如何将语音预训练库的.pth文件转为C#可读的.pt文件？

需要将预训练语音模型的.pth文件转为C#可读的.pt格式，但仅找到torchvision视觉模型的转换资料，PyTorch.audio相关文档较少。已知.pth和.pt本质相同，但Python加载.pth时存在兼容问题，参考torchvision示例写了代码，运行时出现RuntimeError。

代码示例 #

pth_file_path = 'FILE/PATH/model.pth' 
pt_file_path = 'FILE/PATH/model.pt'

import torch
import torchaudio
from unet import UNet
from zmq import device 
model = UNet(3, 2)

model.load_state_dict(torch.load(pth_file_path, map_location=torch.device('cpu')))

错误信息 #

model.load_state_dict(torch.load(pth_file_path, map_location=torch.device('cpu')))
Traceback (most recent call last):
  File ".\IMPORT.py", line 11, in <module>
    model.load_state_dict(torch.load(pth_file_path, map_location=torch.device('cpu')))
  File "C:\Users\user\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\nn\modules\module.py", line 2215, in load_state_dict
    raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for UNet:
        Missing key(s) in state_dict: "encoder.encoding_blocks.0.conv1.conv_layer.weight", "encoder.encoding_blocks.0.conv1.conv_layer.bias", "encoder.encoding_blocks.0.conv1.block.0.weight", "encoder.encoding_blocks.0.conv1.block.0.bias", "encoder.encoding_blocks.0.conv2.conv_layer.weight", "encoder.encoding_blocks.0.conv2.conv_layer.bias", "encoder.encoding_blocks.0.conv2.block.0.weight", "encoder.encoding_blocks.0.conv2.block.0.bias", "encoder.encoding_blocks.1.conv1.conv_layer.weight", "encoder.encoding_blocks.1.conv1.conv_layer.bias", "encoder.encoding_blocks.1.conv1.block.0.weight", "encoder.encoding_blocks.1.conv1.block.0.bias", "encoder.encoding_blocks.1.conv2.conv_layer.weight", "encoder.encoding_blocks.1.conv2.conv_layer.bias", "encoder.encoding_blocks.1.conv2.block.0.weight", "encoder.encoding_blocks.1.conv2.block.0.bias", "encoder.encoding_blocks.2.conv1.conv_layer.weight", "encoder.encoding_blocks.2.conv1.conv_layer.bias", "encoder.encoding_blocks.2.conv1.block.0.weight", "encoder.encoding_blocks.2.conv1.block.0.bias", "encoder.encoding_blocks.2.conv2.conv_layer.weight", "encoder.encoding_blocks.2.conv2.conv_layer.bias", "encoder.encoding_blocks.2.conv2.block.0.weight", "encoder.encoding_blocks.2.conv2.block.0.bias", "encoder.encoding_blocks.3.conv1.conv_layer.weight", "encoder.encoding_blocks.3.conv1.conv_layer.bias", "encoder.encoding_blocks.3.conv1.block.0.weight", "encoder.encoding_blocks.3.conv1.block.0.bias", "encoder.encoding_blocks.3.conv2.conv_layer.weight", "encoder.encoding_blocks.3.conv2.conv_layer.bias", "encoder.encoding_blocks.3.conv2.block.0.weight", "encoder.encoding_blocks.3.conv2.block.0.bias", "bottom_block.conv1.conv_layer.weight", "bottom_block.conv1.conv_layer.bias", "bottom_block.conv1.block.0.weight", "bottom_block.conv1.block.0.bias", "bottom_block.conv2.conv_layer.weight", "bottom_block.conv2.conv_layer.bias", "bottom_block.conv2.block.0.weight", "bottom_block.conv2.block.0.bias", "decoder.decoding_blocks.0.upsample.weight", "decoder.decoding_blocks.0.upsample.bias", "decoder.decoding_blocks.0.conv1.conv_layer.weight", "decoder.decoding_blocks.0.conv1.conv_layer.bias", "decoder.decoding_blocks.0.conv1.block.0.weight", "decoder.decoding_blocks.0.conv1.block.0.bias", "decoder.decoding_blocks.0.conv2.conv_layer.weight", "decoder.decoding_blocks.0.conv2.conv_layer.bias", "decoder.decoding_blocks.0.conv2.block.0.weight", "decoder.decoding_blocks.0.conv2.block.0.bias", "decoder.decoding_blocks.1.upsample.weight", "decoder.decoding_blocks.1.upsample.bias", "decoder.decoding_blocks.1.conv1.conv_layer.weight", "decoder.decoding_blocks.1.conv1.conv_layer.bias", "decoder.decoding_blocks.1.conv1.block.0.weight", "decoder.decoding_blocks.1.conv1.block.0.bias", "decoder.decoding_blocks.1.conv2.conv_layer.weight", "decoder.decoding_blocks.1.conv2.conv_layer.bias", "decoder.decoding_blocks.1.conv2.block.0.weight", "decoder.decoding_blocks.1.conv2.block.0.bias", "decoder.decoding_blocks.2.upsample.weight", "decoder.decoding_blocks.2.upsample.bias", "decoder.decoding_blocks.2.conv1.conv_layer.weight", "decoder.decoding_blocks.2.conv1.conv_layer.bias", "decoder.decoding_blocks.2.conv1.block.0.weight", "decoder.decoding_blocks.2.conv1.block.0.bias", "decoder.decoding_blocks.2.conv2.conv_layer.weight", "decoder.decoding_blocks.2.conv2.conv_layer.bias", "decoder.decoding_blocks.2.conv2.block.0.weight", "decoder.decoding_blocks.2.conv2.block.0.bias", "decoder.decoding_blocks.3.upsample.weight", "decoder.decoding_blocks.3.upsample.bias", "decoder.decoding_blocks.3.conv1.conv_layer.weight", "decoder.decoding_blocks.3.conv1.conv_layer.bias", "decoder.decoding_blocks.3.conv1.block.0.weight", "decoder.decoding_blocks.3.conv1.block.0.bias", "decoder.decoding_blocks.3.conv2.conv_layer.weight", "decoder.decoding_blocks.3.conv2.conv_layer.bias", "decoder.decoding_blocks.3.conv2.block.0.weight", "decoder.decoding_blocks.3.conv2.block.0.bias", "classifier.conv_layer.weight", "classifier.conv_layer.bias", "classifier.block.0.weight", "classifier.block.0.bias".
        Unexpected key(s) in state_dict: "weight", "config", "info", "sr", "f0", "version".

