def transform(self):
if hp.mode == '3d':
if hp.aug:
training_transform = Compose([
# ToCanonical(),
CropOrPad((hp.crop_or_pad_size), padding_mode='reflect'),
# RandomMotion(),
RandomBiasField(),
ZNormalization(),
RandomNoise(),
RandomFlip(axes=(0, )),
OneOf({
RandomAffine(): 0.8,
RandomElasticDeformation(): 0.2,
}),
])
else:
training_transform = Compose([
CropOrPad((hp.crop_or_pad_size, hp.crop_or_pad_size,
hp.crop_or_pad_size),
padding_mode='reflect'),
ZNormalization(),
])
elif hp.mode == '2d':
if hp.aug:
training_transform = Compose([
CropOrPad((hp.crop_or_pad_size), padding_mode='reflect'),
# RandomMotion(),
RandomBiasField(),
ZNormalization(),
RandomNoise(),
RandomFlip(axes=(0, )),
OneOf({
RandomAffine(): 0.8,
RandomElasticDeformation(): 0.2,
}),
])
else:
training_transform = Compose([
CropOrPad((hp.crop_or_pad_size, hp.crop_or_pad_size,
hp.crop_or_pad_size),
padding_mode='reflect'),
ZNormalization(),
])
else:
raise Exception('no such kind of mode!')
return training_transform
def test_reproducibility_oneof(self):
subject1, subject2 = self.get_subjects()
trsfm = Compose([
OneOf([RandomNoise(p=1.0),
RandomSpike(num_spikes=3, p=1.0)]),
RandomNoise(p=.5)
])
transformed1 = trsfm(subject1)
history1 = transformed1.history
trsfm_hist, seeds_hist = compose_from_history(history=history1)
transformed2 = self.apply_transforms(subject2,
trsfm_list=trsfm_hist,
seeds_list=seeds_hist)
data1, data2 = transformed1.img.data, transformed2.img.data
self.assertTensorEqual(data1, data2)
def get_torchio_transformer(mask=False):
if mask:
# Don't apply noise on mask
proba = 0
else:
proba = p
return RandomNoise(std, proba, seed, is_tensor)
def __init__(self, transform1, m1, p1, transform2, m2, p2):
ranges = {
'flip': np.zeros(10),
'affine': np.linspace(0, 180, 10),
'noise': np.linspace(0, 0.5, 10),
'blur': np.arange(10),
'elasticD': np.zeros(10)
}
transforms = {
'flip': lambda magnitude, p: RandomFlip(p=p),
'affine':
lambda magnitude, p: RandomAffine(degrees=(magnitude), p=p),
'noise': lambda magnitude, p: RandomNoise(std=magnitude, p=p),
'blur': lambda magnitude, p: RandomBlur(std=magnitude, p=p),
'elasticD': lambda magnitude, p: RandomElasticDeformation(p=p)
}
self.transform1 = transforms[transform1]
self.t1_input = transform1
self.m1 = ranges[transform1][m1]
self.m1_input = m1
self.p1 = p1
self.transform2 = transforms[transform2]
self.t2_input = transform2
self.m2 = ranges[transform2][m2]
self.m2_input = m2
self.p2 = p2
self.kappa = 0.0
def initialize_transforms_simple(p=0.8):
transforms = [
RandomFlip(axes=(0, 1, 2), flip_probability=1, p=p),
#RandomAffine(scales=(0.9, 1.1), degrees=(10), isotropic=False,
# default_pad_value='otsu', image_interpolation=Interpolation.LINEAR,
# p = p, seed=None),
# *** SLOWS DOWN DATALOADER ***
#RandomElasticDeformation(num_control_points = 7, max_displacement = 7.5,
# locked_borders = 2, image_interpolation = Interpolation.LINEAR,
# p = 0.5, seed = None),
RandomMotion(degrees=10,
translation=10,
num_transforms=2,
image_interpolation='linear',
p=p),
RandomAnisotropy(axes=(0, 1, 2), downsampling=2),
RandomBiasField(coefficients=0.5, order=3, p=p),
RandomBlur(std=(0, 2), p=p),
RandomNoise(mean=0, std=(0, 5), p=p),
RescaleIntensity((0, 255))
]
transform = tio.Compose(transforms)
return transform
def test_transforms(self):
landmarks_dict = dict(
t1=np.linspace(0, 100, 13),
t2=np.linspace(0, 100, 13),
)
random_transforms = (
RandomFlip(axes=(0, 1, 2), flip_probability=1),
RandomNoise(),
RandomBiasField(),
RandomElasticDeformation(proportion_to_augment=1),
RandomAffine(),
RandomMotion(proportion_to_augment=1),
)
intensity_transforms = (
Rescale(),
ZNormalization(),
