---

---

Matlab demo code for UH-BDNN

If you use/adapt our code, please appropriately cite our ECCV 2016 paper: "Learning to Hash with Binary Deep Neural Network", Thanh-Toan Do, Anh-Dzung Doan and Ngai-Man Cheung

This code is for academic purpose only. Not for commercial/industrial activities.

---

---

I. PREREQUISITES

1. A working version of Matlab/mex.
2. A working version of the Yael library (http://yael.gforge.inria.fr)
3. A 3rd party implementation of L-BFGS (subdirectory minFunc included in http://ufldl.stanford.edu/wiki/resources/sparseae_exercise.zip)

We have included, compiled and tested all 3rd party libraries on MATLAB R2014a, OS Ubuntu 14.04 LTS 64-bit

II. DATASET

Download dataset.zip here, extract and place it in the directory of source code

The folder './dataset' contains the mat files used for this demo code + cifar-10/cifar_gist_320.mat: This is GIST 320-D representation of CIFAR-10, it contains: - Xtest (10000x320): 10K query images - Xtrain(50000x320): 50K gallery images - Ytest (10000x1): semantic label of 10K query images (we do not use in unsupervised case) - Ytrain(50000x1): semantic label of 50K gallery images (we do not use in unsupervised case) + cifar-10/groundtruth.mat: groundtruth of CIFAR-10 based on Euclidean distance

	+ mnist/mnist_train.mat: This is 784-D gray-scale training images of MNIST, it contains:
			- Xtrain (784x60000): 60K gallery images
			- Ytrain (60000x1): semantic label of 60K gallery images (we do not use in unsupervised case)
	+ mnist/mnist_test.mat: This is 784-D gray-scale query images of MNIST, it contains:
			- Xtest (784x10000): 10K query images
			- Ytest (10000x1): semantic label of 10K query images (we do not use in unsupervised case)
    + mnist/groundtruth.mat: groundtruth of MNIST based on Euclidean distance

III. USAGE

Run 'demo.m', it will visualize a comparison between our method with ITQ in mAP Please test with MNIST dataset first. If the code works properly, you will get Fig 2(b) (with 2 curves: our UH-BDNN and ITQ) as our ECCV16 paper It should take 3-4 hours to finish, we tested on workstation Intel Xeon(R) CPU E5-1620 v2 @ 3.70GHz × 8, RAM 64GB

IV. CORE FUNCTIONS:

1. learn_all.m: learns binary deep neural network - INPUT: - Xtrain (number of training sample X dimension): training data - Xval (number of validation sample X dimension): validation data - val_gnd_inds (number of validation sample X number of grountruth): groundtruth indices for validation data. - hiddenSize: array contains number of layers in all hidden layers. In our code, we fix number of hidden layers is equal to 3 - lambda1, lambda2, lambda3, lambda4: lambda values respectively correspond to regularization, binary constraint violation, independence and balance - iter_lbfgs: number of L-BFGS iteration - max_iter: maximum number of iteration for alternating optimization over (W,c) and B - OUTPUT: - stack: cell contains (W,c) in our binary deep neural network.

2. feedForwardDeep.m: does feedforward given an input sample, output is obtained from penultimate layer (layer n-1) - INPUT: - stack: cell contains (W,c) in our binary deep neural network (output of learn_all.m function) - data (dimension X number of sample): the input sample.