WIP: Basic feed-forward neural network with back propagation in zig language

This is zig experiment of writing basic feed forward neural network with back-propagation. It is built and tested for digit recognition on MNIST dataset.

Note: This is my first time writing or using neural network, therefore there might be some math errors etc. If you see any, please let me know. 🙂

Setup

$ git clone [email protected]:Deins/zig-nnet.git
$ git submodule init
$ git submodule update

Build

zig build -Drelease-fast

Output program should be built at default location ./zig-out/bin/nn or *.exe etc. depending on platform.

Usage

Download MNIST dataset from here and extract it at ./data/digits/.

(Optional) It is recommended to preprocess input for faster further loading. (has to be done once or after dataset or file format is modified).

./zig-out/bin/nn preprocess

Train new network and save output net:

 ./zig-out/bin/nn.exe --save data/n.net --learn-rate 0.01 --batch-size 64 --epoches 5 --workers 4 train

You can use --load data/n.net afterwards, to load existing net instead of generating new and train it further or with different settings.

Test trained model on test data set, and save classified results inspection:

./zig-out/bin/nn.exe --load data/n.net --img-dir-out ./test-out test