问题分析 #

从错误信息能看出，你加载的.pth文件内容和定义的UNet模型结构完全不匹配：

• 意外键weight、config、info等说明这个.pth不是标准的PyTorch模型state_dict，而是包含权重、配置信息的复合字典，甚至可能是语音库自定义的序列化格式。
• 缺少的键都是UNet模型的子模块参数，说明.pth里的权重结构和你的UNet完全对应不上。

解决步骤 #

1. 先排查.pth文件的结构 #

先不要急于加载到模型，先打印文件内容，确认里面的结构：

import torch
pth_data = torch.load('FILE/PATH/model.pth', map_location='cpu')
print(type(pth_data))
# 如果是字典，打印所有键
if isinstance(pth_data, dict):
    print(pth_data.keys())
    # 查看'weight'键对应的张量形状（如果存在）
    if 'weight' in pth_data:
        print(pth_data['weight'].shape)

这一步能帮你明确.pth里存的是单个权重张量、完整模型还是复合配置数据。

2. 匹配模型结构或提取正确权重 #

• 如果.pth里的weight是模型权重张量，需要确认该权重对应的模型结构和你的UNet是否一致。不一致的话，要么修改UNet结构匹配权重，要么找到该语音库的官方模型定义代码，用正确的模型类加载。
• 如果.pth是完整模型（用torch.save(model, path)保存的），直接加载整个模型即可：

  model = torch.load('FILE/PATH/model.pth', map_location='cpu')
  

• 如果是复合字典，比如权重存在pth_data['model']这类键下，提取对应部分再加载：

  model.load_state_dict(pth_data['model'])
  

3. 转换为C#可读的.pt格式 #

成功加载模型后，直接保存为.pt格式即可（.pt和.pth本质都是PyTorch序列化格式，C#的TorchSharp等库均可读取）：

# 保存模型state_dict
torch.save(model.state_dict(), 'FILE/PATH/model.pt')
# 或者保存完整模型
torch.save(model, 'FILE/PATH/model.pt')

注意事项 #

• 语音模型的.pth格式常带有自定义内容（如采样率、配置信息），不能直接套用视觉模型的加载逻辑。
• 优先参考该预训练语音库的官方文档或加载示例，确保模型结构和权重完全匹配。

内容的提问来源于stack exchange，提问作者Crombopul