def transform(x,y,crop_range,rot_range,shear_range,zoom_range,t):
if t == False: # If t is set to false the input is only cropped
crop_type = np.random.randint(0,5,1)[0] # Randomly crop image from either center or a corner.
x_new = SpecialCrop((crop_range),crop_type=crop_type)(x)
y_new = SpecialCrop((crop_range),crop_type=crop_type)(y)
else:
rot = RandomRotate(rot_range,lazy=True)(x)
shear = RandomShear(shear_range,lazy=True)(x)
zoom = RandomZoom(zoom_range,lazy=True)(x)
flip = RandomFlip(v=True,p = np.random.randint(0,2,1)[0])# Images and label is flipped with 0.5 prob.
crop_type = np.random.randint(0,5,1)[0]
x_new = SpecialCrop((crop_range),crop_type=crop_type)(x)
y_new = SpecialCrop((crop_range),crop_type=crop_type)(y)
x_new = th_affine2d(x_new,rot)
y_new = th_affine2d(y_new,rot)
x_new = th_affine2d(x_new,shear)
y_new = th_affine2d(y_new,shear)
x_new = th_affine2d(x_new,zoom)
y_new = th_affine2d(y_new,zoom)
x_new = flip(x_new)
y_new = flip(y_new)
return AddChannel()(x_new), AddChannel()(y_new) # Add channel for concatenating batch.
def AddChannel_setup():
tforms = {}
tforms['addchannel_axis0'] = AddChannel(axis=0)
tforms['addchannel_axis1'] = AddChannel(axis=1)
tforms['addchannel_axis2'] = AddChannel(axis=2)
return tforms
def giana_data_pipeline(file_path):
## For loading complete tensor data into memory
# image_data_file = 'C:/dev/data/Endoviz2018/GIANA/polyp_detection_segmentation/image_data_all_640x640.npy'
# label_data_file = 'C:/dev/data/Endoviz2018/GIANA/polyp_detection_segmentation/gt_data_all_640x640.npy'
## For using csv file to read image locations and use them to load images at run time
## Transformations ############################################################
## TorchVision for getting the right format
## These transforms are applied by pytorch dataloader directly
image_transform = transforms.Compose(
[ResizePadArray(IMAGE_SIZE),
ChannelsFirst(),
TypeCast('float')])
label_transform = transforms.Compose([
ResizePadArray(IMAGE_SIZE),
RangeNormalize(min_val=0, max_val=1),
AddChannel(axis=2),
ChannelsFirst()
])
joint_transform = None
## Imgaug tranforms to get the proper data augmentation
add_image_transform_list = [
iaa.Sometimes(0.5, [
iaa.GaussianBlur(sigma=(0.5, 2.0)),
iaa.Multiply((0.8, 1.2)),
iaa.ContrastNormalization((0.75, 1.5)),
iaa.AdditiveGaussianNoise(
loc=0, scale=(0.0, 0.05 * 255), per_channel=True),
iaa.MotionBlur(size=(5, 10))
])
]
add_joint_transform_list = [
# iaa.Sometimes(0.5, iaa.Affine(translate_percent={"x": (-0.2, 0.2),
# "y": (-0.2, 0.2)},
# rotate=(-15, 15))),
iaa.Fliplr(0.5)
]
##################################################################################
file_list_df = pd.read_csv(file_path, header=None)
giana_dataset = CSVDataset(file_list_df,
input_transform=image_transform,
target_transform=label_transform,
co_transform=joint_transform)
train_dataset, valid_dataset = giana_dataset.train_test_split(TRAIN_RATIO)
## **Inefficient** XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
## When loading all the data into memory
# image_data = np.load(image_data_file)
# label_data = np.load(label_data_file)
# label_data = np.expand_dims(label_data, axis=3)
# giana_dataset = TensorDataset(image_data, label_data, input_transform=image_transform,
# target_transform=label_transform, co_transform=joint_transform)
## XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
giana_train_loader = DataLoader(train_dataset,
batch_size=BATCH_SIZE,
shuffle=True)
giana_valid_loader = DataLoader(valid_dataset,
batch_size=BATCH_SIZE,
shuffle=False)
giana_imgaug_transform = GianaImgAugTransform(
image_transform_list=add_image_transform_list,
joint_transform_list=add_joint_transform_list)
return (giana_imgaug_transform, giana_train_loader, giana_valid_loader)