HistogramStandardization(landmarks_dict=landmarks_dict),
)
for transform in random_transforms:
sample = self.get_sample()
transformed = transform(sample)
for transform in intensity_transforms:
sample = self.get_sample()
transformed = transform(sample)
def get_motion_transform(type='motion1'):
if 'motion1' in type:
dico_params_mot = {
"maxDisp": (1, 6),
"maxRot": (1, 6),
"noiseBasePars": (5, 20, 0.8),
"swallowFrequency": (2, 6, 0.5),
"swallowMagnitude": (3, 6),
"suddenFrequency": (2, 6, 0.5),
"suddenMagnitude": (3, 6),
"verbose": False,
"keep_original": True,
"proba_to_augment": 1,
"preserve_center_pct": 0.1,
"keep_original": True,
"compare_to_original": True,
"oversampling_pct": 0,
"correct_motion": False
}
dico_params_mot = {
"maxDisp": (1, 4),
"maxRot": (1, 4),
"noiseBasePars": (5, 20, 0.8),
"swallowFrequency": (2, 6, 0.5),
"swallowMagnitude": (3, 4),
"suddenFrequency": (2, 6, 0.5),
"suddenMagnitude": (3, 4),
"verbose": False,
"keep_original": True,
"proba_to_augment": 1,
"preserve_center_pct": 0.1,
"keep_original": True,
"compare_to_original": True,
"oversampling_pct": 0,
"correct_motion": False
}
if 'elastic1' in type:
dico_elast = {
'num_control_points': 6,
'max_displacement': (30, 30, 30),
'proportion_to_augment': 1,
'image_interpolation': Interpolation.LINEAR
}
if type == 'motion1':
transforms = Compose([
RandomMotionFromTimeCourse(**dico_params_mot),
])
elif type == 'elastic1_and_motion1':
transforms = Compose([
RandomElasticDeformation(**dico_elast),
RandomMotionFromTimeCourse(**dico_params_mot)
])
if type == 'random_noise_1':
transforms = Compose([RandomNoise(std=(0.020, 0.2))])
return transforms
def test_transforms(self):
landmarks_dict = dict(
t1=np.linspace(0, 100, 13),
t2=np.linspace(0, 100, 13),
)
transforms = (
CenterCropOrPad((9, 21, 30)),
ToCanonical(),
Resample((1, 1.1, 1.25)),
RandomFlip(axes=(0, 1, 2), flip_probability=1),
RandomMotion(proportion_to_augment=1),
RandomGhosting(proportion_to_augment=1, axes=(0, 1, 2)),
RandomSpike(),
RandomNoise(),
RandomBlur(),
RandomSwap(patch_size=2, num_iterations=5),
Lambda(lambda x: 1.5 * x, types_to_apply=INTENSITY),
RandomBiasField(),
Rescale((0, 1)),
ZNormalization(masking_method='label'),
HistogramStandardization(landmarks_dict=landmarks_dict),
RandomElasticDeformation(proportion_to_augment=1),
RandomAffine(),
Pad((1, 2, 3, 0, 5, 6)),
Crop((3, 2, 8, 0, 1, 4)),
)
transformed = self.get_sample()
for transform in transforms:
transformed = transform(transformed)
def get_torchio_transformer(mask=False):
if mask:
# Don't apply noise on mask
proba = 0
else:
proba = p
return RandomNoise(mean=mean, std=std, p=proba, seed=seed)
def test_reproducibility_no_seed(self):
trsfm = RandomNoise()
subject1, subject2 = self.get_subjects()
transformed1 = trsfm(subject1)
transformed2 = trsfm(subject2)
data1, data2 = transformed1.img.data, transformed2.img.data
seed1, seed2 = transformed1.history[0][1][
'seed'], transformed2.history[0][1]['seed']
self.assertNotEqual(seed1, seed2)
self.assertTensorNotEqual(data1, data2)
def test_reproducibility_from_history(self):
trsfm = RandomNoise()
subject1, subject2 = self.get_subjects()
transformed1 = trsfm(subject1)
history1 = transformed1.history
trsfm_hist, seeds_hist = compose_from_history(history=history1)
transformed2 = self.apply_transforms(subject2,
trsfm_list=trsfm_hist,
seeds_list=seeds_hist)
data1, data2 = transformed1.img.data, transformed2.img.data
self.assertTensorEqual(data1, data2)
def test_rng_state(self):
trsfm = RandomNoise()
subject1, subject2 = self.get_subjects()
transformed1 = trsfm(subject1)
seed1 = transformed1.history[0][1]['seed']
value1_torch, value1_np = torch.rand(1).item(), np.random.rand()
transformed2 = trsfm(subject2, seed=seed1)
