These notes describe the process of training a set of neural nets using the ASAT TestBed software. The ASAT TestBed software is currently designed to use the TIMIT data set and neural network classifiers built on QuickNet software.

Training

train.pl is the training script. For this step, you must have a set of label files and feature files as built with the TestBed building scripts. Training requires two keys in the config file - a [trnprg] key that holds parameters for the training program and a [train] key that holds parameters for the TestBed script itself. Training is done in batch per feature, so a different configuration file is used for each. For QuickNet, the following parameters are used: 

[global]
# In addition to the required keys from the Build programs, the following
# new key is required
outputdir = [ output directory for the trained weight files ]

[train]
prefix = [prefix the output files should use - conventionally the feature name]
trainfeaslist =  [ full path to the list of partitions to use for training ]
trainlabslist =  [ full path name to the list of label files to use ]
trainsize = [ number of partitions to use for training (batch) ]
crosssize = [ number of partitions to use for cross validation (batch) ]
trainfiles = [ which partitions to use for training (single run) ]
crossfiles = [ which partition to use for cross validation (single run) ]

[trnprg]
# These are parameters to pass to the underlying training program.  For 
# QuickNet neural nets, the following parameters are typical.  Output size
# and hidden size should vary by feature.
hardtarget_format=ilab
hardtarget_window_offset=4
window_extent=9
ftr1_window_len=9
ftr1_ftr_count=26
ftr1_norm_mode=file
learnrate_vals=0.08
mlp3_input_size=234
mlp3_output_size=5
mlp3_hidden_size=100
mlp3_bunch_size=16

In batch mode, the program creates sliding windows of partitions of data to train on, keeping the number of partitions in the data the same, but moving the first partition and the cross-validation partition(s) each time. Each time, new weights are created for the new partition, so the final result is a number of weight files equal to the number of partitions declared. In single file mode, only one set of training is performed.