Cross-Domain Cross-Set Few-Shot Learning via Learning Compact and Aligned Representations

This is the PyTorch implementation of stabPA.

Prerequisites

We run this code with Ubuntu 16.04.5 LTS, Python 3.7 and Pytorch 1.4.0. qpth is also needed if you want to test MetaOptNet.

Data preparation

DomainNet

• Download real.zip, painting.zip, clipart.zip and sketch.zip from http://ai.bu.edu/M3SDA/
• unzip data to dataset/DomainNet
• cd dataset/DomainNet and run split.py

Office-Home

• Download OfficeHomeDataset_10072016.zip from https://www.hemanthdv.org/officeHomeDataset.html
• unzip data to dataset/Office-Home
• cd dataset/Office-Home and run split.py

Pretrain the classifier on the source data

python main.py --gpu 0 --dataset DomainNet --cross_domain painting --exp baseline

This will automatically test the 5-way 5-shot accuracy on the validation set of painting-real.

Train

python main.py --gpu 0 --dataset DomainNet --cross_domain painting --exp stabPA --st_align --ts_align --aug_type strong --init_teacher checkpoints/DomainNet/painting/ResNet18_baseline/baseline/50.tar

Test

python main.py --gpu 0 --dataset DomainNet --cross_domain painting --split novel --test --n_shot 1 --exp test --st_align --ts_align --aug_type strong --resume --checkpoint checkpoints/DomainNet/painting/ResNet18_baseline/stabPA/
python main.py --gpu 0 --dataset DomainNet --cross_domain painting --split novel --test --n_shot 5 --exp test --st_align --ts_align --aug_type strong --resume --checkpoint checkpoints/DomainNet/painting/ResNet18_baseline/stabPA/
python main.py --gpu 0 --dataset DomainNet --cross_domain painting --split novel --test --n_shot 1 --reverse_sq --exp test --st_align --ts_align --aug_type strong --resume --checkpoint checkpoints/DomainNet/painting/ResNet18_baseline/stabPA/
python main.py --gpu 0 --dataset DomainNet --cross_domain painting --split novel --test --n_shot 5 --reverse_sq --exp test --st_align --ts_align --aug_type strong --resume --checkpoint checkpoints/DomainNet/painting/ResNet18_baseline/stabPA/

Evaluate other methods

ProtoNet

python main.py --gpu 0 --dataset DomainNet --cross_domain painting --exp protonet --method protonet

RelationNet

python main.py --gpu 0 --dataset DomainNet --cross_domain painting --exp relationnet --method relationnet_softmax

MetaOptNet

python main.py --gpu 0 --dataset DomainNet --cross_domain painting --exp metaoptnet --method metaoptnet

DeepEMD

python main.py --gpu 0 --dataset DomainNet --cross_domain painting --exp deepemd --method deepemd --optimizer sgd --lr 5e-4 --init_model checkpoints/DomainNet/painting/ResNet18_baseline/baseline/50.tar

References

This code is built on

• CloserLookFewShot: https://github.com/wyharveychen/CloserLookFewShot
• FixMatch-pytorch: https://github.com/kekmodel/FixMatch-pytorch
• rfs: https://github.com/WangYueFt/rfs
• MetaOptNet: https://github.com/kjunelee/MetaOptNet
• DeepEMD: https://github.com/icoz69/DeepEMD