Image augmentation in dataset

[antimora]

Feature description

Similar to Pytorch's vision utility, it'd be great if burn provides with interface and a few useful image augmentation utilities, such as random image rotation.

Feature motivation

Instead of having users coming up with their own utils, it would be beneficial for all if Burn's dataset supports data augmentation.

(Optional) Suggest a Solution

For image augmentation we could use these image backends: 1) https://lib.rs/gh/imazen/imageflow/imageflow_core or 2)https://github.com/image-rs/imageproc

[laggui]

Just leaving this here for future reference.

For some reason imageflow is not available on crates.io (ref issue).

I was just browsing the other possible crates and I found photon (extends image and imageproc crates). It already seems to implement different image transformations. Also supports WASM.

Also found zune-image with zune-imageprocs, which seems to have a focus on performance and support for different image types (u8, u16 and f32), as opposed to photon which at first glance seems to currently work on RGBA images only.

[ronofays]

Hello, I have been working on a first pass attempt for getting some vision utility with burn tensors.

I'm somewhat new to the open source space, so I am not sure whether what I have done so far warrants a PR, but you can find it in my fork of burn, under burn-vision: https://github.com/ronofays/burn/tree/burn-vision/burn-vision

Currently, I have three things done:

• read rgba images as tensors (omitting alpha channel for now)
• converting tensors back to rgba images
• padding for tensors with 3 color channels

read/write is in io.rs
padding is in transforms.rs

Obviously a lot needs to be done to handle images that are not u8 rgba images, but I would be thrilled to have some feedback on the start that I have here, particularly if there are some code style things for burn that I am missing.

Thanks!

[antimora]

@ronofays Thanks for sharing your work! Yes, any addition in this space would be awesome.

I suggest you create a draft PR and submit to the project. It's easier to see your new additions and we can context comment.

[antimora]

Linking related ticket: #2361