Keras fails to train on custom generator (Torch backend)

[DLumi]

Keras model would throw an error on training when using Torch backend and a custom (Python) data generator.

Steps to reproduce:

1. Define a model
2. Define a simple data generator that yields one batch of random data
3. Start training with epoch=1 and steps_per_epoch=1
4. Get an UnboundLocalError: local variable 'logs' referenced before assignment suggesting that the training loop itself was skipped, and epoch_iterator.enumerate_epoch() has issues wrapping provided generator.

The same steps would work just fine with TF.

Colab

Edit: torch==2.3 seems to be incompatible with Keras, but it would have the same fail on 2.2 which should be supported

[DLumi]

Seems to be the data adapter issue. Can't seem to make the training work in Torch regardless of what data I use. I tried feeding both torch's DataLoaders and tf's Datasets, and the results are the same.

[Grvzard]

a temporary workaround is to specify the batch_size explicitly in Model.fit().