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
set default none decoder for ASR #4917
Conversation
I added @pyf98 for this PR. |
@pyf98, would it solve all issues? |
Actually, there are a lot of CI errors, @ftshijt |
Yeah, sorry that I did not find time to fix the errors. Will try to fix those later |
It can solve the issue when we run new experiments using prior CTC configs. However, for previously uploaded models, the issue still exists? For example, the decoder was set to If we download and load it, the decoder will be set to RNN? |
Oh, this is complicated. |
I think it does not affect the final result. It just causes some redundant parameters in the model, which is not good. Also, it will cause an error for |
I think we can change the default decoder to But to keep compatibility, we still need to set the decoder as None in the model code when ctc weight is 1.0. |
I see. Good idea. |
Sure, I agree with Yifan on the potential behaviors. I think the only point here is the initialization with strict loading checks. Do you guys know some of the available pre-trained models with only CTC? |
Codecov Report
@@ Coverage Diff @@
## master #4917 +/- ##
==========================================
- Coverage 76.56% 76.56% -0.01%
==========================================
Files 603 603
Lines 53753 53756 +3
==========================================
Hits 41157 41157
- Misses 12596 12599 +3
Flags with carried forward coverage won't be shown. Click here to find out more.
📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more |
@ftshijt, I just want to make sure whether it is working with existing pre-trained CTC models without changing the configurations. |
I'm also curious. I don't see the related modifications in the code. How is it solved? Am I missing anything? |
Sorry, my bad. I did not push that to remote. The current one should prevent issues when loading previous checkpoints that are trained with ctc only. |
(I checked with the |
This PR is ready to merge |
LGTM. |
looks good! |
Thanks, @ftshijt! |
Tries to fix the issues in #4908