Training from scratch on COCO

[factplay1]

Thanks for the good work!

Just wanted to mention that I have tried the two most stared EfficientDet PyTorch repos on Github, and neither reproduce the paper results on COCO, not even close.

They mostly claim they train on custom data and port weights from official TF checkpoints, but fail to train from scratch on COCO.

I tried their implementations with 200+ kinds of hyper-parameter tuning sets & settings - yes 200+ jobs!

Very keen to see your completed training on COCO. Looking forward to that.

Cheers

[sevakon]

Thank you very much for your words! It really motivates me.
I will let you know when everything is ready by commenting on your issue.

Just wondering, what hardware specs do you have? Maybe we could collaborate to obtain checkpoints trained on COCO from scratch on all the models.

Best,
Seva

[factplay1]

Glad to hear that.
Thank you :)

I have access to cloud and a dozen local GPUs - can run jobs say on 8 X V100 GPUs or lower.
I would be happy to run training jobs, can't however send out any files (corporate limitations). But will be able to write you which set of hyper-parameters lead to what mAP...

[tchaton]

Hey @sevakon @factplay1,

We are going to use this code to create an on boarding tutorial within my company to put it on production.
Thanks for the hard work and clean code.

Best regards,
Thomas Chaton.

[sevakon]

@tchaton Excited to hear that!
Please let me know how this one goes!
4 months have passed since I worked on the project, and there are some things that I would have done differently now.
For example, PyTorch-Lihgtning would make the training code much more organized and add seamless multi-GPU and TPU support.

[tchaton]

Dear @sevakon,

We are using Pytorch Lightning and it will be integrated within our PL stack :)

By the way, I will be joining PL next week :)

Best T.C

[tchaton]

Dear @sevakon,

You should switch to Hydra :)

Best regards,
T.C