Provide generic decode/read image io functions

[vfdev-5]

🚀 Feature

Currently, torchvision.io.image module provides read_png/jpeg and decode_png/jpeg separate functions to read and decode images from files. They are not and will not be exposed publicly and the idea is to provide two generic methods like
read and decode in order to handle both formats (png and jpeg): call internally read_png or read_jpeg

Maybe, we can inspire of how does the libraries like PIL to dispatch according to the input image format.

[oke-aditya]

This is PIL Implementation fo Image.open() This works for all imags and is generic.
Source code is here. Line number 2841.

Since PIL is well maintained and generic for all image types, including alpha channel and it is a requirement for torch and torchvision, can we provide a wrapper to read with torchvision using PIL internally?

I'm not sure if I can implement this though.

[vfdev-5]

@oke-aditya thanks for the info, but I'm not sure if we would like to use PIL for reading and decoding the data. Idea is see how PIL manages its png/jpeg specific readers and try to do something similar here in order to dispatch to our read_png, read_jpeg method

[oke-aditya]

It might be better as you say, allowing native support through tochvision with that.

What about C++, I would like to have the same in Libtorch.

[oke-aditya]

@QuantScientist I guess as vfdev-5 said this will be implemented in C++ only.

[fmassa]

@QuantScientist we already have C++ implementation for image decoding functions.

@oke-aditya about using PIL, the idea is being able to have a C++-only implementation that works with torchscript, and PIL's C implementation is very tied to Python.

From the PIL implementation, it seems they compute a prefix, which is used to quickly check if one format could be used for not.

Here is the implementation for jpeg and for png for reference. Both formats seems to have a magic number header that could be used to perform the dispatching (although further investigation would be useful to see if there are no corner cases, as in PIL they have a try: except while loading).

See here:
#2691

So I wrote a small lib for it:
https://github.com/QuantScientist/PngTorch