This repository contains a PyTorch implementation of No Fuss Distance Metric Learning using Proxies as introduced by Google Research.

The same parameters were used as described in the paper, except for the optimizer. In particular, the size of the embedding and batches equals 64 and 32 respectively. Also, BN-Inception is used and trained with random resized crop and horizontal flip and evaluated with resized center crop.

I have ported the PyTorch BN-Inception model from PyTorch 0.2 to 0.4. It's weights are stored inside the repository in the directory net.

You need Python3 and minimum PyTorch 0.4.1 to run the code.

If you want to reproduce the results in the table below, then the only thing you have to do is to execute: python3 train.py.

In this case, the CUB dataset will be automatically downloaded to the directory cub200 (default) and verified with the corresponding md5 hash. If you train the model for the first time, then the images file will be extracted automatically in the same folder. After that you can use the argument --cub-is-extracted to avoid extracting the dataset over and over again.

Training takes about 10 minutes with one Titan X (Pascal).

An example training log file can be found in the log dir, see example.log.