Wav2Vec2 upgrade with Conv1D options #1758
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@minhthuc2502 could you please have a look at this PR and merge it? |
added 9 commits
August 14, 2024 11:01
|
@minhthuc2502 Thank you for your suggestions. I agree that the changes were redundant. I've updated the code as you recommended and made the necessary commits to ensure all checks pass. It appears that the network conditions for build-and-push-docker-images are more favorable in the morning where I am located. |
|
Hello, Thank you for your updates. I'll merge this. I agree that there is some network problems with build-and-push-docker-images , currently, I have rerun it manually. Have to fix this in the future. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR enhances the efficiency of Wav2Vec2 inference within the CTranslate2 framework, specifically improving both speed and memory usage. Compared to the HuggingFace implementation, the int8 quantized model shows an 11% increase in speed and a 70% reduction in memory usage on GPU, as well as a 5% increase in speed and a 71% reduction in memory usage on CPU when processing 300 audio files. Additionally, using an N-gram language model with pyctcdecode further can improve the speech recognition accuracy. My environment includes an NVIDIA GeForce RTX 3090 24GB with CUDA 12.4, torch==2.12+cu12.1, and transformers==4.41.0. Special thanks for the Depthwise convolution process introduced in #1749.