train_one_net.m

From Spinach Documentation Wiki
Jump to: navigation, search

Generates a training library, trains a network and writes it into a file with the specified number. If the specified file exists, the function will load that network file and continue training.

Syntax

    train_one_net(parameters,trsets,file_number)

Arguments

  parameters  - library and network specification struc-
                ture (see the header of deer_lib_gen.m)

  trsets      - sizes of independent randomly generated 
                training libraries to sequentially train
                against. [160k 160k 160k 160k 160k 160k] 
                works well on a Tesla V100 card.

  file_number - the network object will be saved into a
                file with this number as the name, this
                file also serves as a restart checkpoint.

Two more fields are required in the parameters structure:

  parameters.lastlayer   - activation function to use in 
                           the output layer ('tansig',
                           'logsig', or 'purelin').

  parameters.layer_sizes - number of neurons per layer, a
                           row vector where the number of
                           elements is the number of hid-
                           den layers desired.

Outputs

This function writes a MAT file with the network object and the parameters data structure.

Examples

The example below will train a single network using the parameters given in the netset_params.m file for the network ensemble optimized for any peak width.

	% Load training set parameters
	run('net_set_any_peaks/netset_params.m');

	% Specify the sizes of training databases to train against
	trsets=[160e4 160e4 160e4 160e4];

	% Run the network training for a a single network
	train_one_net(parameters,trsets,111);

The function will save the the network into 111.mat file.

Notes

  1. A CUDA capable NVidia GPU is required.
  2. If the file pre-exists, the network will be loaded and used as the initial guess.

See also

Neural network module

DEER/PELDOR experiments


Version 2.5, authors: Ilya Kuprov, Steve Worswick, Jake Keeley, Tajwar Choudhury