readme.parameter

C-package of "Long Short-Term Memory" for Protein classification

This is the documentation for the parameter file.

The parameter file is loaded from LSTM by the parameter -c and consists
of two sections.

:: Memory cell (MC) configuration

Here the number of MCs and the parameters for each MC can be set.

number memory cell
blocks:                : Number of memory cell blocks. More blocks lead to
                         a more complex network but can result in overfitting.
block size             : Block size. A memory cell can have more than one
                         unit. Default 1.
initial input bias     : Initial input bias. A good choice is to set
                         all input bias to different negative values.
                         E.g. -2.0, -3.0, -4.0 etc.
initial input gate bias: Initial input gate bias. Default -1.0.
initial output gate
bias                   : Initial output gate bias. Default -1.0.
initial output weight  : Initial weight from the memory cell to the output.
                         A good choice could be to set half of all MCs negative
                         (-1.0) and the other half positive (1.0) and set for each
                         positive/negative pair the same initial input bias.

:: Other parameters

Here the parameters like window size, learning rate, location of the
datasets etc. are set.

windowsize             : Window size for the LSTM scan process
outputbias             : Output bias for the output neuron.
inputdata ..           : Locations of the datasets.

targetvalue0           : Target value for the positive class
targetvalue1           : Target value for the negative class

learning rate          : Learning rate for the training process. Note that a
                         to high learning rate results in a not converging
                         training process. To low learning rates slow the training
                         process down.

half interval..        : Half interval length for random weight initialization.
                         Default 0.1 so weights are randomly set to [-0.1,0.1].

performing test after
? epochs               : Number of epochs after a test is made within training.
write weight after ?
epochs                 : Number of epochs after a weight file is written

initialization of
random generator       : Initialization of the random generator with
                         a fixed seed given here. '0' seed with time.

reset the net after
each sequence?         : If the net should reset after each sequence.
                         1 : yes 0 : no. 1 is the default.

weight update after
sequence or epoch?     : Online (1) or batch learning (0)
stop learning after n
epochs                 : Number of epochs after the training should stop.

Important settings for different datasets:

size of training set   : Maximum number of training sequences, positive plus negatives.
maxlength of training
set                    : Maximum length of a sequence in the training dataset.
size of test set       : Maximum number of test sequences, positive plus negatives.
maxlength of test set  : Maximum length of a sequence in the training dataset.