Update model_loader.py #7
Merged
Conversation
This file contains hidden or 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
add wav2vec 2.0 [base, large]; hubert [base, large]; wavlm [base, baseplus, large]; whisper [base] to utilize fadtk for speech analysis
Contributor
Author
|
@microsoft-github-policy-service agree |
Collaborator
|
Thank you so much for adding support for more embedding models! I will review the code changes in a moment. |
fadtk/model_loader.py
Outdated
Comment on lines
467
to
535
| class W2V2baseModel(ModelLoader): | ||
| """ | ||
| W2V2base model from https://huggingface.co/facebook/wav2vec2-base-960h | ||
| Please specify the layer to use (1-12). | ||
| """ | ||
| def __init__(self, size='960h', layer=12, limit_minutes=6): | ||
| super().__init__(f"w2v2base" + ("" if layer == 12 else f"-{layer}"), 768, 16000) | ||
| self.huggingface_id = f"facebook/wav2vec2-base-{size}" | ||
| self.layer = layer | ||
| self.limit = limit_minutes * 60 * self.sr | ||
|
|
||
| def load_model(self): | ||
| from transformers import AutoProcessor | ||
| from transformers import Wav2Vec2Model | ||
|
|
||
| self.model = Wav2Vec2Model.from_pretrained(self.huggingface_id) | ||
| self.processor = AutoProcessor.from_pretrained(self.huggingface_id) | ||
| self.model.to(self.device) | ||
|
|
||
| def _get_embedding(self, audio: np.ndarray) -> np.ndarray: | ||
| # Limit to 9 minutes | ||
| if audio.shape[0] > self.limit: | ||
| log.warning(f"Audio is too long ({audio.shape[0] / self.sr / 60:.2f} minutes > {self.limit / self.sr / 60:.2f} minutes). Truncating.") | ||
| audio = audio[:self.limit] | ||
|
|
||
| inputs = self.processor(audio, sampling_rate=self.sr, return_tensors="pt").to(self.device) | ||
| with torch.no_grad(): | ||
| out = self.model(**inputs, output_hidden_states=True) | ||
| out = torch.stack(out.hidden_states).squeeze() # [13 layers, timeframes, 768] | ||
| out = out[self.layer] # [timeframes, 768] | ||
|
|
||
| return out | ||
|
|
||
|
|
||
| class W2V2largeModel(ModelLoader): | ||
| """ | ||
| W2V2large model from https://huggingface.co/facebook/wav2vec2-large-960h | ||
| Please specify the layer to use (1-24). | ||
| """ | ||
| def __init__(self, size='960h', layer=24, limit_minutes=6): | ||
| super().__init__(f"w2v2large" + ("" if layer == 24 else f"-{layer}"), 1024, 16000) | ||
| self.huggingface_id = f"facebook/wav2vec2-large-{size}" | ||
| self.layer = layer | ||
| self.limit = limit_minutes * 60 * self.sr | ||
|
|
||
| def load_model(self): | ||
| from transformers import AutoProcessor | ||
| from transformers import Wav2Vec2Model | ||
|
|
||
| self.model = Wav2Vec2Model.from_pretrained(self.huggingface_id) | ||
| self.processor = AutoProcessor.from_pretrained(self.huggingface_id) | ||
| self.model.to(self.device) | ||
|
|
||
| def _get_embedding(self, audio: np.ndarray) -> np.ndarray: | ||
| # Limit to 9 minutes | ||
| if audio.shape[0] > self.limit: | ||
| log.warning(f"Audio is too long ({audio.shape[0] / self.sr / 60:.2f} minutes > {self.limit / self.sr / 60:.2f} minutes). Truncating.") | ||
| audio = audio[:self.limit] | ||
|
|
||
| inputs = self.processor(audio, sampling_rate=self.sr, return_tensors="pt").to(self.device) | ||
| with torch.no_grad(): | ||
| out = self.model(**inputs, output_hidden_states=True) | ||
| out = torch.stack(out.hidden_states).squeeze() # [25 layers, timeframes, 1024] | ||
| out = out[self.layer] # [timeframes, 1024] | ||
|
|
||
| return out | ||
|
|
Collaborator
There was a problem hiding this comment.
I see that the code for W2V2 base and W2V2 large are mostly identical. Would it be better to reuse some duplicate parts of the code by using abstractions? (e.g. defining and extending from a base class for each model family containing the duplicated functions)
This also applies to the base and large variants of HuBERT, and WavLM
Contributor
Author
There was a problem hiding this comment.
Updated!
I also integrated more model types of whisper. Now, the most widely used speech models, i.e., w2v2, hubert, wavlm, and whisper have been included, which should be sufficient for the majority of speech tasks.
reusing duplicate parts of w2v2, hubert, and wavlm integrating mode types of whisper These four models are the most widely used ones in the speech area.
hykilpikonna
approved these changes
Nov 20, 2023
Collaborator
|
Thanks! I just merged the pull request. By the way, are there any extra dependencies required for these models? From what I see in the code changes, the only conditional import you used is |
Contributor
Author
|
Yep. No other dependencies required. Thanks for merging! |
christhetree
pushed a commit
to christhetree/fadtk
that referenced
this pull request
Mar 24, 2025
add citation
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
add
wav2vec 2.0 [base, large];
hubert [base, large];
wavlm [base, baseplus, large];
whisper [base]
to utilize fadtk for speech analysis