Supernet training and tuning

[pawopawo]

Thanks for your excellent work!

I have not found how to pretrain supernet in ImageNet and tuning supernet in COCO, here is no supernet training, only the training of the searched network.

[yukang2017]

Hi,
for training the supernet on COCO, take 300M FPN models for example,
you can just run
bash scripts/run_detnas_coco_fpn_300M_search.sh
and refer to configs/e2e_faster_rcnn_DETNAS_COCO_FPN_300M_search.yaml for details.
(The '_search' tail means for supernet training.)

For training the supernet on ImageNet, I would pull request to my colleague who is in charge of the ImageNet project. It will not cost too long. Thanks!

[yukang2017]

Hi, for training the supernet on ImageNet, you can find it here. This part is inherited from the ImageNet project, where you can find more details.

Please feel free to reopen the issue if there are any other questions.

[uname0x96]

@yukang2017 Can I train with my custom dataset? And I just have only 1 or maximum is 2 cards RTX-2080Ti so it enough for the train?