Contrastive Prototypical Network with Wasserstein Confidence Penalty

PyTorch implementation of
Contrastive Prototypical Network with Wasserstein Confidence Penalty
Haoqing Wang, Zhi-hong Deng

ECCV 2022

Prerequisites

• Python >= 3.6
• Pytorch >= 1.2.0 and torchvision (https://pytorch.org/)

Datasets

For miniImageNet and tiredImageNet, download them from

• miniImageNet: https://drive.google.com/file/d/1MnYnUUTHX5KdMIyhIFSYnZ31kTzbGA0C/view?usp=sharing

• tiredImageNet: https://drive.google.com/file/d/10l0ev0TF9DjpMTLNzWoCI-nVNhDTpb0C/view?usp=sharing

and put them under their respective paths, e.g., ./Datasets/miniImagenet.

Training

Set method to BarTwins,SimCLR,BYOL,pn, cpn, cpn_cr, cpn_ls, cpn_cp, cpn_js or cpn_wcp for Barlow Twins, SimCLR, BYOL, CPN w/o Pairwise Contrast, CPN, CPN with Consistency Regularization, CPN with Label Smoothing, CPN with Confidence Penalty, CPN with Jensen–Shannon Confidence Penalty or CPN with Wasserstein Confidence Penalty respectively.

python train.py --dataset miniImagenet --backbone Conv4 --batch_size 64 --aug_num 4 --method cpn --alpha 0.1 --gamma 8 --name Exp_name

alpha is the label relaxation factor for Label Smoothing, gamma is the scaling factor for Wasserstein Confidence Penalty.

Evaluation

Set classifier to ProtoNet for prototype-based nearest-neighbor classifier and to R2D2 for ridge regression classifier.

python test.py --testset miniImagenet --backbone Conv4 --name Exp_name --classifier ProtoNet --n_way 5 --n_shot 5

Calibration

Set classifier to ProtoNet for prototype-based nearest-neighbor classifier and to R2D2 for ridge regression classifier.

python calibration.py --testset miniImagenet --backbone Conv4 --name Exp_name --classifier ProtoNet --n_way 5 --n_shot 5