Results 342 comments of François Chollet

This appears to be correct and normal: https://github.com/pytorch/audio/issues/1058 > I found librosa use slaney normalization for the mel-filterbank creation as the default, while torchaudio is no normalization by default. How...

Did you try to compare the output of `mel_weights_matrix` side by side between the two implementations? If it matches, then the implication is that there's an additional difference with Librosa....

> Also do we need tf_data compatibility? I tried adding tf_data_compatibility test. But it fails. It is feasible. The current layer code is not set up for that. Take a...

Ok -- it's definitely better to make it work with tf.data even if that's not the main use case, then. If you're running into major blockers, then we can revert...

Right -- if you want tf.data support with a backend other than TF, all ops applied by the layer must go through `self.backend`. Anything else won't work.

The `weights=` layer constructor argument was dropped in Keras 3. I've decided to add it back in the specific case of the `Embedding` layer, because unlike other layers it gets...

As of now, ONNX is a talking point used by large companies, that doesn't seem to address an existing problem and doesn't seem to have actual users. This may change...

Thanks for the report -- I just fixed it. Please check that the build at HEAD works for your use case.

> This branch has conflicts that must be resolved Please solve conflicts or exclude the file from the PR.