New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Pytorch Lightning Integration #569
Conversation
Next steps are to integrate saving/loading of the model from training to eval. |
fe6583e
to
b6f6572
Compare
This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions. |
…ity, integrated checkpoint support, fixed validation support
Added trains viz logging Precision
…x for adam Trains support, removed autocast since this is handled via lightning
…orrect lightning version
… the epochs used for previous release
…ew ModelCheckpoint class for model saving
…ath, refactor path name and test
…bject for manifest to modify root path
0ba0fb7
to
51fc8fb
Compare
deepspeech_pytorch/model.py
Outdated
) | ||
self.fc = nn.Sequential( | ||
SequenceWise(fully_connected), | ||
) | ||
self.inference_softmax = InferenceBatchSoftmax() | ||
self.criterion = CTCLoss(reduction='sum', zero_infinity=True) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'm currently using reduction='mean'
and zero_infinity=False
(the defaults)
What's the reason for these parameters?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Reduction sum means do not average the loss/gradients. This is reflecting the behaviour of warp-ctc. Zero infinity means if we get an inf gradient, we zero instead of returning infinite. Another behaviour copying warp-ctc!
Co-authored-by: Anas Abou Allaban <aabouallaban@protonmail.com>
I've run into some strange convergence issues with librispeech on this branch. Going to investigate and have a look (might just be the batch size needs to be small). |
FYI, I had to install Otherwise, I would get an error when the model tried to parse the TypeError: issubclass() arg 1 must be a class where the type of My assumption is that this is the relevant issue. Maybe this is also an upstream issue w/ PTL? |
Specify blank index explicitly
…-model Add blank index to ctc loss explicitly
Fix Syntax Warning
* Use torchaudio * Add torchaudio to reqs
Builds will fail due to lack of pre-trained models, but that should be solved once i make the release! I think we're there |
I solved this by reinstalling omegaconf to this version |
Closes #568