-
Notifications
You must be signed in to change notification settings - Fork 173
Closed
Labels
good first issueGood for newcomersGood for newcomers
Description
1、stage1训练框架中
configs['optimizer_args']['lr'] = configs['scheduler_args']['initial_lr']
scheduler: WarmupLR_withStepDecay中没有定义initial_lr,应该可以注释
2、WarmupLR_withStepDecay的定义中,self.warmup_step = warmup_step是不是应该修改为 self.warmup_step = warmup_step * epoch_iter,才能保证是前5个epoch学习率增加到1e-4
3、frontend_type = frontend时代码训练中是否需要apply cmvn
代码中默认采用apply cmvn,但是 with torch.cuda.amp.autocast(enabled=configs['enable_amp']):
features, _ = model.module.frontend(wavs, wavs_len), features的是元组的格式,是需要取最后一层输出用于apply cmvn,还是不采用apply cmvn。
Metadata
Metadata
Assignees
Labels
good first issueGood for newcomersGood for newcomers