Lambda.compute_output_shape generates wrong shape

[AGFACBNNR]

When I use compute_output_shape() to help with shape inferencing, I found the Lambda layer may generate a wrong shape.
The code reproducing it:

from keras.layers import Lambda
import numpy as np
L=Lambda(lambda x:x**2, output_shape=[6])
x=np.zeros((3,2))
print(L(x).shape)
print(L.compute_output_shape(x.shape))

The results would be:
(3, 2)
(3, 6)

[SuryanarayanaY]

Hi @AGFACBNNR ,

Thanks for reporting. It seems users can set output_shape to any value without validating. I tested a demo where a simple model fails in inference when set random output_shape to it. But without setting output_shape it infers from the inputs and inference works fine. Attached gist for same.

I am not actually getting the intent/use case of output_shape here. If users sets output_shape explicitly in a model it will raise an exception if not compatible with the inputs provided. Will escalate to dev team and let's hear from them. Thanks!

[nkovela1]

Hi @AGFACBNNR , since the output_shape was specified to 6, compute_output_shape will return that corresponding shape (3,6). Please try removing the output_shape argument from the Lambda layer. Thanks!

[sampathweb]

Hope this addresses this issue. If you still think its open, please re-open it.

[google-ml-butler]

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