Conversation
fix some bugs about raw waveform codes add make_raw_feats script add ApplyMinMaxToWeights fix the scripts to use max/min_param_value On/Off batchnorm in fft component add a nnet3/get_egs_old.sh backup small fix small fix2
add _separate version fvector code
| --initial-effective-lrate 0.002 \ | ||
| --final-effective-lrate 0.0002 \ | ||
| --max-param-change 0.2 \ | ||
| --minibatch-size 16 \ |
There was a problem hiding this comment.
isn't it too small? Are you sure the training objective converge?
There was a problem hiding this comment.
I will increase it. For now, when I use 40 epochs data, the log-likelihood on valid is -0.29. I will increase it to 64 and 128 firstly.
| --max-snr 20 \ | ||
| --min-snr 10 \ | ||
| --num-epochs 8 --use-gpu $use_gpu --stage $train_stage \ | ||
| --num-jobs-initial 1 --num-jobs-final 3 \ |
There was a problem hiding this comment.
Did you try to increase num-jobs-final?
There was a problem hiding this comment.
I am not sure, what is the difference among your training scripts? Also _separate is not a good name for script, use better informative suffix.
There was a problem hiding this comment.
The num-jobs-final is limited by the number of egs. As the minilibrispeech is small, it only has 4 egs. For now, I use "num-jobs-final=4"
I will think about a better suffix name.
| done | ||
| fi | ||
|
|
||
|
|
2 participants
You can go through along the egs/mini_librispeech/s5/run_fvector_separate.sh for newest version.