# ============================================================================================= #
# Author: Lucia Hradecká #
# Copyright: Lucia Hradecká : lucia.d.hradecka@gmail.com #
# #
# MIT License. #
# #
# Permission is hereby granted, free of charge, to any person obtaining a copy #
# of this software and associated documentation files (the "Software"), to deal #
# in the Software without restriction, including without limitation the rights #
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell #
# copies of the Software, and to permit persons to whom the Software is #
# furnished to do so, subject to the following conditions: #
# #
# The above copyright notice and this permission notice shall be included in all #
# copies or substantial portions of the Software. #
# #
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR #
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, #
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE #
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER #
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, #
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE #
# SOFTWARE. #
# ============================================================================================= #
import time
import numpy as np
# TorchIO: requires pytorch, torchIO (pip) (had networkx 2.8.8)
import torchio
# volumentations: with opencv (pip)
import volumentations
# # gunpowder: requires gunpowder (pip) (installed networkx 3.2.1), python>=3.10
# import gunpowder
# import funlib.persistence
# raw = gunpowder.ArrayKey('RAW') # declare arrays to use in the pipeline
# # torchvision: requires pytorch, torchvision
# import torchvision.transforms.v2 as torchvision_v2
# # albumentations: require albumentations
# import albumentations
# bio-volumentations
import src.core.composition as biovol_compose
import src.augmentations as biovol
# libs = ['biovol', 'torchio', 'volum', 'gunpowder', 'torchvision', 'album']
libs = ['volum', 'biovol', 'torchio']
# libs = ['gunpowder']
num_repeat = 100
# num_repeat = 10
if 'gunpowder' in libs[0]:
out_file_name = f'./runtime-{num_repeat}_iterations-gunpowder.txt'
else:
out_file_name = f'./runtime-{num_repeat}_iterations-all-1.txt'
# image_shape_list = [(1, 256, 256, 256), (3, 256, 256, 256), (1, 256, 256, 256, 10), (3, 256, 256, 256, 10)]
image_shape_list = [(1, 256, 256, 256)]
# image_shape_list = [(1, 128, 128, 128)]
crop_shape = (100, 100, 100)
pad_size = 6
pad_shape = (140, 140, 140)
rotation_limit = (22.5,) * 3
translation_limit = (10,) * 3
scaling_limit = (0.8, 1.2) * 3
anisotropic_spacing = (1, 0.5, 2)
blur_sigma_limit = 2
noise_sigma_limits = (0.001, 1)
brightness_limit = 0.2
constrast_limit = 0.2
def get_transformation_list(lib):
"""
Get list of instantiated transformation objects for the given library.
Transformations that are missing or that behave differently to Bio-Volumentations are indicated by the comments.
"""
if lib == 'biovol':
# https://biovolumentations.readthedocs.io/latest/bio_volumentations.augmentations.html#
return [
biovol.RandomCrop(shape=crop_shape, border_mode='constant', p=1), # crop or pad
biovol.Pad(pad_size=pad_size, border_mode='constant', p=1), # only pad
biovol.RandomFlip(axes_to_choose=None, p=1),
biovol.RandomAffineTransform(angle_limit=rotation_limit, translation_limit=translation_limit,
scaling_limit=scaling_limit, p=1), # isotropic data
biovol.RandomAffineTransform(angle_limit=rotation_limit, translation_limit=translation_limit,
scaling_limit=scaling_limit, spacing=anisotropic_spacing, p=1),
# anisotropic data
biovol.RandomGaussianBlur(max_sigma=blur_sigma_limit, p=1),
biovol.GaussianNoise(var_limit=noise_sigma_limits, mean=0, p=1),
biovol.RandomBrightnessContrast(brightness_limit=brightness_limit, contrast_limit=constrast_limit, p=1),
biovol.Normalize(mean=0, std=1, p=1),
]
if lib == 'torchio':
# https://torchio.readthedocs.io/
return [
torchio.CropOrPad(target_shape=crop_shape, padding_mode=0, p=1), # TODO center or random? no one knows
torchio.Pad(padding=pad_size, padding_mode=0, p=1),
torchio.RandomFlip(axes=(0, 1, 2), flip_probability=0.5, p=1),
torchio.RandomAffine(scales=scaling_limit, degrees=rotation_limit, translation=translation_limit, p=1),
# MISSING: anisotropic affine transform (their 'isotropic' has a different meaning)
torchio.RandomBlur(std=blur_sigma_limit, p=1),
torchio.RandomNoise(mean=0, std=noise_sigma_limits, p=1),
# MISSING: brightness contrast
torchio.ZNormalization(p=1)
]
if lib == 'volum':
# https://github.com/ZFTurbo/volumentations/blob/master/volumentations/augmentations/transforms.py
# TODO: no generated docs + some classes are not documented at all (even in the code)
rotate_limit_vol = (-22.5, 22.5)
return [
volumentations.RandomCrop(shape=crop_shape, p=1), # TODO only crop
volumentations.PadIfNeeded(shape=pad_shape, border_mode='constant', p=1), # TODO only pad ??
