Efficient end-to-end learning for quantizable representations

This repository has the source code for the paper "Efficient end-to-end learning for quantizable representations"(ICML18).

Citing this work

@inproceedings{jeongICML18,
    title={Efficient end-to-end learning for quantizable representations},
    author={Jeong, Yeonwoo and Song, Hyun Oh},
    booktitle={International Conference on Machine Learning (ICML)},
    year={2018}
    }

Installation

• Python3.5
• Deep learning frame work : Tensorflow1.4 gpu Check https://github.com/tensorflow/tensorflow/tree/r1.4
• Ortools(6.6.4656) Check https://developers.google.com/optimization/introduction/download

Prerequisites

1. Make Directory for data and experiment

cd RROOT
mkdir dataset deep_hash_table_processed deep_hash_table_exp_results
mkdir dataset/Imagenet32

2. Change path in config/path.py

RROOT = '(user enter path)'
EXP_PATH = RROOT+'deep_hash_table_exp_results/'
#=============CIFAR100============================#
CIFAR100PATH = RROOT+'dataset/cifar-100-python/'
CIFARPROCESSED = RROOT+'deep_hash_table_processed/cifar_processed/'
#==========================Imagenet32===============================#
IMAGENET32PATH = RROOT+'dataset/Imagenet32/'
IMAGENET32PROCESSED = RROOT+'deep_hash_table_processed/Imagenet32_processed/'

3. Download and unzip dataset Cifar-100 and Downsampled imagenet(32x32)

cd RROOT/dataset
wget https://www.cs.toronto.edu/~kriz/cifar-100-python.tar.gz
tar zxvf cifar-100-python.tar.gz

cd RROOT/dataset/Imagenet32
wget http://www.image-net.org/image/downsample/Imagenet32_train.zip
wget http://www.image-net.org/image/downsample/Imagenet32_val.zip
unzip Imagenet32_train.zip
unzip Imagenet32_val.zip

Processing Data

cd process
python cifar_process.py
python imagenet32_process.py 

Training Procedure

• Cifar-100 experiment(cifar_exps/) and ImageNet experiment(imagenet_exps/).

1. Training metric(metric/)
   • train_metric.py is to train embedding with metric learning losses.
   • test_metric.py is to test the embedding with the hash codes built with vector quantization method(VQ) and thresholding method(Th).
2. Training hash codes(exp1/)
   • train_hash.py is to replace the last layer and fine tune the embedding with the proposed method in paper.
   • test_hash.py is to test the hash codes built with the embedding trained from train_hash.py.

Evaluation

• Evaluation code is in utils/evaluation.py.
• The hash table built with hash code is evaluated with 3 different metric(NMI, precision@k, SUF).

Ortools

• The code to solve the dicrete optimization problem in polynomial time is in utils/ortools_op.py
• The time to solve the discrete optimization problem is calculated with the code ortools_exp/

License

MIT License