start training from a checkpoint

[ghost]

i want train.py to start from the checkpoint provided in pretrained model. but when i start the training , it doeasnt take the checkpoint of the pretrained model. It starts from the first checkpoint. p.s. the check point of pre trained model and the new training data are both in logs-tacotron folder

[Usama0121]

You need to use the --restore_step argument to start training from checkpoint

Try this command

python train.py --restore_step 441000

Where 441000 is the checkpoint number You might need to rename the checkpoint files as model.ckpt-441000.index and similarly other files

[Usama0121]

@saharsyed I used Urdu dataset for that as well which required some changes in code and stuff

for training with English dataset make sure that the sampling rate of audios matches the pre-trained models sampling rate which I think was 22k or 16k probably, and the length of audios should be between 7 to 15 seconds these work better, moreover, a female dataset might produce better results

[Usama0121]

@saharsyed For using a non-english dataset with the pre-trained model you need to add a transliteration layer that maps your language alphabets to the English alphabets. You can try using the transliteration_cleaners. PS having a phonetically rich corpus witch high-quality audios will produce better results

[ghost]

ok. Thanks a lot for your feedback