mnist-c

C implementation of a neural network for classification.

Performance

Using default settings training takes about 1.5s per epoch on my Ryzen 5 2600.

• MNIST Handwritten Digits dataset: ~97% accuracy
• Zalando Fashion-MNIST dataset: ~85% accuracy

Architecture

Three fully connected layers (64-32-10 by default) with sigmoid activations in between, though this is trivially modifiable.

Usage

Training

git clone https://github.com/atzuur/mnist-c
cd mnist-c
make # optionally add `CC=clang EXTRA="-fuse-ld=lld"` to use clang
./a.exe train

Prediction

1. Draw a 28x28 MNIST-like image (see demo)
2. Convert to raw floating point using e.g. FFmpeg:

ffmpeg -v error -i YOUR_IMAGE.png -pix_fmt grayf32 -f rawvideo YOUR_IMAGE

3. Run prediction using the model:

./a.exe YOUR_IMAGE

References

• Michael A. Nielsen, ‘Neural Networks and Deep Learning’. Determination Press, 2015, http://neuralnetworksanddeeplearning.com/.