Simplify shape inference in Keras resize_images

[lgeiger]

This PR simplifies shape inference of keras.backend.resize_images, since the shape of the output tensor is already correctly set in image_ops.resize_images_v2:

tensorflow/tensorflow/python/ops/image_ops_impl.py

Line 1390 in 0921a80

images.set_shape([None, new_height_const, new_width_const, None])

This also removes the need for the tf.shape op in cases where the spatial shape of the input tensor is fully defined which fixes #25086.

[mattdangerw]

Please add a unit test!

[lgeiger]

@mattdangerw the correct output shape computations should already be covered in the following test cases

tensorflow/tensorflow/python/keras/backend_test.py

Lines 599 to 615 in c4d6c69

	def test_resize_images(self):
	height_factor = 2
	width_factor = 2
	data_format = 'channels_last'
	x = backend.variable(np.ones((1, 2, 2, 3)))
	y = backend.resize_images(x, height_factor, width_factor, data_format)
	self.assertEqual(y.shape.as_list(), [1, 4, 4, 3])
	data_format = 'channels_first'
	x = backend.variable(np.ones((1, 3, 2, 2)))
	y = backend.resize_images(x, height_factor, width_factor, data_format)
	self.assertEqual(y.shape.as_list(), [1, 3, 4, 4])
	# Invalid use:
	with self.assertRaises(ValueError):
	backend.resize_images(
	x, height_factor, width_factor, data_format='unknown')

But let me know if there is still a case I am missing.

[mattdangerw]

Look like we are missing a case for a partially defined shape to me. Something that exercises the else block in your change.

[lgeiger]

@mattdangerw Sorry about that, I added a test case for dynamic input shapes.

[mattdangerw]

Thanks! lgtm