Scaling percentage docs are not consistent

[fepegar]

The config spec docs say:

scaling_percentage
Float array indicates a random spatial scaling should be applied (This can be slow depending on the input volume dimensionality).

It's not clear what the numbers should be, and their effect.

The example says:
scaling_percentage | float array | scaling_percentage=0.8,1.2

I wanted many of my training windows to be zoomed out, so I used scaling_percentage=0.25, 1. I didn't see any difference in my results, so I decided to investigate.

According to an INI file in the repo, it seems like these parameters should be given in terms of the difference in percentage with respect to 100%:

NiftyNet/config/highres3dnet_config.ini

Lines 49 to 58 in 777bb61

	[TRAINING]
	sample_per_volume = 32
	rotation_angle = (-10.0, 10.0)
	scaling_percentage = (-10.0, 10.0)
	lr = 0.0001
	loss_type = Dice
	starting_iter = 0
	save_every_n = 5
	max_iter = 6
	max_checkpoints = 20

And in the code:

NiftyNet/niftynet/layer/rand_spatial_scaling.py

Lines 15 to 35 in 777bb61

	class RandomSpatialScalingLayer(RandomisedLayer):
	"""
	generate randomised scaling along each dim for data augmentation
	"""
	def __init__(self,
	min_percentage=-10.0,
	max_percentage=10.0,
	name='random_spatial_scaling'):
	super(RandomSpatialScalingLayer, self).__init__(name=name)
	assert min_percentage < max_percentage
	self._min_percentage = max(min_percentage, -99.9)
	self._max_percentage = max_percentage
	self._rand_zoom = None
	def randomise(self, spatial_rank=3):
	spatial_rank = int(np.floor(spatial_rank))
	rand_zoom = np.random.uniform(low=self._min_percentage,
	high=self._max_percentage,
	size=(spatial_rank,))
	self._rand_zoom = (rand_zoom + 100.0) / 100.0

So I think the example in the docs should be -20, 20 instead of 0.8,1.2, and I should've used -75, 0 instead of 0.25, 1 (I think with those numbers I actually got the -slightly- opposite effect, a scaling between 1.0025 and 1.01).

[fepegar]

I actually think that the (wrong?) example of the docs is a better format. It feels more intuitive to write 0.25, 1 for scaling parameters than -75, 0.