volumentations.Flip(axis=None, p=1), # this is random variant (no deterministic exists)
# MISSING: affine and translation -> using a sequence of rotation and scale at least
[volumentations.Rotate(x_limit=rotate_limit_vol, y_limit=rotate_limit_vol, z_limit=rotate_limit_vol, p=1),
volumentations.RandomScale(scaling_limit[:2], p=1)],
# MISSING: anisotropic affine (or anisotropic anything)
# MISSING: implementation of Blur ... volumentations.Blur(blur_limit=(3, 7), p=1), # TODO: unclear meaning of param blur_limit; is it Gaussian? / alternative: GlassBlur - but docs say it's noise + it isn't gaussian blur only
volumentations.GaussianNoise(var_limit=noise_sigma_limits, mean=0, p=1),
volumentations.RandomBrightnessContrast(brightness_limit=brightness_limit, contrast_limit=constrast_limit,
p=1),
volumentations.Normalize(range_norm=False, p=1),
]
if lib == 'gunpowder':
# https://funkelab.github.io/gunpowder/
return [
[gunpowder.Crop(key=raw, roi=gunpowder.Roi((0,) * 4, (1,) + crop_shape)), gunpowder.RandomLocation()], # only random crop # TODO this probably operates at smaller image from the begining
gunpowder.Pad(key=raw, size=gunpowder.Coordinate((0,) + (pad_size,) * 3)), # only pad # TODO this pads the image but only retrieves the original-sized image at the end
gunpowder.SimpleAugment(transpose_probs=[0, 0, 0, 0], p=1), # only mirror=flip
# MISSING: affine (only have DeformAugment = elastic transform (rotation, scaling, jitter; needs physical units), and deprecated ElasticAugment)
# gunpowder.DeformAugment(control_point_spacing=gunpowder.Coordinate((1, 16, 16, 16)),
# jitter_sigma=gunpowder.Coordinate((0,)*4),
# scale_interval=scaling_limit[:2], rotate=True, spatial_dims=3, p=1), # TODO no way to control rotation + throws some error related to voxel size (WTF?)
# MISSING: affine anisotropic
# MISSING: blur
gunpowder.NoiseAugment(array=raw, mode='gaussian', clip=False, p=1), # TODO image should be of type float and within range [-1, 1] or [0, 1]
gunpowder.IntensityAugment(array=raw, scale_min=1 - constrast_limit, scale_max=1 + constrast_limit,
shift_min=-brightness_limit, shift_max=brightness_limit, z_section_wise=False,
clip=False, p=1),
# MISSING: normalize to 0 mean and 1 variance ... there's only normalization to [0,1] interval: gunpowder.Normalize(array=raw)
]
# elastic_augment = gp.ElasticAugment(
# control_point_spacing=(16, 16),
# jitter_sigma=(4.0, 4.0),
# rotation_interval=(0, math.pi/2))
if lib == 'torchvision':
pass
if lib == 'album':
pass
def init_compose(lib, transform):
"""
Return instantiated transformation pipeline.
"""
if lib == 'biovol':
return biovol_compose.Compose(transforms=[transform], p=1)
if lib == 'torchio':
return torchio.Compose([transform], p=1)
if lib == 'volum':
if isinstance(transform, list):
# the "affine" transform
return volumentations.Compose(transform, p=1.0)
else:
return volumentations.Compose([transform], p=1.0)
if lib == 'gunpowder':
if isinstance(transform, list):
return transform[0] + transform[1] # hardcoded but the easiest way at the moment
else:
return transform
if lib == 'torchvision':
pass
if lib == 'album':
pass
def get_input_data(lib, shape):
"""
Return input data in the correct format.
