Training code?

[xianshunw]

Excellent work! Thanks for your contribution! Will the training code be available later? How many computing resources are needed to reproduce your reported results?

[achalddave]

The training code is technically available, but just not documented. I will try to write up some documentation in the next week.

All of my models were trained on 4x 1080Tis. Because of the stagewise training, the full training process is a bit slow: The objectness model on COCO takes ~2 days, which is the slowest part of training, and I think everything else took ~1 day.

Are you looking to re-train this model soon? If you are interested in one particular aspect of training, I will try to write the docs for that first.

[xianshunw]

Thanks for your reply. The objectness model is the part I'm most interested in.

[FlyingGaifield]

@achalddave What a brilliant work! I'm also very concerned with the training docs! Thanks!

[achalddave]

Re-opening this until I can get the training docs up... Sorry for the delay :(

[jinpkj007]

Great Work !!! Are you planning to upload the training docs soon? Really looking forward to it specially the motion stream.

[alohagin]

Brilliant work! Looking forward to re-train this model too. Please release training docs at your convenience.

[achalddave]

Hi all, sorry again for the delays here. I've documented the commands for training here. Of course, I'd like to flesh these out to more complete docs including how to set up links to the right datasets, but hopefully this initial bit is helpful to someone: achalddave/segment-any-moving_detectron#1

[ChenjieWang]

Excellent work! Thanks for your contribution! Will the training code be available later? How many computing resources are needed to reproduce your reported results?

Did you run the joint model successfully? What is the configuration environment you use, including CUDA and python versions. I hope I can get your help.

[xianshunw]

Excellent work! Thanks for your contribution! Will the training code be available later? How many computing resources are needed to reproduce your reported results?

Did you run the joint model successfully? What is the configuration environment you use, including CUDA and python versions. I hope I can get your help.

The inference phase works for me. But I have never tried their training code.