Data augmentation

def add_normal_noise(input_tensor, mean: int = 0, std: float = 0.5):

# Create the tensor containing the noise

noise_tensor = torch.empty_like(input_tensor)

noise_tensor.normal_(mean=mean, std=std)

# Add noise to input tensor and return results

return input_tensor + noise_tensor

# mean = 0, standard deviation is hyperparameter,

# which has to be figured out

simple_dropout = torch.nn.Dropout(p=p)

from torchvision import transforms

transform_chain = transforms.Compose([

transforms.Resize(size=100),

transforms.ColorJitter(),

transforms.RandomRotation(degrees=180),

transforms.RandomVerticalFlip(),

transforms.RandomHorizontalFlip(),

transforms.ToTensor(),

noise_transform

# use transforms.Lambda, therefore noise has to be wrapped

# in an auxiliary function

)]

• Some transforms work on PIL images, others on tensors -> keep order and position of transformation in mind.

• You might need 2 transform chains, one for training and one for evaluation.

• Option 1: implement the transformations in the getitem() method of the dataset, ideal when we want to apply transformations for training and evaluation

• Option 2: implement the transformation in the getitem() method of a different dataset, since we often want to apply different transformation to training and evaluation