Mode Selection

SNoW's basic usage is as follows:

snow -train -I inputfile -F networkfile [ -AaBbcdEefGgiLlMmOoPpRrSsTtuvWwz ]
snow -test -I inputfile -F networkfile [ -abEefGgiLlmOopRSstvwz ]
snow -interactive -I inputfile -F networkfile [ -AEfLoPpRvWw ]
snow -evaluate -x example -F networkfile [ -befmpRvw ]
snow -server port -F networkfile [ -befLlmopvw ]

Parameters in the bracketed section are optional, and the others are required. All parameters require an argument, and thus must be listed separately on the command line. Their order does not matter except for the first parameter, which must be one of -train, -test, -interactive, -evaluate, or -server. This parameter sets one of the following modes:

-train

: The system is run in training mode. Command line parameters describing the desired SNoW architecture are used to create a new network. The input file is considered to be a set of labeled training examples. Each example in the file is considered a positive example for all targets which are included (active) in the example and a negative example for all other targets (absent from the example). At the end of training, the resulting network (i.e., weight vectors and other such parameters) is written to the network file. If the file already exists, it is overwritten.

-test

: The system is run in a batch test mode. First, the system is initialized with the data in the network file. Next, each example in the input file is presented to the system and classified. The input file can consist of labeled or unlabeled testing examples. If the examples are labeled, the result of the classification can be compared with the label of the example and scored, with a final accuracy reported after all examples are presented. The result of each prediction can also be output to a file in a number of ways and scored externally.

-interactive

: The system is run in an interactive training mode. First, the network is initialized with the data in the network file, just as in test mode. If the network file doesn't exist, a new network is created according to any specified architecture parameters as in training mode. The input file contains unlabeled examples with individual handling instructions. At the end of processing, the resulting network (i.e., weight vectors and other such parameters) is written to the network file. Note that when running in this mode, the following additional command line parameters are implied and cannot be overridden: -e count:1 -g - -l - -r 1 -s f . Also, the Naive Bayes, Constraint Classification, threshold relative updating, Gradient Descent, and thick separator algorithms are not available in this mode.

-evaluate

: The system is run in an ``online'' test mode. A single labeled or unlabeled example is supplied on the command line. SNoW assumes it is unlabeled. The process terminates after making a prediction for this single example. The predicted target ID is output and also returned to the operating system in the process's exit status. Note that running the system in Evaluation mode loads the network once for each example, and thus is not the best way to process large sets of examples.

-server

: The system is run in batch test mode. First, a client program may send a string of command line parameters to the server. Any of the parameters that are legal when instantiating SNoW in -server mode may be passed by a client to the server except the -e parameter. Input to the server can consist of labeled or unlabeled testing examples and is expected to be received over a user specified port. Each example is presented to the system and classified. If the examples are labeled, the result of the classification can be compared with the label of the example and scored. SNoW's output is then sent back over the same port. When instantiating SNoW as a server, the network is loaded once for any number of examples from any number of files to be tested.

See Chapter 7 for more detailed examples of SNoW's usage in each of these modes. Also, see Section 6.1 for precise descriptions of accepted example file formats.