Input/Output Parameters

-c <i>

: This setting controls the generation of results that can be used to create a learning curve. It can only be used in -train mode, and it must be accompanied by the -T parameter. When its integer argument is positive, it denotes an interval of training examples. After each interval of the specified length within a given cycle (see -r), the network is tested with the examples in the file specified by the -T parameter and results are output. If -u - is specified, no testing or results output will occur during the first cycle. The presence of the -T parameter (both with and without the -c parameter) also causes the network to be tested when training has entirely completed. We suggest setting -u - -r 2 when using a positive argument for this parameter. This parameter is not available in -interactive mode. The default is 0, when the network is never tested during training.

-E <errorfile>

: Specifies the name of a file in which to write information about mistakes during testing. If the file already exists it is overwritten. This option is only supported by the -o accuracy output mode (which is the default). If -o is set otherwise, no error file will be created. No default.

-F <networkfile>

: Specifies the name of a file which the resulting network is written to (after training) or read from (for testing). This parameter is required in every mode of execution. No default.

-I <inputfile>

: Specifies the input file from which examples are read. During training and testing the input file specifies the training examples and testing examples, respectively. This parameter is required in -train and -test modes and unavailable in -evaluate and -server modes. No default.

-o <accuracy | winners | softmax | allpredictions | allactivations | allboth>

:
Specifies which output mode to use when reporting results during test mode. All output modes output information on the algorithms used as well as the overall accuracy of the test. If this information is irrelevant to your application, consider the -v off verbosity mode. Several output modes are available, and are described below. The default is accuracy.

accuracy

: This output mode requires labeled examples. SNoW compares each prediction to each example's label and keeps track of correct and incorrect predictions, outputting an accuracy report at the end of testing.

Algorithm information:
Winnow: (1.35, 0.741, 4, 0.2) Targets: 0-2
52 test examples presented
Overall Accuracy - 96.15% (50 / 52)

winners

: This mode outputs the target ID with the highest prediction confidence for each example in the test set. For example, if we had targets with ID's 0, , and , the output appearing in between the algorithm information and the overall accuracy report might appear as:

1
2
0
2
1

On the other hand, if we had a network containing only target node with ID , the single-target testing policy would take effect^5.3, and the output appearing in between the algorithm information and the overall accuracy report might appear as:

negative
positive
negative
negative
positive

softmax

: For each example, this mode outputs the softmax normalized activations of each target. The output is sorted by activation. If the test data is labeled, the correct label for each example is output and the correct target's output is marked with an asterisk.

Example 47 Label: 0
0: 0.77582*
1: 0.14876
2: 0.075420

Example 48 Label: 1
2: 0.41475
1: 0.38784*
0: 0.19741

allpredictions

: This mode outputs, for every example, a list of all targets and their predictions ( or 0, indicating if it was the chosen target for this example or not). The target with the highest prediction confidence is predicted as true (), and the rest are false (0). The output is sorted by ID. If the test data is labeled, the correct label for each example is output and the correct target's output is marked with an asterisk.

Example 47 Label: 0
0: 1*
1: 0
2: 0

Example 48 Label: 1
0: 0
1: 0*
2: 1

allactivations

: For each example, this mode outputs the activations of each target. The first activation is the prediction confidence (see Section 4.3.5 for more details on prediction confidence). The second is the target's raw activation (in other words, it is simply the dot product of the target's weights and the example^5.4). The third is the softmax normalized activation. The output is sorted by activation. If the test data is labeled, the correct label for each example is output and the correct target's output is marked with an asterisk.

Example 47 Label: 0
0: 0.70005 4.8475 0.77582*
1: 0.30915 3.1959 0.14876
2: 0.18493 2.5167 0.075420

Example 48 Label: 1
2: 0.67585 4.7348 0.41475
1: 0.66099 4.6677 0.38784*
0: 0.49811 3.9924 0.19741

allboth

: This mode outputs, for every example, a list of all targets and their activations and predictions as described in the above modes. If the test data is labeled, the correct label for each example is output and the correct target's output is marked with an asterisk.

Example 47 Label: 0
0: 1 0.70005 4.8475 0.77582*
1: 0 0.30915 3.1959 0.14876
2: 0 0.18493 2.5167 0.075420

Example 48 Label: 1
2: 1 0.67585 4.7348 0.41475
1: 0 0.66099 4.6677 0.38784*
0: 0 0.49811 3.9924 0.19741

-R <results_file>

: Specifies the name of a file in which testing results and any extra information generated by a non-default verbosity setting are output (see the -v parameter). If this parameter is unspecified, the output will be directed to the console. No default.

-T <testing_file>

: Using this parameter, a network can be trained and tested in the same invocation of snow. After training the network, test examples are read from the file specified using this parameter, and output is given just as if SNoW was run in -test mode. No default.

-v <off | min | med | max>

: Specifies the verbosity, or the level of detail at which output is given. The default is min.

off

: This mode turns off the output.

min

: This mode outputs minimal informational messages.

med

: This mode outputs a moderate information, such as when target nodes are updated for mistakes during training, and outputs a list of features at the end of training.

max

: This mode outputs lots of information, such as when target nodes are updated for mistakes during training, showing each example before the network begins to process it, and outputting a list of features at the end of training. All predictions with activations are output during testing.

-x <example>

: This parameter is only legal in -evaluate mode, and it must be specified in that mode. The example can take any form that is legal in an example file. It is usually necessary to surround the example in quotes in most shells. No default.

... effect^5.3

See Section 4.4 for more information on testing policies.

... example^5.4

In the case of clouds containing more than one target node, this activation is the unweighted summation over all target nodes that learn the class label of those dot products.