Skip to content

Conversation

@rentainhe
Copy link
Collaborator

@rentainhe rentainhe commented Feb 6, 2023

TODO

  • Support EMAHook as d2go.
  • Testing DINO-R50-4scale-12ep with ModelEMA: 49.4AP.

Note

If you set train.model_ema.enabled=True, after each eval_period, detrex will conduct two evalution on normal model and its ema version.

Usage

  • To enabled model ema training:
python tools/train_net.py --config-file projects/dino/configs/dino_r50_4scale_12ep.py \
                          --num-gpus 8 \
                          --resume \
                          train.model_ema.enabled=True \
  • Testing model with ema states when eval-only:
python tools/train_net.py --config-file projects/dino/configs/dino_r50_4scale_12ep.py \
                          --num-gpus 8 \
                          --resume \
                          train.model_ema.enabled=True \
                          train.model_ema.use_ema_weights_for_eval_only=True \

@rentainhe rentainhe changed the title Support EMAHook [Features] Support EMAHook Feb 6, 2023
@rentainhe rentainhe changed the title [Features] Support EMAHook [Features] Support model ema with EMAHook Feb 6, 2023
@rentainhe rentainhe requested a review from SlongLiu February 6, 2023 03:24
@rentainhe rentainhe changed the title [Features] Support model ema with EMAHook [Feature] Support model ema with EMAHook Feb 7, 2023
@rentainhe rentainhe merged commit 6ebb36d into main Feb 7, 2023
@rentainhe rentainhe deleted the support_model_ema branch February 7, 2023 17:36
Lontoone pushed a commit to Lontoone/detrex that referenced this pull request Jan 8, 2024
* add ema hook

* refine ema usage

* refine code

* update ema links
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants