-
Notifications
You must be signed in to change notification settings - Fork 1.3k
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
Can't add new language to pre-trained spoken language recognition model: Model forgets other languages #1516
Comments
Hi, |
Thank you for answer! However, I did not shuffle my training set and the data was organised such way that I fit to model all samples of one certain language and only after that the next language was fit. Could shuffling training set help? |
That could play an important role. I think it is important to make sure there are data from different languages in each batch. |
I ensured that every batch contains all languages I want to detect and fine-tuned model, it worked pretty fine. Thanks a lot! |
Looks like this is solved; closing this one—please feel free to reopen :) |
Hello,
I am trying to fine-tune existing spoken language recognition model. I chose common voice language and trying to add new language. I did things exactly as they are described in fine-tuning tutorial (and ensured unknown label in label encoder as well).
I also tried to freeze more layers, for example, I froze every modules except classifier. However, when I fine-tune the model, the performance gets worse. For example, during the first several epochs model gives different incorrect outputs. However, around 5th epochs it starts assigning every language the label I want to add.
I also tried to fine-tune model on 19 different languages (including previously unknown), however, the results still same. Is there any way to fine-tune model to predict new languages or this model is not supposed to be fine-tuned? Why model can't learn new languages and forgets old during fine-tuning?
Here is the class I used in fine-tuning
The text was updated successfully, but these errors were encountered: