Perch-Hoplite embedding database integration #1178
Open
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
this doesn't seem ideal, but we can't add "inf" values to the hoplite db, and don't want to rescale or normalize because the hoplite db doesn't contain any additional scaling metadata
untested! need to check implementation and see how it works
Collaborator
Author
|
Still need to implement training shallow classifiers on hoplite embedding db (now done) |
Collaborator
Author
|
this is a pretty big refactor aimed at making things like the BMZ Perch2/Birdnet/TF models and ONNXModel only need to implement batch_forward() and otherwise be able to use SpectrogramPreprocessor methods like .predict() and .embed() The strategy is that you implement batch_forward as appropraite for the class, and it returns a dictionary of outputs. Then predict() and embed() just need to make the dataloader with self.predict_dataloader(), iterate the dataloader to get batches, call batch_forward() on each batch, and aggregate the results across batches. Some things in the tutorials are likely broken after this refactor, I haven't run/tested/checked them.
the DataLoader needs to return lists of AudioSample anyways for other parts of the code to work, so we can specify using collate_audio_samples to get the sample.data and sample.labels of each sample after getting a batch of AudioSamples from a dataloader. This reduces complexity/confusion.
batch_forward is what we want to implement per-model anyways, so it is natural to define collation within batch_forward
instead, retain start time and end time listed in the original input, since this is what is loaded from the audio and we want to retain match between preprocessed/output start/end time and input (eg to know the processed sample matches the input) this resolves issues with existing samples not "matching" embedded samples and re-embedding
still getting surprising mismatches for floating point comparison of offsets, established rounding precision default of just 3 decimals
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.
No description provided.