[WIP] Added Adapters to Whisper Model from openai #572
Conversation
|
Hey, thanks for your efforts in contributing new model architectures to adapter-transformers and sorry for the silence on our side. In the last few weeks, we've been working on a large refactoring of our project, which will ultimately result in the release of Adapters, the next-generation adapters library. See #584. As a consequence, we plan to merge any new model integrations directly to the new codebase, which currently can be found on this branch. Unfortunately, this necessitates some changes in the model integration code (detailed here, see already integrated models such as BERT, BART etc. for reference). If you'd be willing to update your model integration to target the new library yourself, we'd be super happy to help you on this. Otherwise, we might look into upgrading and merging some of the open model integration PRs ourselves in the future. For more details, again see #584. |
|
Hi, thanks again for your efforts in contributing Whisper as a new model to our adapters library. |
This PR adds adapter support for the Whisper model from openai and builds upon work done previously in #572. Key Additions: 1. Adapter Support for Whisper Model: - Incorporates adapter functionality to enhance the flexibility and adaptability of the Whisper model. 2. Enhanced Head Functions: - Expanded the argument options for some heads by adding a layer argument with a default value. 3. Preprocessing Scripts for Audio Datasets: - Added preprocessing scripts tailored for audio datasets. - These scripts are now utilized in the Whisper tests within the test suite, replacing the use of arbitrary samples. --------- Co-authored-by: Leon Engländer <leon.englaender@gmail.com> Co-authored-by: calpt <calpt@mail.de>
This PR adds adapter support for the Whisper model from openai.
This feature has been requested in an issue before (#466).