ModularMLP

ModularMLP is a Java library for building and training Multi-Layer Perceptrons (MLPs) from scratch with a fully modular design. The primary goal of this implementation is to prioritize simplicity and ease of use for the end user, allowing the construction and training of complex MLPs with minimal code.

This library is for educational purposes only and shouldn't be considered for big networks. More specially, the matrix implementation is particularly inefficient and doesn't uses standard java librairies such as EJML or ND4J.

Features

• Fully configurable MLPs: number of layers, neurons per layer, activation functions.
• Trainers with support for different optimizers (Adam, SGD, etc.) and loss functions (Cross-Entropy, MSE…).
• Batch support and training on custom datasets.
• Optional regularization: L1, L2, ElasticNet.
• Implemented from scratch, no external dependencies.
• Modular design: every component (layer, optimizer, dataset, trainer) can be swapped or extended easily.

Usage example

// Create the trainer
Trainer mnistTrainer = Trainer.builder() 
        .setLossFunction(CE) 
        .setOptimizer(new Adam(0.001, 0.99, 0.999))
        .setDataset(mnistDataset)
        .setEpoch(30)
        .setBatchSize(7_000)
        .build();

// Build and train the MLP
MLP mnistMLP = MLP.builder(784)
        .setRandomSeed(420)
        .addLayer(256, ReLU)
        .addLayer(128, ReLU)
        .addLayer(10, SoftMax)
        .build()
        .train(mnistTrainer);

Installation

Clone the repository:

git clone https://github.com/yro7/ModularMLP.git
cd ModularMLP

Library overhead

Here's the call tree of the project, when training a basic MNIST mlp resolver. As expected, the matrix multiplication takes ~95% of CPU time.

Note that the application only ran for a few minutes, JIT by the JVM might change results while training larger models.