Problem executing callback.py

[drillkicker]

I'm trying to train samplernn in Arch Linux in a venv in my home folder instead of running it inside Google Colab and I get the following error when running train.py:

Traceback (most recent call last):
  File "/home/vasya/work/prism/prism-samplernn/train.py", line 335, in <module>
    main()
  File "/home/vasya/work/prism/prism-samplernn/train.py", line 278, in main
    TrainingStepCallback(
  File "/home/vasya/work/prism/prism-samplernn/callbacks.py", line 17, in __init__
    self.model = model
    ^^^^^^^^^^
AttributeError: property 'model' of 'TrainingStepCallback' object has no setter

I used the default train command in the colab notebook for prism samplernn with a roughly 22 minute mono .wav file that has 48000 samples per second, chunked into 175 8 second chunks. I have tried running both python and python3.

[relativeflux]

Hi @drillkicker,

I am the original author of the code, unfortunately I no longer work at PRiSM so I can only offer limited support. I suspect this is due to what looks like a recent change in TensorFlow/Keras - it seems the Callback class no longer has a model attribute, instead it's been renamed to _model. Hence the complaint about there being no setter.

Did the Notebook run OK, or did you just experience this issue on your own machine?

[drillkicker]

@relativeflux I didn't test the command in the notebook because I'm not familiar with Colab and couldn't figure out how to import a wav file into it. However, I tried running the code after editing the line with the error to say self._model instead of self.model and ran the same command. I get a different error message this time:

Traceback (most recent call last):
  File "/home/vasya/work/prism/prism-samplernn/train.py", line 335, in <module>
    main()
  File "/home/vasya/work/prism/prism-samplernn/train.py", line 285, in main
    ModelCheckpointCallback(
  File "/home/vasya/work/prism/prism-samplernn/callbacks.py", line 79, in __init__
    super(ModelCheckpointCallback, self).__init__(*args, **kwargs)
  File "/home/vasya/work/prism/lib/python3.11/site-packages/keras/src/callbacks/model_checkpoint.py", line 183, in __init__
    raise ValueError(
ValueError: When using `save_weights_only=True` in `ModelCheckpoint`, the filepath provided must end in `.weights.h5` (Keras weights format). Received: filepath=./logdir/default/21.04.2024_14.36.58/model.ckpt-{epoch}

[relativeflux]

@drillkicker Hi - yes, it looks as if the new version of Keras (Keras 3) has a lot of breaking changes for the library. Keras 3 is a complete rewrite. As I said I no longer work as PRiSM, and it doesn't look as if anyone there is maintaining this code now, but I'll get in touch with them and see if I can update it. My apologies. I certainly want to get the library back into a useable condition.

Here is someone highlighting the very error you found, it's definitely due to Keras 3 incompatibility.

If there isn't too much to work on I'll fork the library and update it myself. :)