def get_transforms_aug( size_input ):
return transforms.Compose([
#------------------------------------------------------------------
#Resize
mtrans.ToResize( (size_input,size_input), resize_mode='square', padding_mode=cv2.BORDER_REPLICATE),
#------------------------------------------------------------------
#Colors
mtrans.ToRandomTransform( mtrans.RandomBrightness( factor=0.25 ), prob=0.750 ),
mtrans.ToRandomTransform( mtrans.RandomContrast( factor=0.25 ), prob=0.750 ),
mtrans.ToRandomTransform( mtrans.RandomGamma( factor=0.25 ), prob=0.750 ),
# mtrans.ToRandomTransform( mtrans.RandomRGBPermutation(), prob=0.50 ),
# mtrans.ToRandomTransform( mtrans.CLAHE(), prob=0.25 ),
mtrans.ToRandomTransform(mtrans.ToGaussianBlur( sigma=0.05 ), prob=0.25 ),
# mtrans.ToRandomTransform(mtrans.ToGaussianNoise( sigma=0.05 ), prob=0.25 ),
# mtrans.ToRandomTransform( mtrans.RandomBrightness( factor=0.25 ), prob=0.50 ),
# mtrans.ToRandomTransform( mtrans.RandomContrast( factor=0.25 ), prob=0.50 ),
# mtrans.ToRandomTransform( mtrans.RandomGamma( factor=0.25 ), prob=0.50 ),
# mtrans.ToRandomTransform( mtrans.RandomRGBPermutation(), prob=0.50 ),
# mtrans.ToRandomTransform( mtrans.CLAHE(), prob=0.25 ),
# mtrans.ToRandomTransform( mtrans.ToGaussianBlur( sigma=0.05 ), prob=0.25 ),
#------------------------------------------------------------------
mtrans.ToTensor(),
normalize,
])
def get_transforms_fer_aug( size_input ):
return transforms.Compose([
#------------------------------------------------------------------
#Resize input
mtrans.ToResize( (48, 48), resize_mode='square', padding_mode=cv2.BORDER_REFLECT ),
#------------------------------------------------------------------
#Geometric
mtrans.RandomScale(factor=0.2, padding_mode=cv2.BORDER_REPLICATE ),
mtrans.ToRandomTransform( mtrans.RandomGeometricalTransform( angle=30, translation=0.2, warp=0.02, padding_mode=cv2.BORDER_REFLECT ), prob=0.5 ),
mtrans.ToRandomTransform( mtrans.VFlip(), prob=0.5 ),
#mtrans.ToRandomTransform( mtrans.HFlip(), prob=0.5 ),
#------------------------------------------------------------------
#Colors
mtrans.ToRandomTransform( mtrans.RandomBrightness( factor=0.25 ), prob=0.50 ),
mtrans.ToRandomTransform( mtrans.RandomContrast( factor=0.25 ), prob=0.50 ),
mtrans.ToRandomTransform( mtrans.RandomGamma( factor=0.25 ), prob=0.50 ),
mtrans.ToRandomTransform( mtrans.RandomRGBPermutation(), prob=0.50 ),
mtrans.ToRandomTransform( mtrans.CLAHE(), prob=0.25 ),
mtrans.ToRandomTransform( mtrans.ToGaussianBlur( sigma=0.05 ), prob=0.25 ),
#------------------------------------------------------------------
#Resize
mtrans.ToResize( (size_input+5, size_input+5), resize_mode='square', padding_mode=cv2.BORDER_REFLECT ),
mtrans.RandomCrop( (size_input, size_input), limit=2, padding_mode=cv2.BORDER_REFLECT ),
#------------------------------------------------------------------
mtrans.ToGrayscale(),
mtrans.ToTensor(),
normalize,
])
def get_transforms_geom_color(): #7
return transforms.Compose([
#------------------------------------------------------------------
#Resize