value2_torch, value2_np = torch.rand(1).item(), np.random.rand()
data1, data2 = transformed1.img.data, transformed2.img.data
self.assertNotEqual(value1_torch, value2_torch)
self.assertNotEqual(value1_np, value2_np)
self.assertTensorEqual(data1, data2)
def get_brats(
data_root='/scratch/weina/dld_data/brats2019/MICCAI_BraTS_2019_Data_Training/',
fold=1,
seed=torch.distributed.get_rank()
if torch.distributed.is_initialized() else 0,
**kwargs):
""" data iter for brats
"""
logging.debug("BratsIter:: fold = {}, seed = {}".format(fold, seed))
# args for transforms
d_size, h_size, w_size = 155, 240, 240
input_size = [7, 223, 223]
spacing = (d_size / input_size[0], h_size / input_size[1],
w_size / input_size[2])
Mean, Std, Max = read_brats_mean(fold, data_root)
normalize = transforms.Normalize(mean=Mean, std=Std)
training_transform = Compose([
# RescaleIntensity((0, 1)), # so that there are no negative values for RandomMotion
# RandomMotion(),
# HistogramStandardization({MRI: landmarks}),
RandomBiasField(),
# ZNormalization(masking_method=ZNormalization.mean),
RandomNoise(),
ToCanonical(),
Resample(spacing),
# CropOrPad((48, 60, 48)),
RandomFlip(axes=(0, )),
OneOf({
RandomAffine(): 0.8,
RandomElasticDeformation(): 0.2,
}),
normalize
])
val_transform = Compose([Resample(spacing), normalize])
train = BratsIter(csv_file=os.path.join(data_root, 'IDH_label',
'train_fold_{}.csv'.format(fold)),
brats_path=os.path.join(data_root, 'all'),
brats_transform=training_transform,
shuffle=True)
val = BratsIter(csv_file=os.path.join(data_root, 'IDH_label',
'val_fold_{}.csv'.format(fold)),
brats_path=os.path.join(data_root, 'all'),
brats_transform=val_transform,
shuffle=False)
return train, val
def random_augment(x):
'''Randomly augment input data.
Returns: Randomly augmented input
'''
# Data augmentations to be used
transforms_dict = {
RandomFlip(): 1,
RandomElasticDeformation(): 1,
RandomAffine(): 1,
RandomNoise(): 1,
RandomBlur(): 1
}
# Create random transform, with a p chance to apply augmentation
transform = OneOf(transforms_dict, p=0.95)
return augment(x, transform)
def predict_majority(model, x, y):
'''Augments all samples of the original data, and chooses majority predictions predicted by the model.
Usage: predict_majority(model, x_original, y_original)
'''
# Reshape arrays
x = np.reshape(x, (len(x), 40, 40, 4, 1))
y = [x - 1 for x in y]
y = to_categorical(y, 5)
# Predict majority
x_flip = augment(x.copy(), RandomFlip())
x_ed = augment(x.copy(), RandomElasticDeformation())
x_affine = augment(x.copy(), RandomAffine())
x_noise = augment(x.copy(), RandomNoise())
x_blur = augment(x.copy(), RandomBlur())
y_true = pred_list(y)
y_pred = pred_list(model.predict(x.copy()))
y_flip = pred_list(model.predict(x_flip.copy()))
y_ed = pred_list(model.predict(x_ed.copy()))
y_affine = pred_list(model.predict(x_affine.copy()))
y_noise = pred_list(model.predict(x_noise.copy()))
y_blur = pred_list(model.predict(x_blur.copy()))
y_most = []
correct = 0
print(
'\nEntry Number | Prediction (None, Flip, Elastic Deformation, Affine, Noise, Blur) | Actual'
)
for i in range(len(y_true)):
preds = [
y_pred[i], y_flip[i], y_ed[i], y_affine[i], y_noise[i], y_blur[i]
]
most = max(set(preds), key=preds.count)
y_most.append(most)
print('Entry', i, '| Predictions:', preds, '| Most Occuring:', most,
'| Correct:', y_true[i])
if most == y_true[i]:
correct += 1
print('\nTest Accuracy: ', correct / len(y_true))
print('Quadratic Weighted Kappa: ',
cohen_kappa_score(y_true, y_most, weights='quadratic'))
def training_network(landmarks, dataset, subjects):
training_transform = Compose([
ToCanonical(),
Resample(4),
CropOrPad((48, 60, 48), padding_mode='reflect'),
