search

qubvel-org / segmentation_models.pytorch

Semantic segmentation models with 500+ pretrained convolutional and transformer-based backbones.
https://smp.readthedocs.io/
MIT License
9.5k stars 1.66k forks source link

Best way to integrate with fastai? #603

Closed jph00 closed 2 years ago

jph00 commented 2 years ago

@qubvel -- we've been chatting on the fast.ai Discord channel about how cool this lib is, and how much our community loves it! I'm wondering whether there might be ways to integrate it into fastai.

I'm not sure quite the best approach (or whether it's doable at all) but I'd be interested in exploring it. There's quite a bit of overlapping functionality between our libraries at the moment. We also integrate with timm, like you do, but currently we only do so for classification models, not for segmentation. fastai was (AFAIK) the first library to have a dynamic unet that supports a variety of pretrained backbones -- but you've taken it much further! I like the way you've got separate encoder and decoder modules, and also support various metrics and loss functions.

Anywho, I don't have any particularly concrete ideas - just wanted to raise the idea and see if you have any thoughts.

github-actions[bot] commented 2 years ago

This issue is stale because it has been open 60 days with no activity. Remove stale label or comment or this will be closed in 7 days.

jph00 commented 2 years ago

@qubvel -- just checking whether you had any thoughts on this?

qubvel commented 2 years ago

Hi @jph00

Thanks for connecting! I am very sorry for not answering before, the project has not been maintained for a while and I had some other vital issues.

I heard a lot about fastai, and it is widely used on Kaggle, but unfortunately never used it by myself. I guess we can explore the possibility of integration, it can be a nice feature for both libraries! If you have some thoughts about how to do it I am open to this! On my side, I can provide additional API for model creation if you have some standard one. Looking forward to hearing back from you, thanks!

github-actions[bot] commented 2 years ago

This issue is stale because it has been open 60 days with no activity. Remove stale label or comment or this will be closed in 7 days.

github-actions[bot] commented 2 years ago

This issue was closed because it has been stalled for 7 days with no activity.

jph00 commented 2 years ago

Message ID: @.***>No need to apologise. I hope things are OK with you. Since this isn't maintained any more, probably the best thing for us to do is to look to it for inspiration, and incorporate some of those ideas into fastai if they are a good it.