#
#mtrans.CenterCrop((1008, 1008)),
#mtrans.RandomCrop( (size_crop, size_crop), limit=10, padding_mode=cv2.BORDER_REFLECT_101 ),
#mtrans.ToResize( (size_input, size_input), resize_mode='square', padding_mode=cv2.BORDER_REFLECT_101 ),
#mtrans.ToPad( pad, pad, padding_mode=cv2.BORDER_CONSTANT ),
#------------------------------------------------------------------
#Geometric
mtrans.ToRandomTransform(mtrans.VFlip(), prob=0.5),
mtrans.ToRandomTransform(mtrans.HFlip(), prob=0.5),
mtrans.RandomScale(factor=0.3, padding_mode=cv2.BORDER_REFLECT_101),
mtrans.ToRandomTransform(mtrans.RandomGeometricalTransform(
angle=45,
translation=0.2,
warp=0.02,
padding_mode=cv2.BORDER_REFLECT_101),
prob=0.5),
mtrans.ToRandomTransform(mtrans.RandomElasticDistort(
size_grid=32, deform=None, padding_mode=cv2.BORDER_REFLECT_101),
prob=0.5),
mtrans.ToRandomTransform(mtrans.RandomErased(p_grid=0.2, p_img=0.5),
prob=0.1),
mtrans.ToRandomTransform(mtrans.RandomLines(n_lines=5), prob=0.1),
#mtrans.ToResizeUNetFoV(imsize, cv2.BORDER_REFLECT_101),
#------------------------------------------------------------------
#Colors
mtrans.ToRandomTransform(mtrans.RandomBrightness(factor=0.25),
prob=0.50),
mtrans.ToRandomTransform(mtrans.RandomContrast(factor=0.25),
prob=0.50),
mtrans.ToRandomTransform(mtrans.RandomGamma(factor=0.25), prob=0.50),
#mtrans.ToRandomTransform( mtrans.RandomRGBPermutation(), prob=0.50 ),
#mtrans.ToRandomTransform( mtrans.CLAHE(), prob=0.25 ),
mtrans.ToRandomTransform(mtrans.ToGaussianBlur(sigma=0.05), prob=0.1),
#------------------------------------------------------------------
mtrans.ToTensor(),
normalize,
])
def get_transforms_aug( size_input ):
transforms_aug = transforms.Compose([
mtrans.ToResize( (size_input, size_input), resize_mode='square', padding_mode=cv2.BORDER_REPLICATE ) ,
#mtrans.ToResize( (size_input+20, size_input+20), resize_mode='asp' ) ,
#mtrans.RandomCrop( (size_input, size_input), limit=10, padding_mode=cv2.BORDER_REFLECT_101 ) ,
#------------------------------------------------------------------
#Geometric
mtrans.RandomScale(factor=0.2, padding_mode=cv2.BORDER_REPLICATE ),
mtrans.RandomGeometricalTransform( angle=30, translation=0.2, warp=0.02, padding_mode=cv2.BORDER_REPLICATE),
mtrans.ToRandomTransform( mtrans.HFlip(), prob=0.5 ),
#------------------------------------------------------------------
#Colors
mtrans.ToRandomTransform( mtrans.RandomRGBPermutation(), prob=0.30 ),
mtrans.ToRandomTransform( mtrans.RandomBrightness( factor=0.15 ), prob=0.50 ),
mtrans.ToRandomTransform( mtrans.RandomContrast( factor=0.15 ), prob=0.50 ),
mtrans.ToRandomTransform( mtrans.RandomGamma( factor=0.15 ), prob=0.50 ),
mtrans.ToRandomTransform( mtrans.RandomHueSaturation( hue_shift_limit=(-5, 5), sat_shift_limit=(-11, 11), val_shift_limit=(-11, 11) ), prob=0.30 ),
mtrans.ToRandomTransform( mtrans.ToGrayscale(), prob=0.30 ),
#mtrans.ToRandomChoiceTransform( [
