sh/qwen_temperature.py

import os
import random
import numpy as np
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"
)

# 定义基础目录
base_input_dir = "/data/liyangzhuo/test/question"
base_output_dir = "/data/liyangzhuo/test/temperature"

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

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

# 获取所有音频文件
audio_files = []
for filename in os.listdir(base_input_dir):
    if filename.endswith('.wav'):
        audio_files.append(os.path.join(base_input_dir, filename))

# 随机选择20条语音
selected_files = random.sample(audio_files, min(20, len(audio_files)))

# 设置temperature参数
temperature = 0.7

# 处理每个选中的音频文件
for idx, input_file in enumerate(selected_files, start=1):
    print(f"正在处理第 {idx} 个文件: {input_file}")
    
    # 读取输入音频
    input_audio, sr = sf.read(input_file)
    
    # 如果采样率不是24000，需要重采样
    if sr != 24000:
        import librosa
        input_audio = librosa.resample(input_audio, orig_sr=sr, target_sr=24000)
        sr = 24000
    
    # 创建2秒的空白音频（24000 Hz * 2秒）
    silence = np.zeros(24000 * 2)
    
    # 生成4次回复
    for reply_idx in range(1, 5):
        print(f"  生成第 {reply_idx} 次回复...")
        
        # 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)
        
        # 生成输出，添加temperature参数
        text_ids, audio = model.generate(
            **inputs, 
            use_audio_in_video=USE_AUDIO_IN_VIDEO,
            temperature=temperature,
            do_sample=True  # 开启采样以使temperature生效
        )
        full_text = processor.batch_decode(
            text_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
        )

        # 提取模型回复
        generated_text = full_text[0]
        if "assistant" in generated_text:
            model_prediction = generated_text.split("assistant", 1)[1].strip()
        else:
            model_prediction = generated_text.strip()
        print(f"    回复文本: {model_prediction}")
        
        # 获取生成的音频
        generated_audio = audio.reshape(-1).detach().cpu().numpy()
        
        # 拼接音频：输入 + 2秒空白 + 回复
        combined_audio = np.concatenate([input_audio, silence, generated_audio])
        
        # 构造输出文件名，格式为 qwen_0001_1.wav
        output_filename = f"qwen_{idx:04d}_{reply_idx}.wav"
        output_path = os.path.join(base_output_dir, output_filename)
        
        # 保存拼接后的音频
        sf.write(
            output_path,
            combined_audio,
            samplerate=24000,
        )
        
        print(f"    已保存至 {output_path}")
    
    print(f"完成第 {idx} 个文件的处理\n")

print("所有文件处理完成！")