wenet-e2e / wespeaker

Research and Production Oriented Speaker Verification, Recognition and Diarization Toolkit
Apache License 2.0
707 stars 116 forks source link

怎么没有 eval 过程? #130

Closed zuowanbushiwo closed 1 year ago

zuowanbushiwo commented 1 year ago

你好 发现训练过程没有验证集验证的过程?是这种任务不需要吗?能不能从训练集中每个speaker 取10%得数据,然后再自己生成一个trials, 这样每个epoch 后,计算这训练集 trials的eer? 谢谢!

JiJiJiang commented 1 year ago

@czy97 zy可以回答一下哈

czy97 commented 1 year ago

你好 发现训练过程没有验证集验证的过程?是这种任务不需要吗?能不能从训练集中每个speaker 取10%得数据,然后再自己生成一个trials, 这样每个epoch 后,计算这训练集 trials的eer? 谢谢!

是可以的,我们最近也在考虑最终这件事要怎么做,谢谢意见

zuowanbushiwo commented 1 year ago

好的,期待中!

yuGAN6 commented 1 year ago

我感觉是因为声纹trials多,eval时间比较长。如果加到训练过程串行进行,对训练时间影响较大。如果在训练完再测试trials,可以开多进程,出结果会快点。