This repository is an official pytorch implementation of the paper Revisiting the Critical Factors of Augmentation-Invariant Representation Learning (ECCV2022).

We release our framework to the public for the good of reproducibility. We hope it would be helpful for the community to develop new frameworks based on the fair benchmarking settings.

Notation:

1. The models are trained on ResNet-50 for 200 epochs.
2. We release the models with the best linear evaluation results, and the detailed hyper-parameters are packed into the checkpoint files.

First, clone the repository locally:

git clone https://github.com/megvii-research/revisitAIRL.git

Second, install the dependencies:

pip install -r requirements.txt

For example, to train S-MoCov2+ on a single node with 8 GPUs:

python train.py -f exp/moco_based/mcv2p_symmlp.py --output-dir path/to/save -expn S-MoCov2p --data-path path/to/imagenet --total-epochs 200 --batch-size 256

Then, the process will create a directory path/to/save/S-MoCov2p to save all logs and checkpoints.

If you want to train a supervised model, please add the tag --single-aug.

To evaluate a model trained with SGD optimizer on a single node with 8 GPUs:

python linear_eval.py --output-dir path/to/save -expn S-MoCov2p --data-path path/to/imagenet

Then, the process will create a directory path/to/save/S-MoCov2p/linear_eval to save all logs and checkpoints.

If you want to evaluate a model trained with LARS optimizer, please add the tag --large-norm-config.

First, please install detectron2 following detectron2 installation.

Second, convert the pre-trained model into a pickle format:

python convert-pretrain-to-detectron2.py path/to/pretrain_ckpt.pth.tar path/to/save/pickle_file.pkl [online/target/supervised]	

Then, a pickle file will be generated at path/to/save/pickle_file.pkl.

For example, to transfer the pre-trained model to the VOC07 object detection task on a single node with 8 GPUs:

python det/train_net.py  --config-file det/pascal_voc_R_50_C4_24k_moco.yaml --num-gpus 8 MODEL.WEIGHTS path/to/pickle_file.pkl OUTPUT_DIR path/to/save/voc07

Then, the process will create a directory path/to/save/voc07 to save all logs and checkpoints.

We provide a simple demo in norm_rescale.py, please modify and use it as you like.

Part of the code is adapted from previous works:

MoCo

We thank all the authors for their awesome repos.

If you find this project useful for your research, please consider citing the paper.

@misc{huang2022revisiting,
      title={Revisiting the Critical Factors of Augmentation-Invariant Representation Learning}, 
      author={Junqiang Huang and Xiangwen Kong and Xiangyu Zhang},
      year={2022},
      eprint={2208.00275},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

If you have any questions, feel free to open an issue or contact us at [email protected].