RandomMotion(),
HistogramStandardization({'mri': landmarks}),
RandomBiasField(),
ZNormalization(masking_method=ZNormalization.mean),
RandomNoise(),
RandomFlip(axes=(0, )),
OneOf({
RandomAffine(): 0.8,
RandomElasticDeformation(): 0.2,
}),
])
validation_transform = Compose([
ToCanonical(),
Resample(4),
CropOrPad((48, 60, 48), padding_mode='reflect'),
HistogramStandardization({'mri': landmarks}),
ZNormalization(masking_method=ZNormalization.mean),
])
training_split_ratio = 0.9
num_subjects = len(dataset)
num_training_subjects = int(training_split_ratio * num_subjects)
training_subjects = subjects[:num_training_subjects]
validation_subjects = subjects[num_training_subjects:]
training_set = tio.SubjectsDataset(training_subjects,
transform=training_transform)
validation_set = tio.SubjectsDataset(validation_subjects,
transform=validation_transform)
print('Training set:', len(training_set), 'subjects')
print('Validation set:', len(validation_set), 'subjects')
return training_set, validation_set
load_from_dir = [None]
doit = do_training(res_dir, res_name, verbose)
#transforms = get_motion_transform('random_noise_1')
transforms = get_motion_transform('AffFFT_random_noise')
if do_eval:
from torchio.transforms import CropOrPad, RandomAffine, RescaleIntensity, ApplyMask, RandomBiasField, RandomNoise, \
Interpolation, RandomAffineFFT
from utils_file import get_parent_path, gfile, gdir
from utils import get_ep_iter_from_res_name
tc = [RandomNoise(std=(0.020, 0.2))]
if add_affine_rot > 0 or add_affine_zoom > 0:
if add_affine_zoom == 0: add_affine_zoom = 1 # 0 -> no affine so 1
# tc.append(RandomAffine(scales=(add_affine_zoom, add_affine_zoom), degrees=(add_affine_rot, add_affine_rot),
# image_interpolation = Interpolation.NEAREST ))
# name_suffix = '_tAff_nearest_S{}R{}'.format(add_affine_zoom, add_affine_rot)
tc.append(
RandomAffineFFT(scales=(add_affine_zoom, add_affine_zoom),
degrees=(add_affine_rot, add_affine_rot),
oversampling_pct=0.2))
name_suffix = '_tAff_fft_S{}R{}'.format(add_affine_zoom,
add_affine_rot)
else:
name_suffix = '_raw'
def define_transform(transform,
p,
blur_std=4,
motion_trans=10,
motion_deg=10,
motion_num=2,
biascoeff=0.5,
noise_std=0.25,
affine_trans=10,
affine_deg=10,
elastic_disp=7.5,
resample_size=1,
target_shape=0):
### (1) try with different blur
if transform == 'blur':
transforms = [RandomBlur(std=(blur_std, blur_std), p=p, seed=None)]
transforms = Compose(transforms)
### (2) try with different motion artifacts
if transform == 'motion':
transforms = [
RandomMotion(degrees=motion_deg,
translation=motion_trans,
num_transforms=motion_num,
image_interpolation=Interpolation.LINEAR,
p=p,
seed=None),
]
transforms = Compose(transforms)
### (3) with random bias fields
if transform == 'biasfield':
transforms = [
RandomBiasField(coefficients=biascoeff, order=3, p=p, seed=None)
]
transforms = Compose(transforms)
### (4) try with different noise artifacts
if transform == 'noise':
transforms = [
RandomNoise(mean=0, std=(noise_std, noise_std), p=p, seed=None)
]
transforms = Compose(transforms)
### (5) try with different warp (affine transformatins)
if transform == 'affine':
transforms = [
RandomAffine(scales=(1, 1),
degrees=(affine_deg),
isotropic=False,
default_pad_value='otsu',
image_interpolation=Interpolation.LINEAR,
p=p,
seed=None)
]
transforms = Compose(transforms)
### (6) try with different warp (elastic transformations)
if transform == 'elastic':
transforms = [
RandomElasticDeformation(num_control_points=elastic_disp,
max_displacement=20,
locked_borders=2,
image_interpolation=Interpolation.LINEAR,
p=p,
seed=None),
]
transforms = Compose(transforms)
if transform == 'resample':
