A checker is needed in Conv layers

[Shuo-Sun20]

As disscussed in tensorflow/tensorflow#64396, it should be considered as an abnormal input when kernel_size > input_size, while currently the Conv layers will silently generate an unexpected output, therefore maybe a checker is needed here.

The code to reproduce this issue:

from keras.layers import Conv2D
import numpy as np

x=np.random.rand(1,2,2,1)
l=Conv2D(1,3,(1,1),'valid','channels_last', [1,1],1, 'linear', True)
print(l(x).shape)
print(l.compute_output_shape(x.shape))

In my opinion, this code should not generate an output, a warning or error is expected since the kernel(3x3) is larger than input (2x2).

[fchollet]

Thanks for the suggestion. When the size of the inputs is statically known before calling the layer, we can add such a check (like in your code example).

However, when the size is only known at runtime, it would not be practical to add ops to the graph to include this check. Instead, it should be up to the backend framework to raise an exception is the operation is invalid.

[Shuo-Sun20]

You are right. I'll close this issue and report this issue to Tensorflow.

[fchollet]

If you're interested, you can also add a check in the base Conv layer's build() method for the static shape case.