Small bug in tf_utils/layers.py

In line 70 I think there is a small mistake, it should be:
`int(input_shape[2] * strides[2]), int(input_shape[3] * strides[3])]`
It's not a major bug, as it would throw an error if the output shape didn't work out correctly. I think it runs OK just because the input is square.