Guidance for Multi-label classification

[davidrs]

Hello, love the library!

I'm trying to train an image classifier with support for multilabel.

I found this thread from 3+ years ago, and @rwightman alludes to maybe adding support.
#509

• Is multilabel support baked in and I missed it?
• Any recommendations for easy recipes to follow apart from the repo in the old thread?

[rwightman]

@davidrs so yeah, that sorta stayed in limbo, biggest issue was a lack of standardization in terms of multi-label datasets to use as canonical examples / benchmarks. I've actually just started to address this on the data side

• https://huggingface.co/datasets/timm/plant-pathology-2021
• https://huggingface.co/posts/rwightman/673896178494185

So, the plan is to now finally support other tasks, like multi-label, also multi-task (multiple multi-label and/or multi-class heads), AND finally support object detection, segmentation tasks. It's a fair bit to take on but I'm making headway, you'll probably see a bit more on the data handling side (annotation reading & transform to inputs & targets) before I get final train & modelling code sorted as I need to data part solid to refine & test the rest...

[rwightman]

Oh yeah, and other tasks require more metrics, so will be work needed there too for meaningful multi-label support, and the other proposed tasks require more involved metrics too.

[davidrs]

thank you, I appreciate the update and completely get what you mean about lack of standardization on multi-label datasets.

for anyone else trying to roll-their-own in the meantime the previously linked repo and article is a helpful jumping off point:

• https://learnopencv.com/multi-label-image-classification-with-pytorch/
• https://github.com/yang-ruixin/PyTorch-Image-Models-Multi-Label-Classification