A Brief Introduction to the Transforms in this Package

• 5 Minute Read

Summary

• In this tutorial, we will briefly demonstrate some of the parameterized transforms that we provide in this package.

• For each torchvision-based transform, we provide one parameterized version with the same name that exposes the parameters of this transform.

• In addition, we provide useful tools and wrappers in order to ease working with parameterized transforms and porting to-and-from torch/torchvision-based code.

Transforms from torchvision but Parametrized!

• Please refer to the files transforms.py for more information while reading through this section.

• Note that torchvision.transforms.transform provides 35 transforms. Out of these, 31 are atomic transforms and the remaining 4 are composing transforms. All these transforms are listed below.

Atomic transforms from torchvision/transforms/transforms.py (31 transforms)

CenterCrop	ColorJitter	ConvertImageDtype	ElasticTransform	FiveCrop
GaussianBlur	Grayscale	Lambda	LinearTransformation	Normalize
PILToTensor	Pad	RandomAdjustSharpness	RandomAffine	RandomAutocontrast
RandomCrop	RandomEqualize	RandomErasing	RandomGrayscale	RandomHorizontalFlip
RandomInvert	RandomPerspective	RandomPosterize	RandomResizedCrop	RandomRotation
RandomSolarize	RandomVerticalFlip	Resize	TenCrop	ToPILImage
ToTensor

Composing transforms from torchvision/transforms/transforms.py (4 transforms)

Compose	RandomApply	RandomChoice	RandomOrder

• For each transform provided in torchvision/transforms/transforms.py, we provide its parameterized counterpart in parameterized_transforms/transforms.py. These parameterized transforms behave exactly identical to their torchvision counterparts in terms of the augmentation functionality but provide access to their augmentation parameters as well. All these transforms that are provided in parameterized_transforms/transforms.py are listed below–

Atomic transforms from parameterized_transforms/transforms.py (31 transforms)

CenterCrop	ColorJitter	ConvertImageDtype	ElasticTransform	FiveCrop
GaussianBlur	Grayscale	Lambda	LinearTransformation	Normalize
PILToTensor	Pad	RandomAdjustSharpness	RandomAffine	RandomAutocontrast
RandomCrop	RandomEqualize	RandomErasing	RandomGrayscale	RandomHorizontalFlip
RandomInvert	RandomPerspective	RandomPosterize	RandomResizedCrop	RandomRotation
RandomSolarize	RandomVerticalFlip	Resize	TenCrop	ToPILImage
ToTensor

Composing transforms from parameterized_transforms/transforms.py (4 transforms)

Compose	RandomApply	RandomChoice	RandomOrder

Useful Wrapper Transforms

• Please refer to the file wrappers.py for more information while reading through this section.

• In addition to these parameterized transforms, we provide extra wrapper transforms that allow for miscellaneous useful functionality. In particular, we provide 4 useful wrapper transforms which are listed below–

ExtractDefaultParams	ApplyDefaultParams	CastParamsToTensor	DropParams

• These transforms are mostly intended to ease manipulation of train/test augmentation stacks and for porting to-and-from torch/torchvision-based codebases. We briefly describe their usage below.

• ExtractDefaultParams wrapper transform stores an atomic or a composing transform as its core transform. When given an image img and parameters params as input, it fetches the default parameters of the core transform, appends them to the input parameters to obtain concatenated parameters concat_params, and returns the tuple of 1. the input image img as it is, and 2. the concatenated parameters concat_params. The purpose of this wrapper transform is to expose the default parameters of any transform of interest without actually applying that transform.

• ApplyDefaultParams wrapper transform stores an atomic or a composing transform as its core transform. It behaves identical to ExtractDefaultParams except for a subtle difference as suggested by its name. When given an image img and parameters params as input, it fetches the default parameters of the core transform, applies the core transform defined by these default parameters on the img to get the default-augmented image aug_img, appends the default parameters to the input parameters in order to obtain concatenated parameters concat_params, and returns the tuple of 1. the default augmented image aug_img, and 2. the concatenated parameters concat_params.

  • The reason to provide this subtly different wrapper transform is that there are some transforms where default parameters can preserve the image information but still change the image itself (for instance, ToTensor, ConvertImageDtype, Resize, etc.) and there are others where the information in the image is indeed lost ( for instance, CenterCrop, Grayscale, etc.). Given these possibilities, this wrapper is often helpful in easily generating the test transforms corresponding to the given train transforms; wrapping each component transform of a train transform stack often gives the desired test transform!

• CastParamsToTensor wrapper transform stores an atomic or a composing transform as its core transform. This transform is intended to wrap a finalized stack of core transforms, where the output parameter tuple is converted to a torch-tensor with dtype=torch.float32. In particular, when given an image img and parameters params as input, it converts the parameters to a torch-tensor params_torch with dtype=torch.float32 and returns the tuple of 1. the input image img as it is, and 2. the parameters converted to torch-tensor params_torch with dtype=torch.float32.

  • The utility of this wrapper transform will become clear in the tutorial on migrating code from torch/torchvision to use this package!

• DropParams wrapper transform stores an atomic or a composing transform as its core transform. This transform is intended to help convert a parameterized transform into its torchvision-counterpart to help port code from our package back to torch/torchvision. As the name suggests, given given an image img and parameters params as input, it just returns the image img as it is, thereby dropping parameters. Thus, wrapping any parametrize transform with this wrapper transform, we get its torchvision-counterpart!

• In the subsequent tutorials, we will illustrate some of the atomic, composing, and wrapper transforms provided in this package so that readers fill comfortable working with them!

About the Next Tutorial

• Till now, we have studied the core structure of parameterized transforms, we have writte our own custom parameterized atomic and composing transforms, and we have briefly gone through the transforms provided in this package.

• So, in the next tutorial 004-Parametrized-Transforms-in-Action, we will see the parameterized transforms in action; we will see how parameterized transforms implemented in the structure prescribed in this package allow us to perform cool things!

• If you were wondering why do we need to define our transform with the specific templates illustrated in the previous tutorials, this tutorial would answer your doubts by showing how these transforms would fit with each other seamlessly and with other transform provided in the package. This should also help demonstrate the power and the capabilities of the parameterized transforms!