditionalGeneration, Qwen2_5OmniProcessor
from qwen_omni_utils import process_mm_info

# 加载模型和处理器
model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
    "/data/liyangzhuo/models/Qwen2.5-Omni-7B", 
    torch_dtype="auto", 
    device_map="auto"
)
processor = Qwen2_5OmniProcessor.from_pretrained(
    "/data/liyangzhuo/models/Qwen2.5-Omni-7B"
)

# 定义输入输出目录
input_audio_dir = "/data/liyangzhuo/wavbench_volume/speed"
output_dir = "/data/liyangzhuo/wavbench_volume/result"
json_file_path = "/data/liyangzhuo/wavbench_volume/speed/explicit_understanding_speed.json"

# 设置是否在视频中使用音频
USE_AUDIO_IN_VIDEO = False

# 设置temperature参数列表（用于生成多样性）
TEMPERATURES = [0.8, 0.8, 0.8, 0.8]

# 确保输出目录存在
os.makedirs(output_dir, exist_ok=True)


# 读取JSON文件，获取所有音频文件信息
with open(json_file_path, 'r', encoding='utf-8') as f:
    volume_data = json.load(f)

# 创建id到数据的映射
data_map = {item['id']: item for item in volume_data}

# 获取输入目录中的所有wav文件
input_files = []
for filename in os.listdir(input_audio_dir):
    if filename.endswith('.wav'):
        input_files.append(os.path.join(input_audio_dir, filename))

# 按文件名排序
input_files.sort()

# 准备存储结果的列表
results = []

print("开始处理Qwen音频对话任务...")
print(f"输入目录: {input_audio_dir}")
print(f"输入文件数量: {len(input_files)}")
print(f"输出目录: {output_dir}")
print(f"JSON文件: {json_file_path}")
print("=" * 70)

success_count = 0
error_count = 0

# 依次处理每个音频文件
for i, input_file in enumerate(input_files, 1):
    try:
        basename = os.path.basename(input_file)
        file_id = os.path.splitext(basename)[0]  # 获取文件名（不含扩展名）
        
        print(f"\n🔄 处理第 {i}/{len(input_files)} 个文件: {basename}")
        
        # 检查输入文件是否存在
        if not os.path.exists(input_file):
            print(f"   ❌ 错误: 文件不存在 {input_file}")
            error_count += 1
            continue
        
        # 检查音频文件（仅为了验证文件有效性）
        original_audio, original_sr = sf.read(input_file)
        audio_duration = len(original_audio)/original_sr
        print(f"   📁 音频信息: 采样率 {original_sr}Hz, 时长 {audio_duration:.2f}s")
        
        # 获取对应的问题文本和标签
        question_text = ""
        label = ""
        if file_id in data_map:
            question_text = data_map[file_id].get('text', '')
            label = data_map[file_id].get('label', '')
            print(f"   📝 问题: {question_text[:80]}...")
            print(f"   🏷️  标签: {label}")
        else:
            print(f"   ⚠️  警告: 在JSON文件中未找到 {file_id} 的信息")
        
        # 为当前音频文件创建对话
        conversation = [
            {
                "role": "system",
                "content": [
                    {"type": "text", "text": "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."}
                ],
            },
            {
                "role": "user",
                "content": [
                    {"type": "audio", "audio": input_file},
                ],
            },
        ]
        
        # 准备模型输入
        text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
        audios, images, videos = process_mm_info(conversation, use_audio_in_video=USE_AUDIO_IN_VIDEO)
        inputs = processor(text=text, audio=audios, images=images, videos=videos, return_tensors="pt", padding=True, use_audio_in_video=USE_AUDIO_IN_VIDEO)
        inputs = inputs.to(model.device).to(model.dtype)
        
        # 存储多次生成的回复
        responses = []
        
        print("   🤖 正在生成4次回复（不同temperature）...")
        for temp_idx, temperature in enumerate(TEMPERATURES, 1):
            print(f"   📝 生成第 {temp_idx}/4 次 (temperature={temperature})...")
            
            # 生成文本输出（禁用音频生成）
            text_ids = model.generate(
                **inputs, 
                use_audio_in_video=USE_AUDIO_IN_VIDEO,
                return_audio=False,
                temperature=temperature,
                do_sample=True,  # 启用采样以使temperature生效
                max_new_tokens=512  # 限制生成长度
            )
            
            full_text = processor.batch_decode(
                text_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
            )
            
            # 提取模型回复文本
            generated_text = full_text[0]
            if "assistant" in generated_text:
                model_response = generated_text.split("assistant", 1)[1].strip()
            else:
                model_response = generated_text.strip()
            
            responses.append({
                "temperature": temperature,
                "response": model_response
            })
            
            print(f"      回复预览: {model_response[:80]}...")
        
        print(f"   ✅ 完成4次文本生成")
        
        
        # 保存结果到列表
        result_entry = {
            "id": file_id,
            "input_file": basename,
            "question": question_text,
            "label": label,
            "audio_duration_seconds": round(audio_duration, 2),
            "responses": responses  # 包含4次不同temperature的回复
        }
        results.append(result_entry)
        
        success_count += 1
        
    except Exception as e:
        print(f"   ❌ 处理失败: {e}")
        error_count += 1
        continue

# 保存所有结果到JSON文件
result_json_path = os.path.join(output_dir, "speed_results.json")
with open(result_json_path, 'w', encoding='utf-8') as f:
    json.dump(results, f, ensure_ascii=False, indent=2)

print("\n" + "=" * 70)
print("📊 处理完成统计:")
print(f"   ✅ 成功处理: {success_count} 个文件")
print(f"   ❌ 处理失败: {error_count} 个文件")
print(f"   📁 输出位置: {output_dir}")
if success_count > 0:
    print(f"   📋 结果JSON: {result_json_path}")
print("=" * 70)
print("🎉 任务完成！")