sklearn2c

Transpile trained scikit-learn models to dependency-free C code, and deploy ML on any microcontroller.

python transpile_simple_model.py model.joblib
# → generated_model.c  (pure C, no libraries, compiles anywhere)

---

Why

scikit-learn doesn't run on microcontrollers, and existing solutions (ONNX, TFLite Micro) require a runtime that adds significant flash/RAM overhead. sklearn2c walks the trained model's internal parameters and hard-codes them as a pure C function. There is no interpreter, no memory allocation, and no dependency beyond stdio.h.

The output compiles with gcc, arm-none-eabi-gcc, or any standard C compiler and drops directly into Arduino, STM32, ESP32 firmware.

Supported Models

Model	Output type	Generated function
LinearRegression	float	float predict(float *features, int n)
LogisticRegression	int	int predict(float *features, int n)
DecisionTreeClassifier	int	int predict(float *features, int n)

Quick Start

# 1. Install
pip install scikit-learn pandas joblib numpy

# 2. Train a model
python train_model.py             # LinearRegression on house prices
python train_decision_tree.py     # DecisionTree on student data

# 3. Transpile to C
python transpile_simple_model.py regression.joblib
# → generated_model.c

# 4. Compile and run
gcc generated_model.c -o model && ./model

# 5. Copy predict() into your Arduino/firmware sketch

Generated Code Example

Decision tree with depth 3 produces clean, readable C:

int predict(float *features, int n_features) {
    if (features[1] <= -0.2990f) {
        if (features[0] <= -0.0770f) {
            return 1;
        } else {
            return 0;
        }
    } else {
        return 2;
    }
}

Linear/logistic regression:

float prediction(float *features, int n_features) {
    float result = 150432.123f;           // intercept
    result += features[0] * 85.34f;
    result += features[1] * -312.77f;
    return result;
}

Note: regression models generate a function named prediction(), decision trees generate predict().

Deploy on Arduino

Copy the generated predict() into your sketch, no libraries needed:

// paste generated predict() here

void loop() {
    float sensors[2] = { analogRead(A0) / 1023.0f, analogRead(A1) / 1023.0f };
    float pred = predict(sensors, 2);
    Serial.println(pred);
    delay(1000);
}

How It Works

1. Load the .joblib model with Python
2. Introspect internal parameters (coefficients, tree nodes, thresholds)
3. Generate C source with parameters as compile-time constants
4. The predict() function is a series of branches / dot products, with no loops over parameters

Project Structure

sklearn2c/
├── transpile_simple_model.py    # Core transpiler
├── train_model.py               # Example: LinearRegression
├── train_logistic.py            # Example: LogisticRegression
├── train_decision_tree.py       # Example: DecisionTree
├── generated_model.c            # Example output
├── houses.csv / students.csv    # Sample datasets
└── README.md

Roadmap

• RandomForestClassifier (ensemble of trees)
• SVC with RBF kernel
• MLPClassifier (dense layers)
• Automated flash/RAM size estimation
• Arduino .ino file output

License

MIT