transforms = [
Resample(target=resample_size,
image_interpolation=Interpolation.LINEAR,
p=p),
CropOrPad(target_shape=target_shape, p=1)
]
transforms = Compose(transforms)
return transforms
def noise(mean, std, p=1):
return RandomNoise(mean=mean, std=std, p=p)
def get_data_loader(cfg: DictConfig, _) -> dict:
log = logging.getLogger(__name__)
transform = Compose([
RandomMotion(),
RandomBiasField(),
RandomNoise(),
RandomFlip(axes=(0, )),
])
log.info(f"Data loader selected: {cfg['dataset']}")
try:
log.info("Attempting to use defined data loader")
dataset = getattr(datasets, cfg["dataset"])(cfg, transform)
except ImportError:
log.info(
"Not a defined data loader... Attempting to use torchio loader")
dataset = getattr(torchio.datasets,
cfg["dataset"])(root=cfg["base_path"],
transform=transform,
download=True)
for subject in random.sample(dataset._subjects, cfg["plot_number"]):
plot_subject(
subject,
os.path.join(os.environ["OUTPUT_PATH"], cfg["save_plot_dir"],
subject["subject_id"]),
)
sampler = GridSampler(patch_size=cfg["patch_size"])
samples_per_volume = len(sampler._compute_locations(
dataset[0])) # type: ignore
with open_dict(cfg):
cfg["size"] = dataset[0].spatial_shape
val_size = max(1, int(0.2 * len(dataset)))
test_set, train_set, val_set = split_dataset(
dataset, [21, len(dataset) - val_size - 21, val_size])
train_loader = __create_data_loader(
train_set,
queue_max_length=samples_per_volume * cfg["queue_length"],
queue_samples_per_volume=samples_per_volume,
sampler=sampler,
verbose=log.level > 0,
batch_size=cfg["batch"],
)
val_loader = __create_data_loader(
val_set,
queue_max_length=samples_per_volume * cfg["queue_length"],
queue_samples_per_volume=samples_per_volume,
sampler=sampler,
verbose=log.level > 0,
batch_size=cfg["batch"],
)
test_loader = __create_data_loader(
test_set,
queue_max_length=samples_per_volume * cfg["queue_length"],
queue_samples_per_volume=samples_per_volume,
sampler=sampler,
verbose=log.level > 0,
batch_size=cfg["batch"],
)
return {
"data_loader_train": train_loader,
"data_loader_val": val_loader,
"data_loader_test": test_loader,
}
num_images=100,
size_range=(193, 229),
force=False,
)
# Each element of subjects_list is an instance of torchio.Subject:
# subject = Subject(
# one_image=torchio.Image(path_to_one_image, torchio.INTENSITY),
# another_image=torchio.Image(path_to_another_image, torchio.INTENSITY),
# a_label=torchio.Image(path_to_a_label, torchio.LABEL),
# )
# Define transforms for data normalization and augmentation
transforms = (
ZNormalization(),
RandomNoise(std=(0, 0.25)),
RandomAffine(scales=(0.9, 1.1), degrees=10),
RandomFlip(axes=(0, )),
)
transform = Compose(transforms)
subjects_dataset = ImagesDataset(subjects_list, transform)
# Run a benchmark for different numbers of workers
workers = range(mp.cpu_count() + 1)
for num_workers in workers:
print('Number of workers:', num_workers)
# Define the dataset as a queue of patches
queue_dataset = Queue(
subjects_dataset,
queue_length,
def get_motion_transform(type='motion1'):
if 'motion1' in type:
from torchio.metrics import SSIM3D, MetricWrapper, MapMetricWrapper
from torchio.metrics.ssim import functional_ssim
from torchio.metrics.old_metrics import th_pearsonr, NCC
from torch.nn import MSELoss, L1Loss
#from torch_similarity.modules import NormalizedCrossCorrelation
metrics = {
# "L1": MetricWrapper("L1", L1Loss()), #same as L1_map
#"NCC_c": MetricWrapper("L1", NCC()),
"L1_map":
MapMetricWrapper("L1_map",
lambda x, y: torch.abs(x - y),
average_method="mean",
mask_keys=['brain']),
# "L2": MapMetricWrapper("L2", MSELoss(), mask_keys=['brain']),
# "SSIM": SSIM3D(average_method="mean"),
"SSIM_mask":
