torchvision

1. How is the C++ code compiled and how is it loaded in PyTorch.

from torchvision.io import read_file to read a binary file and it returns a 1-d tensor of dtype torch.uint8.

There is also a write_file to write a 1-d tensor of dtype torch.uint8.

from torchvision.io import read_image to read an image. Supported formats are jpg, png, gif.

Color images are represented as a uint8 tensor of shape (channels, height, width).

transforms.v2._utils contains functions for checking and parsing sizes, paddings, fill.

Use v2.ToDtype(torch.float32, scale=True) to convert dtype.

• RandomAffine calls F.affine. RandomAffine is in transforms/v2/_geometry.py. Affine has 4 parameters: angle for rotation, translate, scale, shear. RandomAffine._get_params() just return the 4 parameters. Note that we can call the class method RandomAffine.get_params(), which is defined in transform/transform.py

• F.affine() is defined in transforms/v2/functional/_geometry.py with name def affine_image _get_inverse_affine_matrix() is used to get the matrix for affine transform.

  Inside F.affine(), it calls _affine_grid to get a grid from the affine transform matrix and then use _apply_grid_transform to transform the input.

TODO: look at torch.nn.functional.affine_grid() and grid_sample(). See also https://github.com/wuneng/WarpAffine2GridSample/blob/master/main.py#L56

Deep Learning Paper Implementations: Spatial Transformer Networks - Part I