A tutorial to use randomised permutation during training (#807)

* A tutorial to use randomised permutation during training

Author: vikashg

modules/generate_random_permutations/Creating dataset with randomized transform chain.ipynb

@@ -0,0 +1,110 @@
+import numpy as np
+import os
+import torch
+import glob
+import random
+from monai.data import DataLoader
+from monai.transforms.transform import Transform
+from monai.transforms import (Affine, LoadImage, Rotate, NormalizeIntensity, Transpose, Compose, Resize, AsChannelFirst, AsChannelLast, ScaleIntensity, RandFlip, Rotate90, AddChannel, GaussianSmooth, AdjustContrast)
+from random import shuffle
+
+class Dataset(torch.utils.data.Dataset):
+	def __init__(self, image_file_list, transforms, shuffle_transforms=1):
+		self.image_file_list = image_file_list
+		if shuffle_transforms:
+			transform_list = [LoadImage(image_only=True), AddChannel(), Resize((299, 299))] + shuffle(transforms)
+			self.transform = Compose(transpose_list)
+		else:
+			self.transform = Compose([LoadImage(image_only=True), AddChannel(), Resize((299, 299))] + transforms)
+
+	def __len__(self):
+		return len(self.image_file_list)
+
+	def __getitem__(self, index):
+		return self.transform(self.image_file_list[index])
+
+
+class AugmentData(object):
+	def __init__(self, image_loading_transforms = [LoadImage(image_only=True)], augmentation_dict = {}, num_augmentations=5, output_size=(200, 200), batch_size=3):
+		self.output_size = output_size
+		self.batch_size = batch_size
+		self.augmentation_dict = augmentation_dict
+		self.aug_seq = self.create_augmentation_sequence()
+		self.image_loading_transforms = image_loading_transforms
+		self.num_augmentations = num_augmentations
+
+	def create_augmentation_sequence(self):
+		augmentation_transforms = []
+		for aug, num_aug in self.augmentation_dict.items():
+			_x = [aug]*num_aug
+			augmentation_transforms = augmentation_transforms + _x
+		return augmentation_transforms
+
+
+	def create_transform_list(self, augmentation_sequence):
+		transform_list = self.image_loading_transforms
+		for _aug in augmentation_sequence:
+			if _aug == 'rotate':
+				transform_list.append(Rotate(random.randint(0, 100)))
+			if _aug == 'flip':
+				transform_list.append(RandFlip())
+			if _aug == 'rotate90':
+				transform_list.append(Rotate90())
+			if _aug == 'intensityGaussian':
+				transform_list.append(GaussianSmooth(sigma=random.randint(0, 10)))
+			if _aug == 'adjustContrast':
+				transform_list.append(AdjustContrast(gamma=random.randint(0, 10)))
+
+		transform_list.append(ScaleIntensity())
+		transform_list.append(Resize(self.output_size))
+		return transform_list
+
+
+	def create_native_transform_list(self):
+		transform_list = Compose(self.image_loading_transforms + [ScaleIntensity(), Resize(self.output_size)])
+		return transform_list
+
+
+	def __call__(self, image_file_list, *args, **kwargs):
+		image_file_list = image_file_list
+
+		IMG = []
+		for img in zip(image_file_list):
+			native_transform_list = self.create_native_transform_list()
+			native_img = native_transform_list(img)
+			IMG = IMG + native_img
+			for i in range(self.num_augmentations):
+				shuffle(self.aug_seq)
+				transform_list = self.create_transform_list(self.aug_seq)
+				img_augmentated = Compose(transform_list)(img)
+				IMG  = IMG + img_augmentated
+
+		random.shuffle(IMG)
+		ALLIMG_NP = np.stack(IMG, axis=0)
+		OUT_IMAGE_NP = ALLIMG_NP[0:self.batch_size, :]
+		return OUT_IMAGE_NP
+
+
+
+def main():
+
+	image_dir='./exampleImages'
+	image_file_list = glob.glob(image_dir + '/*.png')
+	output_size = (400, 400)
+	transform_list = [RandFlip(), Rotate(20), NormalizeIntensity(), Rotate90()]
+
+	#print(LoadImage(image_only=True)(image_file_list[0]).shape)
+	#train_dataset=Dataset(image_file_list, transform_list, shuffle_transforms=0)
+	#train_dataloader = DataLoader(train_dataset,  batch_size=4, num_workers=2)
+	#for _batch_data in train_dataloader:
+	#	img = _batch_data[0]
+
+	image_loading_transforms = [LoadImage(image_only=True), AddChannel()]
+	augmentation_dict = {'rotate': 3, 'flip': 2, 'rotate90': 1, 'intensityGaussian': 2, 'adjustContrast' : 2}
+
+	img = AugmentData(image_loading_transforms=image_loading_transforms, augmentation_dict = augmentation_dict)(image_file_list)
+	print(img.shape)
+
+
+if __name__ == '__main__':
+	main()

modules/generate_random_permutations/exampleImages/image1.png

@@ -74,6 +74,7 @@ pattern="-and -name '*' -and ! -wholename '*federated_learning*'\
  -and ! -wholename '*nuclick_infer*'\
  -and ! -wholename '*nuclick_training_notebook*'\
  -and ! -wholename '*full_gpu_inference_pipeline*'\
+ -and ! -wholename '*generate_random_permutations*'\
  -and ! -wholename '*get_started*'"
 kernelspec="python3"