# mtrans.RandomBrightness( factor=0.15 ),
# mtrans.RandomContrast( factor=0.15 ),
# #mtrans.RandomSaturation( factor=0.15 ),
# mtrans.RandomHueSaturation( hue_shift_limit=(-5, 5), sat_shift_limit=(-11, 11), val_shift_limit=(-11, 11) ),
# mtrans.RandomGamma( factor=0.30 ),
# mtrans.ToRandomTransform(mtrans.ToGrayscale(), prob=0.15 ),
# ]),
mtrans.ToRandomTransform(mtrans.ToGaussianBlur( sigma=0.0001), prob=0.25 ),
#------------------------------------------------------------------
mtrans.ToTensor(),
normalize,
])
return transforms_aug
def main():
# parameters
parser = arg_parser()
args = parser.parse_args()
imsize = args.image_size
parallel = args.parallel
num_classes = 2
num_channels = 3
view_freq = 2
network = SegmentationNeuralNet(
patchproject=args.project,
nameproject=args.name,
no_cuda=args.no_cuda,
parallel=parallel,
seed=args.seed,
print_freq=args.print_freq,
gpu=args.gpu,
view_freq=view_freq,
)
network.create(arch=args.arch,
num_output_channels=num_classes,
num_input_channels=num_channels,
loss=args.loss,
lr=args.lr,
momentum=args.momentum,
optimizer=args.opt,
lrsch=args.scheduler,
pretrained=args.finetuning,
size_input=imsize)
# resume
network.resume(os.path.join(network.pathmodels, args.resume))
cudnn.benchmark = True
# datasets
# training dataset
train_data = tgsdata.TGSDataset(
args.data,
tgsdata.train,
count=16000,
num_channels=num_channels,
transform=transforms.Compose([
mtrans.ToRandomTransform(mtrans.HFlip(), prob=0.5),
mtrans.ToRandomTransform(mtrans.VFlip(), prob=0.5),
mtrans.ToResize((300, 300),
resize_mode='squash',
padding_mode=cv2.BORDER_REFLECT_101),
mtrans.RandomCrop((256, 256),
limit=10,
padding_mode=cv2.BORDER_REFLECT_101),
mtrans.RandomScale(factor=0.2,
padding_mode=cv2.BORDER_REFLECT_101),
mtrans.RandomGeometricalTransform(
angle=30,
translation=0.2,
warp=0.02,
padding_mode=cv2.BORDER_REFLECT_101),
#mtrans.ToResizeUNetFoV(imsize, cv2.BORDER_REFLECT_101),
mtrans.ToRandomTransform(mtrans.RandomBrightness(factor=0.15),
prob=0.50),
mtrans.ToRandomTransform(mtrans.RandomContrast(factor=0.15),
prob=0.50),
mtrans.ToRandomTransform(mtrans.RandomGamma(factor=0.15),
prob=0.50),
mtrans.ToRandomTransform(mtrans.ToGaussianBlur(), prob=0.15),
mtrans.ToTensor(),
mtrans.ToMeanNormalization(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
)
#mtrans.ToNormalization(),
]))
train_loader = DataLoader(train_data,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=network.cuda,
drop_last=True)
# validate dataset
val_data = tgsdata.TGSDataset(
args.data,
tgsdata.test,
count=4000,
num_channels=num_channels,
transform=transforms.Compose([
mtrans.ToResize((256, 256), resize_mode='squash'),
#mtrans.RandomCrop( (255,255), limit=50, padding_mode=cv2.BORDER_CONSTANT ),
#mtrans.ToResizeUNetFoV(imsize, cv2.BORDER_REFLECT_101),
mtrans.ToTensor(),
mtrans.ToMeanNormalization(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
)
#mtrans.ToNormalization(),
]))
val_loader = DataLoader(val_data,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=network.cuda,
drop_last=True)
# print neural net class
print('SEG-Torch: {}'.format(datetime.datetime.now()))
print(network)
# training neural net
network.fit(train_loader, val_loader, args.epochs, args.snapshot)
print("Optimization Finished!")
print("DONE!!!")