"""
# Shape is given as ([C], D, H, W, [T])
if lib == 'biovol':
# ([C], D, H, W, [T])
return {'image': np.random.uniform(low=0, high=1, size=shape)}
if lib == 'torchio':
# (C, W, H, D)
shape_torchio = (shape[0], shape[3], shape[2], shape[1])
return np.random.uniform(low=0, high=1, size=shape_torchio)
if lib == 'volum':
# (H, W, D, [C]) or (D, H, W, [C])
shape_volum = (shape[2], shape[3], shape[1], shape[0])
return {'image': np.random.uniform(low=0, high=1, size=shape_volum)}
if lib == 'gunpowder':
# ([C], [D], H, W)
# source = gunpowder.ArraySource(key=raw, array=gunpowder.Array(np.random.uniform(low=0, high=1, size=shape),
# spec=gunpowder.ArraySpec()))
source = gunpowder.ArraySource(key=raw,
array=funlib.persistence.Array(np.random.uniform(low=0, high=1, size=shape)))
# formulate a request for "raw"
request = gunpowder.BatchRequest()
request[raw] = gunpowder.Roi((0,)*len(shape), shape) # an offset and a size
# request[raw] = gunpowder.Roi((0,)*4, (None,)*4) # TODO can I get "the whole image"? Why does this fail?
return {'source': source, 'request': request}
if lib == 'torchvision':
pass
if lib == 'album':
pass
def transform_data(lib, data, pipeline):
"""
Run the transformation pipeline and return shape of the resulting image.
"""
if lib == 'biovol':
augm_data = pipeline(**data)
return augm_data['image'].shape # do something to enforce performing the action
if lib == 'torchio':
return pipeline(data).shape # do something to enforce performing the action
if lib == 'volum':
augm_data = pipeline(**data)
return augm_data['image'].shape # do something to enforce performing the action
if lib == 'gunpowder':
whole_pipeline = data['source'] + pipeline
if 'Crop' in str(pipeline):
data['request'][raw] = gunpowder.Roi((0,)*(1+len(crop_shape)), (1,) + crop_shape)
with gunpowder.build(whole_pipeline):
batch = whole_pipeline.request_batch(data['request'])
return batch[raw].data.shape # batch[raw].data is np.ndarray - do something to enforce performing the action
if lib == 'torchvision':
pass
if lib == 'album':
pass
def single_transform(iterations, shape, augmentation, lib) -> np.ndarray:
times = np.zeros(iterations)
for i in range(iterations):
# prepare data and transformation pipeline
transformation_pipeline = init_compose(lib, augmentation)
data = get_input_data(lib, shape)
# run and measure
t_0 = time.time()
_ = transform_data(lib, data, transformation_pipeline)
time_spent = time.time() - t_0
times[i] = time_spent * 1000 # convert from seconds to milliseconds
return times
def transformation_speed_benchmark(iterations):
f = open(out_file_name, 'w')
for lib in libs:
print(f'*************** LIBRARY {lib} ***************')
for i, augmentation in enumerate(get_transformation_list(lib)):
aug_name = augmentation.__class__.__name__
if isinstance(augmentation, list):
aug_name = ''.join(['Composed'] + ['-' + a.__class__.__name__ for a in augmentation])
print(aug_name)
for shape in image_shape_list:
# the first run (prepare the environment)
transformation_pipeline = init_compose(lib, augmentation)
data = get_input_data(lib, shape)
first_time = time.time()
_ = transform_data(lib, data, transformation_pipeline)
first_result = time.time() - first_time # this is in seconds !!!
# the measured runs
times_arr = single_transform(iterations, shape, augmentation, lib) # this is in milliseconds
# logging
log_message = f'Runtime in seconds. ' \
f'FirstRun: {first_result:.3f}, Average: {times_arr.mean():.3f}, ' \
f'Maximum: {times_arr.max():.3f}, Std: {times_arr.std():.3f}. ' \
f'(Library: {lib}, Transform: {aug_name}, Iterations: {iterations}, ImageSize: {shape})\n'
f.write(log_message)
print(log_message)
print()
f.close()
if __name__ == '__main__':
transformation_speed_benchmark(num_repeat)