Proto2Proto [arxiv]

To appear in CVPR 2022

Creating conda Environment

conda env create -f environment.yml -n myenv python=3.6
conda activate myenv

Preparing Dataset

• Refer https://github.com/M-Nauta/ProtoTree to download and preprocess cars dataset
• For augmentation, run lib/protopnet/cars_augment.py (Change the dataset paths if required)
• Create a symbolic link to the dataset folder as datasets
• We need the dataset paths as follows

  trainDir: datasets/cars/train_augmented # Path-to-dataset
  projectDir: datasets/cars/train # Path-to-dataset
  testDir: datasets/cars/test # Path-to-dataset

Training

sh train_teacher.sh  # For teacher training
sh train_baseline.sh # For baseline student training
sh train_kd.sh       # For proto2proto student training

NOTE: For proto2proto student training, set the teacher path in Experiments/Resnet50_18_cars/kd_Resnet50_18/args.yaml: backbone.loadPath. Use the teacher model trained previously. For eg.

loadPath: Experiments/Resnet50_18_cars/teacher_Resnet50/org/models/protopnet_xyz.pth

Evaluation

Set model paths in Experiments/Resnet50_18_cars/eval_setting/args.yaml: Teacherbackbone.loadPath, StudentBaselinebackbone.loadPath, StudentKDbackbone.loadPath. And Run

sh eval_setting.sh

Things to remember

• Dataset path should be set appropriately
• Model path should be set in KD (1 place) and eval setting (3 places)
• Set CUDA_VISIBLE_DEVICES depending on the GPUs, change batchSize if required

Acknowledgement

Our code base is build on top of ProtoPNet

Citation

If you use our work in your research please cite us:

@inproceedings{Keswani2022Proto2ProtoCY,
  title={Proto2Proto: Can you recognize the car, the way I do?},
  author={Monish Keswani and Sriranjani Ramakrishnan and Nishant Reddy and Vineeth N. Balasubramanian},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2022)},
  eprint={2204.11830},
  archivePrefix={arXiv},
  year={2022}
}