Coco dataset , fix for grayscales images, convert them to RGB

[ExtReMLapin]

No description provided.

[Copilot]

Pull Request Overview

This PR fixes the handling of grayscale images in the COCO dataset by converting them to 3-channel RGB before further processing.

• Added conversion of grayscale images to RGB by duplicating the single channel.
• Ensured that images with an alpha channel are reduced to RGB.

[tlpss]

@ExtReMLapin thanks for your contribution, looks like a useful addition! I'll merge it soon.

[ExtReMLapin]

🖖🏻

A pleasure.
I also have more changes in staging for annother PR which adds max_image_size param
I'm training it on forensic images that have different resolutions and often high ones which causes :

1. OOM during training (because of big resolution)
2. Error during validation because of torch.stack trying to stack up different sizes.

[tlpss]

Hi @ExtReMLapin

Sounds like an interesting project.

I consider these steps part of the preprocessing to reduce the burden on the ML codebase (can't support everything in the training loops) and to increase data loading speeds (loading a huge image from disk and then resizing it can bottleneck the GPU because it has to wait on the CPU, which is not desirable).

I will probably not accept a PR that does image resizing in the dataloader (as a separation of concerns).

You should consider resizing the images upfront into a separate dataset and only then training a detector on them.

I have some code for this here if you are interested.

[tlpss]

@ExtReMLapin can you take a look at the CI failures? apparently one of the tests was broken by an update in torch ,but the fix should be straightforward.

Btw, I'm on a conference next week so will take some time for me to get back to you! But I do appreciate the PRs 🙂

[ExtReMLapin]

No worry with the delay.

To be frank i've been working on this forensic minutiae detector for two years and you have no idea how sometimes it's a pain in the ass to :

• set up the whole repository env
• transform your dataset
• discover their undocumented training examples are not working

here it's just working with wandb integration, few issues with DDP but it's fine tbf