Getting error when calculating entropy for each images in the batch in the input tensor in a custom layer in tensorflow/Keras. about tf-keras

[tilakrayal]

@sachinprasadhs,
I was able to reproduce the issue on tensorflow v2.11 and tf-nightly. Kindly find the gist of it here.

[haifeng-jin]

@DeveshS215 Seems the problem is tf.get_static_value(image_ith_filter_entropy) returns None because of the tensor value cannot be calculated.
Would you please help simplify the code to the specific function and its arguments?

Thanks.

[DeveshS215]

@haifeng-jin

Thank you for your help. But the thing is I already provided the simplified code as much I can. I can't simplify more because I need to calculate the entropy of each feature map of each image. But in keras layer the batch size is None during model building and so we can't calculate entropy and this is the reason of getting error and I am unable to resolve it.

[Frightera]

@DeveshS215 I think the above gist is not complete as entropy_sort_index is not defined. Also it is not clear to me what .item() method should do in this case? Can you clarify it a bit more?

Thanks.

[haifeng-jin]

After consulting the team, we do not aware of any good solution. Let's leave this issue open to see if we can either find some community help on how to implement or consolidate the feature request to a more concrete one.

[DeveshS215]

@Frightera Thanks for pointing my mistake. Actually its not entropy_sort_index instead it is entropy_array . I have corrected it. I am really sorry for it.

For .item() method : I need value of image_ith_filter_entropy to append in entropy list as number, so I used .item() to get the value as number because tf.get_static_value() returns value in ndarray format.

Thanks.

[DeveshS215]

@haifeng-jin

OK.
Thanks for your help.

[Frightera]

@DeveshS215 I see, so if it is a scalar value, maybe you can employ some logic using tf.reduce.xxx.