Scyte is a neural network framework in C – it is fast, lightweight and easy to install! The framework is inspired by TensorFlow, and is based on computational graphs and automatic differentiation.

To clone and build scyte execute:

git clone [email protected]:marvrez/scyte.git
cd scyte
make -j

If no errors popped up during the installation you can run the provided examples. For instance, if you want to train or run inference a simple model that learns to act as an XOR-gate, you can simply do the following.

# train XOR-model for 1000 epochs, and save model weights to 'xor.weight'
./scyte xor -e 1000 xor.weight

# test the newly trained XOR-model
./scyte xor -p xor.weight

Similarly, you can also train and/or run inference on a model that learns the MNIST dataset. The simplest way to do this is to download and extract the data from here. You can then do the following steps to train and/or run inference on the model.

# train mnist model for 100 epochs, and save model weights to 'mnist.weight'
./scyte mnist -e 100 mnist.weight --label_path ./mnist/mnist.labels --data_path ./mnist/mnist.train

# test the newly trained XOR-model
./scyte mnist -p mnist.weight -i <test_image>

In general, the --label_path file is a new-line seperated list of the labels, and --data_path file contains a newline-separated list of the image paths relative to label file.

For more options and help, pass -h or the --help flag.

For best performance OpenBLAS is recommended to be installed and used. You can enable it by setting OPENBLAS=1 in the Makefile.

If you don't want to install OpenBLAS but do have OpenMP installed, you can still speed up the processing by setting OPENMP=1 in the Makefile.