To configure the project you can run the configure.sh file, which will create build directories debug and release.

To build, simply run

cmake --build [debug|release]

From the root project directory

Usage:

Each neural network is held in a directory, containing 4 principle files.

[network]/
          .
          ..
          config.txt
          topology.txt
          training_examples.txt
          weights.bin

config.txt: Text file containing hyperparameters for the network, in this order:

top_learning_rate bot_learning_rate decay_rate learning_rate_cycle_length hidden_unit_activation output_unit_activation

Supported activation functions are sigmoid, tanh, binary and none

topology.txt: Text file containing the size of each layer in order, starting with the input layer, and finishing with the output layer.

training_examples.txt: Text file containing training examples to be used, a delineated file of floating point numbers, with the input vector followed by the expected output vector for each example.

ie for the xor dataset the file could look like:

0 0 0
0 1 1
1 0 1
1 1 0

Finally, weights.bin olds any saved weights for this particular networks.

Commands:

exit: exit the program.

train <epoch> <output csv>: train the neural network on the dataset for <epoch> epochs, and write statistics to a csv file.

save: save weights from ram to disk. (If you don't want to overwrite, you have to rename the old weights file as a backup)

test <test file> <output csv> Test, followed by the input vector to manually test the program, writes the output to stdout.

To run the given networks:

To run a network yourself, once the binary is built, just execute and pass the root directory of a network.

For example, you can first run the following to build the release binary;

$ cmake --build release

Then run the arithmetic network for testing:

$ release/main networks/arithmetic