Error in transform_to_normalized_grayscale

[matheushent]

I have been debugging some functions and I noticed that transform_to_normalized_grayscale defined in tf_explain/utils/image doesn't return a 4D tensor as described in function.

I think the channels axis is being totally disconsidered. So, the keepdims argument inside reduce_sum must be set to True, furthermore tf.reduce_sum would be properly changed by tf.math.reduce_sum.

In other words:

grayscale_tensor = tf.math.reduce_sum(tensor, axis=-1, keepdims=True)
instead of
grayscale_tensor = tf.reduce_sum(tensor, axis=-1)

[matheushent]

I am a little bit lost about what exactly transform_to_normalized_grayscale does. The output of this function just gives me completely black images instead of gray scale images.

[RaphaelMeudec]

@matheushent Totally agree with you. This function does too much things and should be split into multiple ones. At the moment:

• apply sum along axis and transforms the input of shape (H, W, C) into (H, W, 1)
• cast from float to integers

Hope it helps a bit, will try to make it cleaner