have a ready-made onnx file?

[zmoth]

有现成的onnx文件吗？类似于wespeaker在huggingface上的。我使用魔搭社区上面的pt模型一直转换失败.

Traceback (most recent call last):
  File "wekws/bin/export_onnx.py", line 103, in <module>
    main()
  File "wekws/bin/export_onnx.py", line 51, in main
    load_checkpoint(model, args.checkpoint)
  File "/Users/moth/proj/voice/wekws/wekws/utils/checkpoint.py", line 30, in load_checkpoint
    checkpoint = torch.load(path, map_location='cpu')
  File "/Users/moth/proj/voice/wekws/.conda/lib/python3.8/site-packages/torch/serialization.py", line 1040, in load
    return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)
  File "/Users/moth/proj/voice/wekws/.conda/lib/python3.8/site-packages/torch/serialization.py", line 1262, in _legacy_load
    magic_number = pickle_module.load(f, **pickle_load_args)
_pickle.UnpicklingError: invalid load key, 'v'.

[mlxu995]

@zmoth 方便的话发一下你转 onnx 失败的 pt 的地址

[zmoth]

在WIndows WSL和macOS上都是这个结果，使用miniconda，python=3.8.19 ，macOS上pytorch=2.3.0，torchaudio=2.3.0，WSL是按照README环境安装的。

https://www.modelscope.cn/studios/thuduj12/KWS_Nihao_Xiaojing

使用的是run_fsmn_ctc.sh

./run_fsmn_ctc.sh 4 4

checkpoint=
dir=exp/KWS_Nihao_Xiaojing/model/hixiaowen
average_model=true
num_average=30
if $average_model ;then
  score_checkpoint=$dir/avg_${num_average}.pt
else
  score_checkpoint=$dir/final.pt
fi

download_dir=./data/local # your data dir

其实上面的模型我都尝试过，有些是bin文件不知道怎么转换，还有的(https://www.modelscope.cn/iic/speech_charctc_kws_phone-xiaoyun)报错是

Traceback (most recent call last):
  File "wekws/bin/export_onnx.py", line 103, in <module>
    main()
  File "wekws/bin/export_onnx.py", line 42, in main
    feature_dim = configs['model']['input_dim']
KeyError: 'model'

[mlxu995]

ctc loss 的模型暂时还没有支持导出 onnx 和 c++ 推理，你可以先试试 maxpooling loss 的模型