nc_ps

Image classification with Neighborhood Clustering and Pseudo-Labeling

Requirements

Python 3.6.9, Pytorch 1.6.0, Torch Vision 0.7.0, Apex. We used the nvidia apex library for memory efficient high-speed training. You also need sklearn (0.23.2 is used).

Dataset Preparation

Office Dataset OfficeHome Dataset VisDA DomainNet

Prepare dataset in data directory.

./data/amazon/images/ ## Office
./data/dslr/images/ ## Office
./data/webcam/images/ ## Office
./data/Real ## OfficeHome
./data/Clipart ## OfficeHome
./data/Art ## OfficeHome
./data/Product ## OfficeHome
./data/DomainNet/real ## DomainNet real
./data/DomainNet/clipart ## DomainNet clipart
./data/visda/train ## VisDA synthetic images
./data/visda/validation ## VisDA real images

File list is stored in ./txt.

Training and evaluation

All training script is stored in scripts_exp directory.

sh scripts_exp/run_a2d_nc.sh $gpu-id

The script defines the search space of the hyper-parameter.