Deep Metric Learning Using Triplet Network

This code replicates the results from the paper “Deep metric learning using Triplet network” (http://arxiv.org/abs/1412.6622).

It can train a TripletNet on any of the {Cifar10/100, STL10, SVHN, MNIST} datasets.

##Data You can get the needed data using the following repos:

• CIFAR10/100: https://github.com/soumith/cifar.torch.git
• STL10: https://github.com/eladhoffer/stl10.torch
• SVHN: https://github.com/torch/tutorials/blob/master/A_datasets/svhn.lua
• MNIST: https://github.com/andresy/mnist

##Dependencies

• Torch (http://torch.ch)
• "eladtools" (https://github.com/eladhoffer/eladtools) for optimizer.
• "nngraph" (https://github.com/torch/nngraph) for TripletNet configuration.
• "cudnn.torch" (https://github.com/soumith/cudnn.torch) for faster training. Can be avoided by changing "cudnn" to "nn" in models.

##Models Available models are at the “Models” directory. The basic Model.lua was used in the paper, while NiN based models achieve slightly better results.

##Training You can start training using:

th Main.lua -dataset Cifar10 -LR 0.1 -save new_exp_dir

##Additional flags