saveWeights <file-name> [-values <num-values> | -text | -noFrozen | -onlyFrozen]
This writes the weights for every link in the network into a special file. The file can be later reloaded into a network of the same architecture. If the file name has the extension .gz, .bz, .bz2, or .Z it will automatically be compressed. It is customary to give weight files the extension .wt.
When training networks with momentum, deltaBarDelta, or other
"sophisticated" algorithms, it is usually important to save more
information for each link than just its weight. If num-values is
2, both the weights and the deltas (used by momentum) will be saved. A
value of 3 will save the weights, deltas, and "last values", which are
used by deltaBarDelta in
If a num-values of 3 is specified in standard
only the weights and deltas will be written (as there are no
lastValue fields). Therefore, you can use 3 to store all of the
link values in either version of Lens.
The default is to write a binary weight file. If the -text flag is given, values will be written in text format.
Ordinarily, all links in the network will be saved. If -noFrozen is specified, frozen links will not be saved. If -onlyFrozen is specified, thawed links will not be saved. These are useful if you only wish to save part of the network.
When using networks with randomized structures, it is important to record the random seed used when building the network so the same architecture can be reconstructed later. Otherwise, weight files saved with this command may be useless later.
To save the weights in the compressed binary file greatNet.wt.gz:
lens> saveWeights greatNet.wt.gzTo save all of the link values of non-frozen weights in text format:
lens> saveWeights foo.wt.gz -v 3 -n -t