GUI application for visualising Tsetlin machine (TM) states and learning process.

• Download: tsetlinviewer.jar
• Download example data: data.zip

Requires Java 8 or later to run.

Running from command line:

java -jar tsetlinviewer.jar <config_file.cfg>

Example configurations:

• BinaryIris.cfg live learning demo using IRIS dataset.
• NoisyXor.cfg live learning demo using Noisy XOR dataset.
• BinaryIrisLog.cfg live learning IRIS that logs TA states and metrics in data/logs directory.
• BinaryIrisReplay.cfg replays IRIS learning from data/logs directory.

See Control keys on how to operate the GUI.

TsetlinMachineViewer uses configuration files to set up TM and viewer options. Configuration file consists of key : value pairs, one pair per line.

• classes (int, required) the number of classes. This is specific to the training data set: each training data point is classified into one of these classes.
• features (int, required) the number of input features. This is specific to the training data set: each training data point represents a Boolean vector of this size.
• clauses (int, required) the number of TM clauses per class.
• numStates (int, required) the number of TA states per decision, e.g. numStates : 10 means there are 10 include states and 10 exclude states.

• threshold (int, required for live learing) learning threshold parameter.
• s (float, required for live learing) learning rate parameter.
• countTrain (int, required for live learing) the number of training data points in the input dataset.
• countTest (int, required for live learing) the number of test data points in the input dataset.
• trainData (path string) training data file; see Training data format. You must specify either combinedData or both trainData and testData.
• testData (path string) test data file; see Training data format. You must specify either combinedData or both trainData and testData.
• combinedData (path string) input data file; see Training data format. The app will randomly divide these data into training and test data sets based on the countTrain and countTest options. You must specify either combinedData or both trainData and testData.

Recording CSV during live training:

• logData (boolean) enable CSV logging.
• logPath (string) path to write files to. Default is data/logs. The directory must exist.
• logName (string) file name prefix, default is tm.
• logFrequency (int) the number of training inputs between making log entries.
• logTAStates (boolean) save full TM state into *-tastates.csv. (This file is currently incompatible with the CSV-view mode, but it should be!)
• logStatus (boolean) save status variables (number of inclusions, number of decision flips, number of Type I and Type II feedbacks) into *-status.csv.

Providing viewCSV toggles CSV-view mode. CSV-view mode doesn't use live learning options.

• viewCSV (path string) path to CSV (or TSV) file containing the state of TAs. This can be a *-tastates.csv file saved by using logData option in combination with logTAStates.

Providing legacyCSVPathFormat toggles legacy CSV-view mode for older version CSV format (now deprecated).

• legacyCSVPathFormat (path string) path to CSV (or TSV) files, one file per learning epoch. Use %d in the path format to indicate epoch indices. Each line of the file corresponds to a literal, 2 × features lines in total: first features literals are positive, last features literals are negative. One line contains classes × clauses TA state values in order.

• uiDrawLiterals (boolean, optional) start GUI in literal display mode (true) or TA state display (false, default).
• uiClassesVertical (boolean, optional) start GUI in vertical layout (true, default) or horizontal layout (false).
• uiScale (float, optional) initial zoom, default is 1.

Training and test data is expected in the text format. Each line contains one data point (input vector and a label). Input vector is given as 0 or 1 values, the number of values should match features. The last value in a line is the label - a class that corresponds to this input. The label value must be within range 0 ≤ label < classes.

Use Right Mouse Button to pan the view and Control+Scroll to zoom in and out.

The easiest way to build this project is to use Eclipse IDE.

The project is dependent on ZoomUI library (use the latest sources from the master branch, not the JAR). Clone ZoomUI as another project in the same workspace:

.../Workspace/TsetlinMachineViewer
.../Workspace/ZoomUI

Once both projects are added to the workspace, Eclipse should see the dependency.