Help linear_loss in new version about tacotron-pytorch HOT 3 CLOSED

Yes, I think you can modify it as follows

if self.step % self.config['solver']['checkpoint_interval'] == 0 and local_step != 0:
  self.save_ckpt()   
  # For monitoring the training process with the tensorboard
  with torch.no_grad():
    self.validate()
    

I will check that later. Thanks for posting this issue.

from tacotron-pytorch.

In my case, the last saved checkpoint is in step 138k.
I suggest you save the model every 10k steps and run validation for all previously saved models.
Do you change the hyperparameters in config.yaml?

from tacotron-pytorch.

No, i don't change config.yaml -> checkpoint_interval: 2000
but not save every 10K for solver.py:

if val_err < self.best_val_err:
# Save checkpoint
self.save_ckpt()
self.best_val_err = val_err

Should I comment on that condition?

Thanks!

from tacotron-pytorch.