neka-nat / pytorch-hdml

Hardness-Aware Deep Metric Learning (CVPR2019) in pytorch

Hardness-Aware Deep Metric Learning

This is an unofficial implementation of "Hardness-Aware Deep Metric Learning" (CVPR 2019 Oral) in Pytorch.

Installation

cd pytorch-hdml
pip install pipenv
pipenv install

Download dataset

cd data
python cars196_downloader.py
python cars196_converter.py

Train CARS196 dataset

Execute a training script. When executed, the tensorboard log is saved.

pipenv shell
python train_triplet.py

Result triplet HDML

CARS196 result on training(99 classes, 30000 iterations)

Loss

t-SNE

CARS196 result on testing(97 classes)

t-SNE

Todo

• Implementation of Npair loss HDML

Reference

Official tensorflow implementation https://github.com/wzzheng/HDML