Add Keras 3 example for "Audio track separation" #2003
Conversation
| import numpy as np | ||
| import soundfile as sf | ||
| from IPython import display | ||
| from keras import callbacks, layers, ops, optimizers, saving, utils |
There was a problem hiding this comment.
Only keep layers, ops, callbacks -- they are used many times. However optimizers, saving, utils are only used 1-2x so you can just do e.g. keras.optimizers.Adam.
There was a problem hiding this comment.
I ended up needing many more uses of saving in refactor, so I kept that import, removed the rest
|
|
||
|
|
||
| @saving.register_keras_serializable() | ||
| class TDF(layers.Layer): |
There was a problem hiding this comment.
Can there be a more explicit name?
|
|
||
|
|
||
| @saving.register_keras_serializable() | ||
| class TFC(layers.Layer): |
| else: | ||
| model = tfc_tdf_net(keras.Input(sample_batch_x.shape[1:]), name="tfc_tdf_net") | ||
|
|
||
| model.summary() |
There was a problem hiding this comment.
Would it be useful to plot the model, or is the result too busy? e.g. keras.utils.plot_model
There was a problem hiding this comment.
This is how it would look that after a refactoring i've made grouping the TFC_TDF blocks.
It's a little long, but actually simple in structure. Regardless, It can be further refactored to group into decoder and encoder blocks, but perhaps it wont be as informative.
There was a problem hiding this comment.
Ok -- your call on whether to include it.
|
Hi @fchollet , Thanks for the feedback! I did a small refactor to further group TFC_TDF and Downsample/Upsample blocks into custom layers so they could be better visualized in a plot. I also renamed symbols to further avoid abbreviations and updated some docstrings to reflect the changes. The code is currently working on all frameworks after these changes. Also the convert script still runs correctly. |
|
Hi again, I have just pushed the generated files. I also updated layer |
Hi, I saw "Audio track separation" in the call for contributions, so I implemented an example that separates the vocal track from songs in the MUSDB18 dataset.
Some notes:
Please let me know if there is anything wrong with the provided script I may have missed.
Thanks!