Torchvision

Torchvision package

PyTorch has torchvision package designed to prepare visual images for learning process.

The torchvision package in turn has additional subpackages.

datasets 
models 
transforms 
utils

The torchvision.datasets subpackage contains most important datasets. At the current moment these are:

cifar
cityscapes
coco
fakedata
flickr
folder
lsun
mnist
omniglot
phototour
sbu
semeion
stl10
svhn
utils
voc

The torchvision.models subpackage contains these models at the current moment:

alexnet
densenet
inception
resnet
squeezenet
vgg

The torchvision.utils help us save Tensors to a file. These tensors are of shape:

 BxCxHxW : number of mini batches, channels, height, width

and create grids of images.

But the most interesting sub-package today is the torchvision.transforms package. This package has exactly two sub pakages torchvision.transforms.functional and torchvision.transforms.transforms that holds the classes behind the torchvision.transforms.functional methods.

The torchvision.transforms.functional package depends on PIL.Image functionality. Contains methods to detect the image type:

_is_numpy_image
_is_pil_image
_is_tensor_image

Methods to adjust the image:

adjust_brightness
adjust_contrast
adjust_gamma
adjust_hue
adjust_saturation

Methods to transform the image

affine (keeps the center in place)
center_crop    
crop    
five_crop
hflip
pad
resize
resized_crop
rotate
scale
ten_crop
vflip

Some handy methods to convert the image:

to_grayscale
to_pil_image
to_tensor

And also the method to normalize the image.

normalize