Some parameters are required, some are optional and some may be required,
optional, or superfluous depending on the mode in which the system is
run. Refer to Chapter 7 for simple usage examples. Note that
some training parameter settings are not stored in the network. Therefore,
when invoking SNoW in a mode other than -train or interactive, it
is good practice to specify the same optional command line parameter settings
that were used when training the specified network. Alternatively, reading
the descriptions of each parameter below will give insight as to which
parameters are written to the network file. The user can then decide
precisely which parameters need to be set in each mode for his particular
purposes.
There are four kinds of SNoW command line parameters: (1) architecture definition, (2) training, (3) testing, and (4) I/O parameters.
Architecture definition parameters instantiate target nodes and algorithms, tweak algorithm parameters, and modify the way algorithms behave while training. They are also training parameters, and they conform to all the rules of training parameters. Unless otherwise specified, their settings are stored in the network file at the end of -train or -interactive mode. You may find it useful to define an architecture file and mention it in the command line or in a shell script (using the -A option) instead of describing the entire architecture directly in the command line. See Chapter 7 for examples.
Training parameters can always be specified in -train mode and may or may not be written to the network file at the end of execution. They are also available in -interactive mode unless otherwise specified. If a training parameter is written to the network file at the end of -train or -interactive mode, its setting cannot be overridden by specifying a new setting on the command line in another mode. Otherwise, it can (and often should be, since its setting during the initial training was lost when it wasn't written to the network file). In -test mode, training parameters that aren't written to the network after -train or -interactive mode can be specified on the command line and will take effect when incrementally learning with the -i + testing parameter. Settings that were stored in the network will also take effect in this scenario. Training parameters never have an effect on -evaluate or -server modes.
Testing parameters are available in any mode other than -train unless otherwise specified and are never written to the network. In -train mode, they can also take effect when testing after training is complete with the -T I/O parameter.
I/O parameters are for specifying the names of SNoW's input and output files and for changing the type and amount of information presented to the user. Unless otherwise specified, they are available in every execution mode.