This project aims to build a framework for training and running artificial neural networks using tensors and dynamic computational graphs.

Currently the main purpose is for learning how to implement each of a major framework's (such as TensorFlow) components from scratch.

https://github.com/gokadin/ml-framework/wiki

• mat package with common vector and matrix operations.
• tensor package with dynamic computational graph building.
• dataset package for automatically getting popular datasets from the net or for processing local files.
• modules package implementing the most common ANN layers and activation functions.
• models package for building ANN models.
• runners package for running models and training them.
• GPU support partially there...

The following demonstrates how to train the mnist digits dataset.

// get the dataset
dataset := datasets.From("mnist").SetBatchSize(1000)

// build the model
runner := runners.BuildModelRunner(
		modules.Linear(128),
		modules.Relu(),
		modules.Linear(10))

// configure the model
runner.Configure(runners.ModelConfig{
		Epochs: 50,
		Loss:   modules.LossSoftmaxCrossEntropy})

// run and validate the model
runner.Fit(dataset)
runner.Run(dataset)

This is a personal project with the aim of learning the ANN world and is not intended to be used in any production applications.

However, any contribution is welcome as it will most likely teach someone something new. Feel free to open a PR anytime.

You can also contact me for any questions on how to contribute or just for general comments about the project.