Tidf

A tiny, simple, stupid but powerful deep-learning framework in c++.

Why Tidf?

• Powerful: Tidf provides a Keras-like api.
• Simple and stupid: Support brandcast and made in pure c++11.
• Tiny but fast: The implementation is under 1,000 semicolons.

Components

Layers

• Dense Layer (Linear Layer)

Regularizations

• L1 regularization
• L2 regularization
• Dropout

Activations

• sigmoid
• relu
• tanh

Optimizers

• SGD (Stochastic gradient descent)

Cast Functions

• CrossEntropy
• MSE (L2Loss)
• L1Loss

more to come! :)

Examples

Train a 3-Layers neural network

#include "../tidf/core.h"

int main() {
    _TIDF_INIT_;
    NEW_MAT(train_inputs, double, 
      ({{0.0, 0.0, 1.0},
        {1.0, 1.0, 1.0},
        {1.0, 0.0, 1.0},
        {0.0, 1.0, 1.0}}));

    NEW_MAT(train_outputs, double,
      ({{0.0, 1.0, 1.0, 0.0}}));
    
    Net<double>* net = new Net<double>(train_inputs.transpose(), train_outputs);
    net->addLayer< 4 >(LayerType::Dense, "sigmoid"); 
    net->addLayer< 4 >(LayerType::Dense, "tanh"); 
    net->addLayer< 1 >(LayerType::Dense, "sigmoid");
    net->compile("CrossEntropyLoss", "SGD");
    net->fit(100000);
    std::cout << net->predict(
      MAT( double, ({{1.0}, {1.0}, {1.0}})) ) << std::endl;
    return 0;
}

Result:

Matrix(1 x 1):
0.998889

You can also load matrix from .txt files:

  ...
  Matrix<double> train_inputs = load_mat<double>("train_X.txt");
  Matrix<double> train_outputs = load_mat<double>("train_Y.txt");
  ...

Matrix

tidf supports brandcast and provides a simple but powerful API to make it easier to work with matrices:

#include "tidf/core.h"

int main() {
    NEW_MAT(m1, double, ({{1, 4, 3}, {1, 2, 3}}));
    NEW_MAT(m2, double, ({{1, 2, 3}}));
    NEW_MAT(m3, double, ({{2, 3, 4}}));

    std::cout << m1.dot(m2.transpose()) << std::endl;
    std::cout << m2 * 2.0 + m3 << std::endl;
    std::cout << m2.apply([](double x) -> double {return x * 10.0;}) << std::endl;
    std::cout << m2.join(m3) << std::endl;
    std::cout << m1.row(1).sum() << std::endl;
    return 0;
}

License

MIT License
Copyright (c) 2023 Stepfen Shawn