Awesome

This repository contains the code associated to the DISCO Nets paper that can be found on arxiv.

Requirements:

• Python + classic packages (numpy)

• Chainer library for Python

If you use this work, please cite:

D. Bouchacourt, M. P. Kumar, S. Nowozin, "DISCO Nets: DISsimilarity COefficient Networks", NIPS 2016

Repository contents

• utils/scores.py : define here additional scoring function if needed. We have implemented the \alpha -\beta norm with \alpha = 2 (as used in our experiment)
• examples/HandPoseEstimation/train.py : launching script
• examples/HandPoseEstimation/hand_pose.py : defines the DISCO Nets and runs training
• examples/HandPoseEstimation/hand_pose_testing.py : testing utils specific to hand pose estimation

Run a simple example :

The "example" folder allows you to use the DISCO Nets on the NYU Hand Pose dataset NYU Hand Pose dataset. Data should be pre-processed using the code from Markus Oberweger and require the installation of DeepPrior to load the data.

If you want to use a GPU, set up gpu = True and your gpu ID in the file hand_pose.py. To run the example, set up your parameters in the file train.py and run python train.py from your terminal.

Parameters:

• beta : same as in the paper.
• seed : random seed to initialise the network weights. All biases are initialised to 0.
• alpha : dissimilarity coefficient hyper-parameter, referred as gamma in the paper.
• C : weight decay
• savedir : folder to save the model + monitored values at each iteration
• datadir : folder to find the data, in the NYU Hand Pose example "../"
• nrand : size of the noise vector
• finger_w : used in our fingers experiment, leave it to 1.0 if you want to consider all 5 fingers
• fingers : used in our fingers experiment, leave it to the full list of fingers if you want to consider all 5 fingers