Yet another TensorFlow with Keras reimplementation of the paper No Fuss Distance Metric Learning using Proxies as introduced by Google Research.

Take advantage of the a little recent network VGG19 with somewhat naive optimizer SGD. These two parts could be easily changed to any other compatible components.

Provided two training processes. One is direct and the other applies some data augmentation tricks. Both produce applicable results.

The dataset applied is Cars 196. But it could be easily changed to any other datasets under Keras framework.

The anaconda environment to run is provided as the environment.yaml file.

A custom version of Keras is included which is necessary for data augmentation.

Do the following:

git clone https://github.com/terU3760/tensorflow_proxy-nca.git

In the cloned project root directory:

mkdir data
cd data

Download cars_data.tar.xz (slightly modified version of Stanford University CARS196 dataset) into this directory, and then type

tar -xf cars_data.tar.xz

Edit file config_cars_tensorflow.json, let "root" point to your "cars_data" directory.

To train directly, in the root directory if the cloned project, type:

python train.py

To train with data augmentation, type:

python train_data_augmentation.py

Since it is run under Keras framework, no need to mind the validation loss, a custom evaluation process is adopted.

Since using a different network VGG19 instead of BN-Inception, the result is not listed here. But the base network could be very easily converted back into it.