## color
#mtrans.RandomSaturation(),
#mtrans.RandomHueSaturationShift(),
#mtrans.RandomHueSaturation(),
#mtrans.RandomRGBShift(),
#mtrans.ToNegative(),
#mtrans.RandomRGBPermutation(),
#mtrans.ToGrayscale(),
## blur
#mtrans.ToRandomTransform( mtrans.ToLinealMotionBlur( lmax=1 ), prob=0.5 ),
#mtrans.ToRandomTransform( mtrans.ToMotionBlur( ), prob=0.5 ),
mtrans.ToRandomTransform( mtrans.ToGaussianBlur(), prob=0.5 ),
## geometrical
#mtrans.ToRandomTransform( mtrans.HFlip(), prob=0.5 )
#mtrans.ToRandomTransform( mtrans.VFlip(), prob=0.5 )
#mtrans.RandomScale(factor=0.2, padding_mode=cv2.BORDER_REFLECT101 ),
#mtrans.RandomGeometricalTranform( angle=360, translation=0.2, warp=0.02, padding_mode=cv2.BORDER_REFLECT101),
#mtrans.RandomElasticDistort( size_grid=50, padding_mode=cv2.BORDER_REFLECT101 ),
## tensor
mtrans.ToTensor(),
#mtrans.RandomElasticTensorDistort( size_grid=10, deform=0.05 ),
## normalization
def test_ellipse():
data = SyntethicCircleDataset(
count=300,
generate=SyntethicCircleDataset.generate_image_mask_and_weight,
imsize=(512, 612),
sigma=0.01,
bdraw_grid=True,
transform=transforms.Compose([
## resize and crop
mtrans.ToResize((400, 400),
resize_mode='square',
padding_mode=cv2.BORDER_REFLECT_101),
#mtrans.CenterCrop( (200,200) ),
#mtrans.RandomCrop( (255,255), limit=50, padding_mode=cv2.BORDER_REFLECT_101 ),
#mtrans.ToResizeUNetFoV(388, cv2.BORDER_REFLECT_101),
## color
mtrans.ToRandomChoiceTransform([
mtrans.RandomSaturation(),
mtrans.RandomHueSaturationShift(),
mtrans.RandomHueSaturation(),
#mtrans.RandomRGBShift(),
#mtrans.ToNegative(),
#mtrans.RandomRGBPermutation(),
#mtrans.ToRandomTransform( mtrans.ToGrayscale(), prob=0.5 ),
mtrans.ToGrayscale(),
]),
## blur
#mtrans.ToRandomTransform( mtrans.ToLinealMotionBlur( lmax=1 ), prob=0.5 ),
#mtrans.ToRandomTransform( mtrans.ToMotionBlur( ), prob=0.5 ),
mtrans.ToRandomTransform(mtrans.ToGaussianBlur(), prob=0.75),
## geometrical
#mtrans.ToRandomTransform( mtrans.HFlip(), prob=0.5 )
#mtrans.ToRandomTransform( mtrans.VFlip(), prob=0.5 )
mtrans.RandomScale(factor=0.2, padding_mode=cv2.BORDER_REFLECT101),
#mtrans.RandomGeometricalTransform( angle=360, translation=0.2, warp=0.02, padding_mode=cv2.BORDER_REFLECT101),
#mtrans.RandomElasticDistort( size_grid=50, padding_mode=cv2.BORDER_REFLECT101 ),
## tensor
mtrans.ToTensor(),
mtrans.RandomElasticTensorDistort(size_grid=10, deform=0.05),
## normalization
mtrans.ToNormalization(),
#mtrans.ToWhiteNormalization(),
#mtrans.ToMeanNormalization(
# mean=[0.485, 0.456, 0.406],
# std=[0.229, 0.224, 0.225]
# ),
]))
dataloader = DataLoader(data, batch_size=3, shuffle=True, num_workers=1)
label_batched = []
for i_batch, sample_batched in enumerate(dataloader):
print(i_batch, sample_batched['image'].size(),
sample_batched['label'].size(), sample_batched['weight'].size())
image = sample_batched['image']
label = sample_batched['label']
weight = sample_batched['weight']
print(torch.min(image), torch.max(image), image.shape)
print(torch.min(label), torch.max(label), image.shape)
print(torch.min(weight), torch.max(weight), image.shape)
print(image.shape)
print(np.unique(label))
print(image.min(), image.max())
image = image.permute(2, 3, 1, 0)[:, :, :, 0].squeeze()
label = label.permute(2, 3, 1, 0)[:, :, :, 0].squeeze()
weight = weight.permute(2, 3, 1, 0)[:, :, :, 0].squeeze()
plt.figure(figsize=(16, 4))
plt.subplot(131)
plt.imshow(image)
plt.title('image + [grid]')
plt.axis('off')
plt.ioff()
plt.subplot(132)
plt.imshow(label)
plt.title('gt')
plt.axis('off')
plt.ioff()
plt.subplot(133)
plt.imshow(weight)
plt.title('weight map')
plt.axis('off')
plt.ioff()
#print('save figure ...')
#plt.savefig('../out/image_{}.png'.format(i_batch) )
plt.show()
# observe 4th batch and stop.
if i_batch == 1:
break