ML

A stb-style (for now) ML library built while learning. The Git commit history shows my learning path. Future updates depend on what I want to learn and what I want to change.

About

ML is a small machine learning library written in C, inspired by tsoding's videos and project.

• https://www.youtube.com/playlist?list=PLpM-Dvs8t0VZPZKggcql-MmjaBdZKeDMw
• https://github.com/tsoding/nn.h

This project is not intended to be a production-ready machine learning framework. It is a learning-oriented project where I implement the basic parts of neural networks from scratch.

Current Features

1. Matrix structure and operations (alloc, free, mul, transpose, slice, etc.)
2. Multi-layer neural network with forward propagation
3. Activation functions: sigmoid, ReLU, Leaky ReLU, tanh, GELU, swish, linear, softmax
4. Loss functions: SSE, MSE, BCE, CCE, softmax + CCE
5. Weight initialization: Xavier and uniform random
6. Backpropagation and finite difference gradient estimation
7. SGD optimizer with mini-batch training
8. Model save and load (binary format)
9. Data loading: CSV and binary (MLDB) formats
10. Model verification: binary (threshold) and multi-class (argmax) accuracy
11. Label utilities: one-hot encoding and argmax

Tools

• tools/img2bin — PNG images → MLDB binary data
• train_visualizer.c — real-time training visualization (cost curve + decision boundary)
• check.c — load model + run single-image inference

Data Loading

Function	Format	Description
ml_data_load_csv	.csv	CSV file (header row + comma-separated floats)
ml_data_load_bin	.bin	MLDB binary ("MLDB" magic + rows/cols + raw floats)

Both produce a Mat where each row is [input features ... | target values ...].

MLDB files are generated by the preprocessing tool:

./tools/img2bin --max 100 train train_1k.bin 10

Activation Functions

Symbol	Function	Derivative
ML_SIGMOID	sigmoid	dsigmoidf
ML_RELU	ReLU	dReLUf
ML_LRELU	Leaky ReLU (α=0.01)	dLReLUf
ML_TANH	tanh	dtanhf
ML_GELU	GELU	dGeLUf
ML_SWISH	swish	dswishf
ML_LINEAR	identity	dlinearf
softmaxf()	row-wise softmax	—

Loss Functions

Forward	Backward	Use
ml_model_loss_sse	ml_model_loss_dsse	Sum of squared errors
ml_model_loss_mse	ml_model_loss_dmse	Mean squared error
ml_model_loss_bce	ml_model_loss_dbce	Binary cross-entropy
ml_model_loss_cce	ml_model_loss_dcce	Categorical cross-entropy (one-hot)
ml_model_loss_softmax_cce	ml_model_loss_dsoftmax_cce	Softmax + CCE combined

Logging

Define ML_LOG_LEVEL before including ml.h to control log output:

Level	Output	Usage
0	none (zero overhead)	Release builds
1	LOG_ERROR	Production (default)
2	+ LOG_WARN	Tuning
3	+ LOG_INFO	Training progress
4	+ LOG_DEBUG	Debugging

example:

gcc -DML_LOG_LEVEL=3 ml.c -o ml -lm

Without -D, defaults to level 1.

Build

A simple build script is provided:

./build.sh

You can also compile manually:

gcc -Wall -Wextra -O3 ml.c -o ml -lm

Some experiments may require external libraries such as raylib.

For example, if you want to build the visualization program, you may need something like:

gcc -Wall -Wextra -O3 train_visualizer.c -o train_visualizer -lraylib -lm

The exact command may depend on your system and installed libraries.

The preprocessing tool (tools/img2bin) requires stb_image.h, which is not tracked in this repository. Download it before building:

wget -O tools/stb_image.h https://raw.githubusercontent.com/nothings/stb/master/stb_image.h
gcc -O3 tools/img2bin.c -o tools/img2bin -lm

Project Style

This project is currently written in a stb-style direction. This may change in the future. The current structure is not final, and future changes will depend on what I want to learn next.

Notes

This repository is experimental. Code may be rewritten, renamed, removed, or reorganized at any time. The main purpose of the project is learning, not API stability. That said, you are welcome to fork this project and experiment with it yourself.

Possible Future Directions

Some possible future directions include:

1. Cleaner stb-style organization
2. More flexible layer abstraction
3. Separate loss and optimizer modules
4. Better model serialization
5. GPU backend experiments
6. A more general ML library architecture

(None of these are guaranteed)

License

This project is licensed under the MIT License.