ProxyNCA++: Revisiting and Revitalizing Proxy Neighborhood Component Analysis

This repo consists of the source code for the ProxyNCA++ paper

Make sure to download the corresponding dataset to the correct folder as specified in dataset/config.json We also include script to convert the dataset to hdf5 format.

To create data

> conda activate pytorch_p36

# create Shoes_train & Shoes_test files. This command outputs TRAIN_DATA_SIZE & TEST_DATA_SIZE
> python data_to_txt.py --path DATA_DIR

# create hdf5 file
> python dataset/make_mixed_hdf5.py --nb_train_all TRAIN_DATA_SIZE --nb_test_all TEST_DATA_SIZE --source DATA_DIR --output OUTPUT_PATH

To start training

nb_train_all: 106974 nb_test_all: 106550

> conda activate pytorch_p36

# SOP
CUDA_VISIBLE_DEVICES=0,1 python train.py --dataset sop  --config config/sop.json --mode train --apex --seed 0
CUDA_VISIBLE_DEVICES=0,1 python train.py --dataset sop  --config config/sop.json --mode trainval --apex --seed 0

# MIXED
> python train.py --dataset mixed_h5  --config config/mixed.json --mode train --seed 0
> python train.py --dataset mixed_h5  --config config/mixed.json --mode trainval --seed 0

The following is the Bibtex of our paper:

@article{teh2020proxynca++,
  title={ProxyNCA++: Revisiting and Revitalizing Proxy Neighborhood Component Analysis},
  author={Teh, Eu Wern and DeVries, Terrance and Taylor, Graham W},
  journal={arXiv preprint arXiv:2004.01113},
  year={2020}
}