Resetting Learning Rate for train_from


I read in previous posts that in OpenNMT-py resetting or changing the learning rate with train_from was not possible. Is this still the case, or is there a way to do this?


Hey Yasmin,
You have to use the -reset_optim flag if you want to change the learning rate. Else, you can modify the checkpoint “by hand”, loading it with torch.load and manually setting the values you need to change.

Thanks, François! I will try this.