MLask

MLask

A custom c++ deep learning library.
Explore the docs »

View Demo · Report Bug

‍

Table of Contents

1. About The Project
   • Features
     • Built With
   • Getting Started
     • Prerequisites
     • Installation
   • Usage
     • Defining a model
       • Adding your own layer
     • Training the model
     • Exporting the model
   • License

About The Project

This is a custom c++ deep learning library. It was meant to be a challenge to understand how all the professional ML libraries work under the hood.

Features

• Fully connected layers
• Activation functions
• Backpropagation algorithm
• Export to ONNX format

(back to top)

Built With

• CMake
• Eigen
• ONNX
• Protobuf

(back to top)

Getting Started

Prerequisites

• CMake 3.28 or higher
• c++23 compatible compiler

Installation

# add these to install faster, however protobuf is pretty big, so it will take a while. Be patient.
set(protobuf_BUILD_TESTS OFF CACHE BOOL "" FORCE)
set(protobuf_BUILD_EXAMPLES OFF CACHE BOOL "" FORCE)
set(protobuf_BUILD_SHARED_LIBS OFF CACHE BOOL "" FORCE)
set(ONNX_BUILD_TESTS OFF CACHE BOOL "" FORCE)
set(EIGEN_BUILD_TESTING OFF CACHE BOOL "" FORCE)
 
include(FetchContent)
FetchContent_Declare(MLask
    GIT_REPOSITORY https://gitlab.com/Kamzie1/MLask.git
    GIT_TAG 1.0.3
    GIT_SHALLOW ON
    EXCLUDE_FROM_ALL
    SYSTEM)
FetchContent_MakeAvailable(MLask)

(back to top)

Usage

Defining a model

You need to create a model with in, out, parameters. Optionally you can add number of layers, and whether to draw progress bar while learning. Then you add layers.

Model model(1, 1, 1,); // input of size 1, output of size 1, 1 layer
// You can add specific mlask layers
model.addFullyConnectedLayer<1, 1>(); // Correct output->input flow is validated
model.addActivationFunction(InternalActivationFunction::Relu); //Only builtin activation functions listed in InternalActivationFunctioon enum
model.addLambdaActivationFunction([](float_t x){ return 0.5*x*x; }, [](float_t x){ return x; } ); // A layer that should be used only for prototyping.

Adding your own layer

You can define and add your own layer. All you need to do is create a class that derives from abstract Layer class.

class CustomLayer : public Layer{
public:
    // You need to override these functions. However if you want your layer to be exportable to ONNX for example, you also need to override tryConvertToONNX.
    vectorOut forward(vectorIn input) override;
    vectorIn backward(vectorOut error) override;
    void fit(float_t learning_rate) override {}
};
 
model.addLayer<CustomLayer>();// You can add your custom layer to the model like this.
model.addLayer(std::make_unique<CustomLayer>()); // or like this

Training the model

float_t learning_rate = 0.001;
for(std::size_t epochs=0; epochs < EPOCHS; epochs++){
  for(int x = 0;x<=SIZE;x++){
    model.backprop<DerivedStandardMean>(vectorIn{{(x-(SIZE/2.f))/(SIZE/2.f)}}, vectorOut{{Y[x]}}); // in arguments we construct an Eigen vector using mlask predefined type for simplicity
  }
  model.fit(learning_rate); // if your model has log set to true and it has correct number of epochs, then on fit it will update progress bar.
}

Exporting the model

std::cout<<model.str();
model.exportToONNX("onnx_format.onnx", "One Neuron Neural Network");

For more examples, please refer to the Documentation

(back to top)

License

Distributed under the MIT. See LICENSE for more information.

(back to top)