MTRNN

Description

This Aquila-compliant module trains multiple time-scales recurrent neural network (MTRNN) using back-propagation through time training (BPTT) algorithm

Libraries

• CUDA 5.0
• YARP

Parameters

Launch this module with –help parameter to see the list of possible options

Ports Accessed

• none

Ports Created

Input ports:

• /mtrnn/[id]:i
  • input port to which Aquila connects

Output ports:

• /mtrnn/[id]:o
  • output port to which Aquila connects

Input Data Files

Needs a training file, which can be set from:

• terminal via the following command: '–training_set [FILENAME]'
• Aquila via the top-level menu

Output Data Files

By default, trained MTRNN is saved to 'mtrnn.txt' file. This can be changed from:

• terminal via the following command: 'set output [FILENAME]'
• Aquila via the top-level menu

Configuration Files

Default values for various paramters are loaded from a configuration file, which can be modified in /conf/config.ini

The file consists in a few sections:

output mtrnn.txt
showProgress 1
maxThreads 256
iterations 10000
ioDelta 1
fastDelta 1
slowDelta 1
numFastNeurons 10
numSlowNeurons 10
weightRange 0.025
threshold 0.0001
learningRate 0.005
momentum 0.0

output - output file used for saving trained MTRNN show_progress - if 1 then training progress will be shown if 0 then not maxThreads 256 - maximum number of GPU threads per block iterations 10000 - maximum number of iterations ioDelta 1 - delta_t value that of input-output neurons fastDelta 1 - delta_t value that of fast neurons slowDelta 1 - delta_t value that of slow neurons numFastNeurons 10 - number of fast neurons numSlowNeurons 10 - number of slow neurons weightRange 0.025 - initial range of synapses (in this case from -0.025 to 0.025) threshold 0.0001 - error threshold that, once reached, will cause the training to stop learningRate 0.005 - learning rate used druing the training momentum 0.0 - momentum used during the training

Tested OS

Linux, OSX and Windows.

Example Instantiation of the Module

This will launch the training on 3 different GPUs (IDs: 0,1,2), which will divide the sequences provided in 'training_set.txt':

mtrnn –gpu 0 1 2 –trainingSet training_set.txt –iterations 10000 –learningRate 0.005

Author

Martin Peniak Copyright (C) 2008 RobotCub Consortium CopyPolicy: Released under the terms of the BSD This file can be edited at src/main.cpp.