SSIM3D(average_method="mean", mask_keys=["brain"]),
"NCC":
MetricWrapper("NCC_th_brain",
lambda x, y: th_pearsonr(x, y),
use_mask=True,
mask_key='brain'),
"NCC2":
MetricWrapper("NCC_th",
lambda x, y: th_pearsonr(x, y),
use_mask=False),
# "SSIM": MetricWrapper("SSIM", lambda x, y: functional_ssim(x, y, return_map=False),
# use_mask=True, mask_key="brain"),
"ssim_base":
MapMetricWrapper('SSIM_base',
lambda x, y: ssim3D(x, y, size_average=True),
average_method="mean",
mask_keys=['brain'])
}
# metrics = {"L1_map": MapMetricWrapper("L1_map", lambda x, y: torch.abs(x - y), average_method="mean",
# mask_keys=['brain'])}
dico_params_mot = {
"maxDisp": (1, 6),
"maxRot": (1, 6),
"noiseBasePars": (5, 20, 0.8),
"swallowFrequency": (2, 6, 0.5),
"swallowMagnitude": (3, 6),
"suddenFrequency": (2, 6, 0.5),
"suddenMagnitude": (3, 6),
"verbose": False,
"proba_to_augment": 1,
"preserve_center_pct": 0.1,
"compare_to_original": True,
"oversampling_pct": 0,
"correct_motion": False
}
dico_params_mot = {
"maxDisp": (1, 4),
"maxRot": (1, 4),
"noiseBasePars": (5, 20, 0.8),
"swallowFrequency": (2, 6, 0.5),
"swallowMagnitude": (3, 4),
"suddenFrequency": (2, 6, 0.5),
"suddenMagnitude": (3, 4),
"verbose": False,
"proba_to_augment": 1,
"preserve_center_pct": 0.1,
"compare_to_original": True,
"metrics": metrics,
"oversampling_pct": 0,
"correct_motion": False
}
if 'elastic1' in type:
dico_elast = {
'num_control_points': 6,
'max_displacement': (30, 30, 30),
'p': 1,
'image_interpolation': Interpolation.LINEAR
}
if type == 'motion1':
transforms = Compose([
RandomMotionFromTimeCourse(**dico_params_mot),
])
if type == 'elastic1':
transforms = Compose([
RandomElasticDeformation(**dico_elast),
])
elif type == 'elastic1_and_motion1':
transforms = Compose([
RandomElasticDeformation(**dico_elast),
RandomMotionFromTimeCourse(**dico_params_mot)
])
if type == 'random_noise_1':
transforms = Compose([RandomNoise(std=(0.020, 0.2))])
if type == 'AffFFT_random_noise':
transforms = Compose([
RandomAffineFFT(scales=(0.8, 1.2),
degrees=10,
oversampling_pct=0.2,
p=0.75),
RandomNoise(std=(0.020, 0.2))
])
if type == 'AffFFT_random_noise':
transforms = Compose([
RandomAffine(scales=(0.8, 1.2),
degrees=10,
p=0.75,
image_interpolation=Interpolation.NEAREST),
RandomNoise(std=(0.020, 0.2))
])
return transforms
training_batch_size = 12
validation_batch_size = 6
patch_size = 32
samples_per_volume = 20
max_queue_length = 80
training_name = "denseNet3D_torchIO_patch_{}_samples_{}_ADAMOptim_{}Epochs_BS{}_GlorotWeights_SSIM_1511".format(
patch_size, samples_per_volume, Epochs, training_batch_size)
train_writer = SummaryWriter(
os.path.join("runs", "Densenets", training_name + "_training"))
validation_writer = SummaryWriter(
os.path.join("runs", "Densenets", training_name + "_validation"))
training_subjects, test_subjects, validation_subjects = train_test_val_split()
training_transform = Compose([RescaleIntensity((0, 1)), RandomNoise(p=0.05)])
validation_transform = Compose([RescaleIntensity((0, 1))])
test_transform = Compose([RescaleIntensity((0, 1))])
training_dataset = tio.SubjectsDataset(training_subjects,
transform=training_transform)
validation_dataset = tio.SubjectsDataset(validation_subjects,
transform=validation_transform)
test_dataset = tio.SubjectsDataset(test_subjects, transform=test_transform)
'''Patching'''
patches_training_set = tio.Queue(
subjects_dataset=training_dataset,
max_length=max_queue_length,
samples_per_volume=samples_per_volume,
sampler=tio.sampler.UniformSampler(patch_size),
def main():
opt = parsing_data()
print("[INFO]Reading data")
# Dictionary with data parameters for NiftyNet Reader
if torch.cuda.is_available():
print('[INFO] GPU available.')
