search

mindspore-lab / mindaudio

A toolbox of audio models and algorithms based on MindSpore
Apache License 2.0
36 stars 10 forks source link

[LibriSpeech-deepspeech2] [Ascend910] [GRAPH] Unable to replicate the precision of the weight file #106

Closed 787918582 closed 1 year ago

787918582 commented 1 year ago

deepspeech2在ascend上无法复现权重文件精度 版本信息：mindspore:1.9.0 mindaudio:2023.03.16

  1. readme中的精度标准 deepspeech2_5
  2. 实测精度 deepspeech2_6
787918582 commented 1 year ago

回归时间：2023.03.29 回归版本：mindspore-1.9.0 mindaudio-20230328 回归步骤：python eval.py --config ./hparams/DeepSpeech2.yaml --device_target GPU 回归现象：Test Summary Average WER 10.244 Average CER 3.461 回归结论：回归通过