The .fit() method passes unsized batches to a custom loss function. about keras

[fchollet]

x.shape is the "static shape" of x. It is often not a number. It can be None. If you want the actual number value, you must use keras.ops.shape(x), e.g.

keras.ops.sum(keras.ops.square(x - y), axis=0) / keras.ops.shape(x)[0]

[SuryanarayanaY]

Hi @Ybisalt ,

I have tested the given code with Keras 3.3.3v and it executes fine. Please note that I have changed the code K.sum to keras.ops.sum and same for K.square also. Please refer to attached gist.

[Ybisalt]

I have tested the given code with Keras 3.3.3v and it executes fine. Please note that I have changed the code K.sum to keras.ops.sum and same for K.square also.

No! Same problem here. You didn't notice "Batch size = None! (None, 4)" line in the last output log.
The number of batches without size depends on the order in which the method is called.

Try just one run of the fit() method (comment out the other fit lines):
log = model.fit(inp_data, out_data, epochs=3, batch_size=30) # Batch size = None!

My gist

[mehtamansi29]

Hi @Ybisalt -

I am not able to reproduce this issue in lates keras(3.5.0) version. Attached gist for the reference.
Are you still able to reproduce this issue ?