MobileNets

[FabioNaecht]

Thanks for sharing this repo, @n-poulsen !
Would it be possible to add an option to train MobileNets using the DLC PyTorch version and deploy them in the DLC-live PyTorch version, similar to the TensorFlow versions?

[n-poulsen]

Hi @FabioNaecht - MobileNets can be added relatively easily in DLC3.0. We don't have them available quite yet, but I would try the CSPNeXt backbones instead. They're very fast and produce good results. I just add the backbone to the branch that is being worked on.

You can train a CSPNeXt model by settings net_type="cspnext_s", net_type="cspnext_m" or net_type="cspnext_x" when creating the training dataset. I'd recommend taking a look at those models before trying a MobileNet :)

Also the docs are a WIP - but I'm working on improving things now that everything is working (also don't hesitate to report bugs if you find any!)

[FabioNaecht]

Thanks a lot! :) I’ll give it a try next week—excited to see how CSPNeXt performs!

[n-poulsen]

Awesome @FabioNaecht! Let me know if you run into any issues I'd be glad to help!

[n-poulsen]

@FabioNaecht I opened #121 in DeepLabCut-live which is better documented and already has some improvements for DLC-Live with PyTorch models. I'd recommend giving that a go! Let me know if you have any questions or if you run into any issues!

[FabioNaecht]

cool, thanks for the update @n-poulsen !