FunASR icon indicating copy to clipboard operation
FunASR copied to clipboard

微调后模型speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch后,推理无输出

Open linshoubiao opened this issue 1 year ago • 3 comments

微调模型后,加载./checkout微调输出目录下模型,无推理结果,用训练集测试也无结果。code如下:

infer_params = {} infer_params["model_dir"] = "/workspace/asr/FunASR/egs_modelscope/asr_vad_punc/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch/checkpoint" infer_params["decoding_model_name"] = "10epoch.pb" infer_params["test_data_dir"] = "/workspace/asr/FunASR/egs_modelscope/asr_vad_punc/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch/data/validation" modelscope_infer(infer_params)

linshoubiao avatar Nov 06 '23 09:11 linshoubiao

OS: e.g. linux Python/C++ Version:3.8.18 Package Version:pytorch、torchaudio、modelscope、funasr version(pip list) torch 2.0.1 torchaudio 2.0.2 modelscope 1.9.4 funasr 0.8.2 Model:speech_paraformer-large-vad-punc_asr Command: Details: 其中 --model文件目录是原始的预训练模型路径,其中的model.pb 是微调后19epoch.pb靠过来改成model.pb的 def modelscope_infer(args): os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpuid) inference_pipeline = pipeline( task=Tasks.auto_speech_recognition, model=args.model, output_dir=args.output_dir, batch_size=args.batch_size, update_model=False, param_dict={"decoding_model": args.decoding_mode, "hotword": args.hotword_txt} ) rlst = inference_pipeline(audio_in=args.audio_in) print("rlst") print(rlst)

if name == "main": parser = argparse.ArgumentParser() parser.add_argument('--model', type=str, default="/workspace/asr/FunASR/egs_modelscope/asr_vad_punc/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch/checkpoint") parser.add_argument('--audio_in', type=str, default="/workspace/data/asr-data/asr_data_1000-1101/asr_data_1000-1101/74_2021111203.wav_12.wav") parser.add_argument('--output_dir', type=str, default="./results/") parser.add_argument('--decoding_mode', type=str, default="normal") parser.add_argument('--hotword_txt', type=str, default=None) parser.add_argument('--batch_size', type=int, default=4) parser.add_argument('--gpuid', type=str, default="1") args = parser.parse_args() modelscope_infer(args)

Error log:

linshoubiao avatar Nov 06 '23 10:11 linshoubiao

原始得预训练模型是在huggingface上hezhaoqia / speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch下得。 微调过程中loss已经很低了,使用部分训练集合进行测试,没有输出,results = speech2text(**batch)中间结果全为 funasr代码版本为dev_bat_modelscope

linshoubiao avatar Nov 06 '23 13:11 linshoubiao

遇到相似的问题,请问解决了吗?

h310558606 avatar Nov 22 '23 03:11 h310558606