Implementation of a simple MLP from scratch in C. It does not use any external libraries. It has limited features as I want to keep it simple and easy to learn from.

Micrograd by Andrej Karpathy was a big help.

Build and Run (Windows)

To build the train.c file with gcc:

$  gcc -o train train.c src/*.c

then run with:

$  train.exe

If you want to run the tests, use the following code and replace tests/test.c with the test you want e.g.

$  gcc -o test_tensor_ops tests/test_tensor_ops.c src/*.c

Demo

The train.c file contains the training loop for a binary classifier with two hidden layers of size 16. Binary cross entropy loss is used with SGD as the optimizer. The dataset is the moons dataset. This setup is identical to the demo from the previously mentioned micrograd so that I can compare performance; however, I used binary cross entropy instead of hinge loss. Here is an example decision boundary after 100 iterations using 100 data samples:

Visualising

The visualise_decision_boundary.ipynb file will display the decision boundary of the binary classifier by reading data that train.c exports. Train.c will pass points on a grid through the trained model and export the points along with the models prediction. Ensure that "n_x_steps" and "n_y_steps" in both train.c and visualise_decision_boundary.ipynb match. Matplotlib and Numpy must be installed.