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
else:
raise Exception(
"[INFO] No GPU found or Wrong gpu id, please run without --cuda")
# FOLDERS
fold_dir = opt.model_dir
fold_dir_model = os.path.join(fold_dir, 'models')
if not os.path.exists(fold_dir_model):
os.makedirs(fold_dir_model)
save_path = os.path.join(fold_dir_model, './CP_{}.pth')
output_path = os.path.join(fold_dir, 'output')
if not os.path.exists(output_path):
os.makedirs(output_path)
output_path = os.path.join(output_path, 'output_{}.nii.gz')
# LOGGING
orig_stdout = sys.stdout
if os.path.exists(os.path.join(fold_dir, 'out.txt')):
compt = 0
while os.path.exists(
os.path.join(fold_dir, 'out_' + str(compt) + '.txt')):
compt += 1
f = open(os.path.join(fold_dir, 'out_' + str(compt) + '.txt'), 'w')
else:
f = open(os.path.join(fold_dir, 'out.txt'), 'w')
sys.stdout = f
# SPLITS
split_path_source = opt.dataset_split_source
assert os.path.isfile(split_path_source), 'source file not found'
split_path_target = opt.dataset_split_target
assert os.path.isfile(split_path_target), 'target file not found'
split_path = dict()
split_path['source'] = split_path_source
split_path['target'] = split_path_target
path_file = dict()
path_file['source'] = opt.path_source
path_file['target'] = opt.path_target
list_split = [
'training',
'validation',
]
paths_dict = dict()
for domain in ['source', 'target']:
df_split = pd.read_csv(split_path[domain], header=None)
list_file = dict()
for split in list_split:
list_file[split] = df_split[df_split[1].isin([split])][0].tolist()
paths_dict_domain = {split: [] for split in list_split}
for split in list_split:
for subject in list_file[split]:
subject_data = []
for modality in MODALITIES[domain]:
subject_data.append(
Image(
modality,
path_file[domain] + subject + modality + '.nii.gz',
torchio.INTENSITY))
if split in ['training', 'validation']:
subject_data.append(
Image('label',
path_file[domain] + subject + 'Label.nii.gz',
torchio.LABEL))
#subject_data[] =
paths_dict_domain[split].append(Subject(*subject_data))
print(domain, split, len(paths_dict_domain[split]))
paths_dict[domain] = paths_dict_domain
# PREPROCESSING
transform_training = dict()
transform_validation = dict()
for domain in ['source', 'target']:
transform_training[domain] = (
ToCanonical(),
ZNormalization(),
CenterCropOrPad((144, 192, 48)),
RandomAffine(scales=(0.9, 1.1), degrees=10),
RandomNoise(std_range=(0, 0.10)),
RandomFlip(axes=(0, )),
)
transform_training[domain] = Compose(transform_training[domain])
transform_validation[domain] = (
ToCanonical(),
ZNormalization(),
CenterCropOrPad((144, 192, 48)),
)
transform_validation[domain] = Compose(transform_validation[domain])
transform = {
'training': transform_training,
'validation': transform_validation
}
# MODEL
norm_op_kwargs = {'eps': 1e-5, 'affine': True}
dropout_op_kwargs = {'p': 0, 'inplace': True}
net_nonlin = nn.LeakyReLU
net_nonlin_kwargs = {'negative_slope': 1e-2, 'inplace': True}
print("[INFO] Building model")
model = Generic_UNet(input_modalities=MODALITIES_TARGET,
base_num_features=32,
num_classes=nb_classes,
num_pool=4,
num_conv_per_stage=2,
feat_map_mul_on_downscale=2,
conv_op=torch.nn.Conv3d,
norm_op=torch.nn.InstanceNorm3d,
norm_op_kwargs=norm_op_kwargs,
nonlin=net_nonlin,
nonlin_kwargs=net_nonlin_kwargs,
convolutional_pooling=False,
convolutional_upsampling=False,
final_nonlin=torch.nn.Softmax(1))
print("[INFO] Training")
train(paths_dict, model, transform, device, save_path, opt)
sys.stdout = orig_stdout
f.close()
def compose_transforms() -> Compose:
print(f"{ctime()}: Setting up transformations...")
"""
# Our Preprocessing Options available in TorchIO are:
* Intensity
- NormalizationTransform
- RescaleIntensity
- ZNormalization
- HistogramStandardization
* Spatial
- CropOrPad
- Crop
- Pad
- Resample
- ToCanonical
We should read and experiment with these, but for now will just use a bunch with
the default values.
"""
preprocessors = [
ToCanonical(p=1),
ZNormalization(masking_method=None,
p=1), # alternately, use RescaleIntensity
]
"""
# Our Augmentation Options available in TorchIO are:
* Spatial
- RandomFlip
- RandomAffine
- RandomElasticDeformation
* Intensity
- RandomMotion
- RandomGhosting
- RandomSpike
- RandomBiasField
- RandomBlur
- RandomNoise
- RandomSwap
We should read and experiment with these, but for now will just use a bunch with
the default values.
