What is correct behavior of keras.ops.eye?

[yubori]

Hi,

I have a question about keras.ops.eye of Keras3.

With Tensorflow, the keras.ops.eye accepts integers and floats.
However, with torch and jax, only integers are accepted.

Which behavior is correct?

There is no limitation description on documents.
Because Keras 2 accepts integers and floats, I was confused with it during migration.

[dhantule]

Hi @yubori,
thanks for reporting this. keras.ops.eye will return a 2-D tensor with ones on the diagonal and zeros elsewhere (identity matrix). However, you can control the data type of the resulting matrix to be float32 or float64. The values on the diagonal will still be 1.0 and the other elements will be 0.0 . Attaching gist.

[yubori]

@dhantule Sorry for the lack of my description.

This is a very simple issue.

Please see the following gist.

[dhantule]

@yubori, Thanks for reporting this. I have reproduced this error here and keras.ops.eye produces error when using float values as input dimensions with Jax and PyTorch as backends. Although in Jax the jax.numpy.eye operation doesn’t support float values, similarly with torch.eye. We will look into this issue more. Thanks !

[yubori]

Hi @dhantule
I am looking forward to it being fixed. Thank you.

[mattdangerw]

I think in general we should probably let numpy be the guide here. np.eye will not accept floats or numpy array values, the same is true for jax and torch. This seems correct to me.

The bug is that tf is too permissive here, we could add errors so that keras.ops.eye errors for floats on the tf backend similar to jax and torch.

I'll mark this as contributors welcome. This is a good first issue to bring tf in line with jax and torch.