DeepLearningStack

DeepLearningStack allows easy definition of deep neural networks with XML configuration files. The package contains modules for definition of MLP, convolutional networks, reinforcement learning and recurrent neural networks.

A library for deep learning in Thenao. It allows easy definition of deep networks using XML configuration files, and has different packages for reinforcement learning and convolutional networks.

Installation

Download the package, cd into the directory and type:

sudo python setup.py install

Alternatively, you can use pip:

pip install DeepLearningStack

Tutorials

AlexNet

This tutorial shows how to create a simple convnet using an XML configuration file. The network is defined in AlexNet.xml, which defines different layers and connect them to each other to form the computation graph of the network. Each layer is deined as

<layer type="" name="">
<input>input1_layer_name</input>
<input>input2_layer_name</input>
...
<param1>value1</param1>
<param2>
  <param2_tag1>val11<param2_tag1>
  ...
</param2>
...

</layer>

Each layer should have a type, e.g. convolution, and an arbitrary but unique name. The name of the layer is used to direct its output to another layer through the <input> tag. Each layer type has layer-specific parameters, such as kernel size for convolution layer. See config/Layers.xml for a complete list of layers and their parameters.

Deep Q-learning

This tutorial implements the deep Q-learning method for active object recognition described in [1]. Active object recognition deals with smart exploration of objects to achieve higher accuracies with smaller number of images processed. A neural network is trained, using Q-learning iterative update rule, to perform a series of actions on GEMRS [2] objects to increase the accuracy of label prediction. The architecture of the network defined in DQLArch.xml is composed of two fully connected layers followed by softmax, which predicts the action-values for 10 different actions. The actions rotate the objects inward or outward w.r.t camera with different magnitudes.

To run this tutorial, you need to download the GERMS dataset from [3]. These are the belief vectors over different object labels, calculated from features extracted by VGG 7-layer network. This network was pre-trained on ImageNet, and was not fine-tuned for GERMS. After downloading the belief files, put them in Tutorials/DeepQL/ and execute DQL.py script.

[1] Malmir M, Sikka K, Forster D, Movellan J, Cottrell GW. Deep Q-learning for Active Recognition of GERMS: Baseline performance on a standardized dataset for active learning. InProceedings of the British Machine Vision Conference (BMVC), pages 2016 Apr 13 (pp. 161-1).

[2] http://rubi.ucsd.edu/GERMS

[3] https://drive.google.com/folderview?id=0BxZOUQHBUnwmQUdWRGlPMGw4WHM&usp=sharing