"""
augments = [
RandomFlip(axes=(0, 1, 2), flip_probability=0.5),
RandomAffine(image_interpolation="linear",
p=0.8), # default, compromise on speed + quality
# this will be most processing intensive, leave out for now, see results
# RandomElasticDeformation(p=1),
RandomMotion(),
RandomSpike(),
RandomBiasField(),
RandomBlur(),
RandomNoise(),
]
transform = Compose(preprocessors + augments)
print(f"{ctime()}: Transformations registered.")
return transform
def random_stuff(self, seed=42):
transform = RandomNoise(std=(100, 100))
transformed = transform(self.sample_subject, seed=seed)
value = transformed.img.data.sum().item()
seed = transformed.history[0][1]['seed']
return value, seed
# Mock PyTorch model
model = lambda x: x
# Define training and patches sampling parameters
num_epochs = 4
patch_size = 128
queue_length = 100
samples_per_volume = 1
batch_size = 2
# Define transforms for data normalization and augmentation
transforms = (
ZNormalization(),
RandomAffine(scales=(0.9, 1.1), degrees=10),
RandomNoise(std_range=(0, 0.25)),
RandomFlip(axes=(0, )),
)
transform = Compose(transforms)
# Populate a list with dictionaries of paths
one_subject_dict = {
'T1':
dict(path='../BRATS2018_crop_renamed/LGG75_T1.nii.gz',
type=torchio.INTENSITY),
'T2':
dict(path='../BRATS2018_crop_renamed/LGG75_T2.nii.gz',
type=torchio.INTENSITY),
'label':
dict(path='../BRATS2018_crop_renamed/LGG75_Label.nii.gz',
type=torchio.LABEL),
HistogramStandardization,
OneOf,
Compose,
)
d_size, h_size, w_size = 155, 240, 240
input_size = [7, 223, 223]
spacing = (d_size / input_size[0], h_size / input_size[1],
w_size / input_size[2])
training_transform = Compose([
# RescaleIntensity((0, 1)), # so that there are no negative values for RandomMotion
# RandomMotion(),
# HistogramStandardization({MRI: landmarks}),
RandomBiasField(),
# ZNormalization(masking_method=ZNormalization.mean),
RandomNoise(),
ToCanonical(),
Resample(spacing),
# CropOrPad((48, 60, 48)),
RandomFlip(axes=(0, )),
OneOf({
RandomAffine(): 0.8,
RandomElasticDeformation(): 0.2,
}),
])
fold = 1
data_root = '../../dld_data/brats2019/MICCAI_BraTS_2019_Data_Training/'
torch.manual_seed(0)
torch.cuda.manual_seed(0)
tc = CenterCropOrPad(target_shape=(182, 218,212))
tc = CropOrPad(target_shape=(182, 218,182), mask_name='maskk')
#dico_elast = {'num_control_points': 6, 'deformation_std': (30, 30, 30), 'max_displacement': (4, 4, 4),
# 'proportion_to_augment': 1, 'image_interpolation': Interpolation.LINEAR}
#tc = RandomElasticDeformation(**dico_elast)
dico_p = {'num_control_points': 8, 'deformation_std': (20, 20, 20), 'max_displacement': (4, 4, 4),
'p': 1, 'image_interpolation': Interpolation.LINEAR}
dico_p = { 'num_control_points': 6,
#'max_displacement': (20, 20, 20),
'max_displacement': (30, 30, 30),
'p': 1, 'image_interpolation': Interpolation.LINEAR }
t = Compose([ RandomElasticDeformation(**dico_p), tc])
t = Compose([RandomNoise(std=(0.020,0.2)), RandomElasticDeformation(**dico_p) ])
t = Compose([RandomNoise(), RandomElasticDeformation() ])
dataset = ImagesDataset(suj, transform=t); dataset0 = ImagesDataset(suj);
from torch.utils.data import DataLoader
dl = DataLoader(dataset, batch_size=2,
collate_fn=lambda x: x, # this creates a list of Subjects
)
samples = next(iter(dl))
s = dataset[0]; s0=dataset0[0]
ov(s['image']['data'][0])
for i in range(1,50):
s=dataset[0]
dataset.save_sample(s, dict(image='/home/romain/QCcnn/random_motion/random{:.2}.nii'.format(100*s['random_noise'])))
'num_control_points': 8,
'deformation_std': (20, 20, 20),
'max_displacement': (4, 4, 4),
'proportion_to_augment': 1,
'image_interpolation': Interpolation.LINEAR
}
dico_p = {
'num_control_points': 6,
#'max_displacement': (20, 20, 20),
'max_displacement': (30, 30, 30),
'proportion_to_augment': 1,
'image_interpolation': Interpolation.LINEAR
}
t = Compose([RandomElasticDeformation(**dico_p), tc])
t = Compose([RandomNoise(std=(0.020, 0.2))])
dataset = ImagesDataset(suj, transform=t)
s = dataset[0]
ov(s['image']['data'][0])
for i in range(1, 50):
s = dataset[0]
dataset.save_sample(
s,
dict(image='/home/romain/QCcnn/random_motion/random{:.2}.nii'.format(
100 * s['random_noise'])))
t = dataset.get_transform()
type(t)
isinstance(t, Compose)