Is there a master list of features/improvements that need implementing?

[brendanartley]

As above :)

[rwightman]

@brendanartley

I probably should keep a better list of near / far goals, but change my mind a lot as I tug on various threads until I find some momentum on one or two.

Best bet would be open issues. I'm generally pretty quick to close issues or PR's that I don't feel have merit, so that said, any open PR or Issue (especially [FEATURE] issues) could be worthwhile. I have left some for quite some time though, either because they weren't highest priority, still issues to sort out or things to verify in impl (PR), or significant undertaking to get right so effort vs reward not clear...

Some people have asked for better documentation, the hfdocs is current (docs are deprecated, should remove that soon) so anything added there will go straight up to https://huggingface.co/docs/timm/en/index

Some of the bigger ideas I still would like to tackle:

• object detection, segmentation (integrate a task wrapper for train loops, necks, heads, losses, new dataset readers) but a focus on best single stage and avoid 2-stage. I'd probably start with variants of YOLO v8/v9/v10 ideas, esp anchor / nms free and also DETR variants, esp the faster real-time focused ones.
• still open to video but the data handling has always been a bit of a grind with open source video decode/processing libs
• multi-task learning (n-heads for classification), supervised + self-supervised combo ... so associated loss + dataset muxing and train loop mods
• distillation training (cover a few core options) with an associated task wrapper or train_dist script

I've actually started sketching out the object detection idea recently. I got inspired having a chat with Claude Sonnet, pretty much sketched out my scaffolding and decided on a few design points...

[brendanartley]

Thanks @rwightman for the ideas.

Are the .yaml configs for a few of the popular backbones (resnet18, efficientnetB0, etc.) stored anywhere?

I would like test out some of these ideas locally w/ a sample of imagenet. Sorry if this is a silly question, I have not used the library to train from scratch before..