Behavior of solarize() depends on integer overflow

[EIFY]

I am not 100% sure about the intention but I do want to raise the alarm. The solarize() transform here

big_vision/big_vision/pp/autoaugment.py

Lines 180 to 184 in 01edb81

	def solarize(image, threshold=128):
	# For each pixel in the image, select the pixel
	# if the value is less than the threshold.
	# Otherwise, subtract 255 from the pixel.
	return tf.where(image < threshold, image, 255 - image)

inverts the pixel when its value is greater or equal to the threshold, so one would think that higher augmentation magnitude needs lower threshold. However, the threshold increases linearly with magnitude:

big_vision/big_vision/pp/autoaugment.py

Line 513 in 01edb81

'Solarize': lambda level: (int((level/_MAX_LEVEL) * 256),),

Counterintuitively, it still works as expected with magnitude=_MAX_LEVEL because of integer overflow. Given

t = tf.constant([[[0,0,0]]], dtype=tf.uint8)

t < i evaluates to tf.Tensor([[[False False False]]], shape=(1, 1, 3), dtype=bool) iff not (i % 256). In other words, magnitude=_MAX_LEVEL means int((level/_MAX_LEVEL) * 256) = 256, which is equivalent to 0 in tf.uint8. Given the following tf_gradient that goes from (0, 0, 0) to (255, 255, 255) in alternating directions

Both solarize(tf_gradient, 256) and solarize(tf_gradient, 0) indeed fully invert the image:

But if magnitude is 9, int((9/10) * 256) = 230, and solarize(tf_gradient, 230) "abruptly" only inverts a small portion of the image: