CategoryEncoding layer lacks consideration of 0-dimension inputs about tensorflow HOT 6 CLOSED

Hi @AGFACBNNR ,

Thanks for reporting. I have checked and acknowledge the reported behaviour. The code works fine with Keras2 but has inconsistent behaviour with Keras3. Attached keras2-gist and keras3-gist for reference.

As per documentation the inputs should be:

inputs: A 1D or 2D tensor of integer inputs.

The behaviour as follows:

# rank-0 Tensor & output_mode='multi_hot'   --> Different results wrt Keras2 & 3
# rank-1 Tensor & output_mode='multi_hot'   --> Same results wrt Keras2 & 3
# rank-2 Tensor & output_mode='multi_hot'   --> Same results wrt Keras2 & 3

# rank-0 Tensor & output_mode='one_hot'     --> Different results wrt Keras2 & 3
# rank-1 Tensor & output_mode='one_hot'     --> Different results wrt Keras2 & 3
# rank-2 Tensor & output_mode='one_hot'     --> Same results wrt Keras2 & 3

IMO, This behaviour of Keras3 should be consistent with keras2. Looking for probable fix.

from tensorflow.

Update: Since Keras3 supports multi backend the Keras3(TF backend) behaviour is different wrt Keras2(TF backend) but inline with Torch and Jax.

But there is issue with compute_output_shape method with below cases which is not same as actual output shape:

# rank-0 Tensor input & output_mode='multi_hot'
# rank-0 Tensor input & output_mode='one_hot'
# rank-1 Tensor input & output_mode='one_hot'

I am going to propose a fix for same. Thanks!

from tensorflow.

Shall be addressed in Keras PR-19388

from tensorflow.

Hi @AGFACBNNR ,

The proposed PR got merged into master. Could you please test in Keras-nightly and let us know. Thanks!

from tensorflow.

I've tested it in Keras-nightly. Thanks a lot.

from tensorflow.

Are you satisfied with the resolution of your issue?
Yes
No

from tensorflow.