arhat

README

Arhat Prequel

This repository contains code of Arhat Prequel, an experimental deep learning framework developed at the early stage of Arhat project. We have used Arhat Prequel as a pure research tool for experimenting with various engineering concepts. Upon completion of these experiments, Arhat Prequel development has been discontinued and definitive implementation of Arhat project has been started from scratch.

We have published the Arhat Prequel code in hope that it might be of interest to the community as a rare example of a deep learning framework implemented in Go programming language.

Unlike most mainstream frameworks that perform training and inference computations directly, Arhat Prequel translates neural network descriptions into standalone lean executable platform-specific code. This approach allows direct embedding of the generated code into user applications and does not require deployment of sophisticated machine learning software stacks on the target platforms.

Arhat Prequel supports swappable platform-specific code generators (backends). Currently two backends are available:

• CPU: a reference backend generating C++ code for use on the CPU;
• CUDA: a backend generating C++/CUDA code for use on the NVIDIA GPU.

Implementation of more backends for other platforms is planned.

Arhat Prequel implementation is based on neon, Intel� Nervana� reference deep learning framework. We have ported sections of neon code from Python to Go and partly redesigned it implementing code generators in place of the original backends.

Requirements

The following hardware and software components are required for using Arhat Prequel:

• OS: Windows or Linux (macOS should suit too but was not yet tested);
• NVIDIA GPU device (Kepler of newer);
• CUDA Toolkit (8.0 or higher);
• C++ compiler toolchain supporting CUDA;
• Go (1.11 or higher);
• Python (only for obtaining data sets for examples).

Obtaining data sets

Examples included in this distribution require two data sets:

• MNIST
• CIFAR10

These data sets can be obtained and converted to Arhat Prequel format using the supplied Python scripts:

• get_mnist.py
• get_cifar10.py

Runtime libraries

Arhat Prequel provides few run-time libraries required to build executables from the generated code:

• runtime: runtime components common for all backends;
• runtime_cpu: runtime components required for the CPU backend;
• runtime_cuda: runtime components required for the CUDA backend.

These libraries must be built and linked with the respective generated code.

Quick start

To start, download this distribution, obtain the data sets, and build the runtime libraries. Set your GOPATH environment variable to the path of the downloaded go directory.

Create an empty working directory, make it you current directory, create a subdirectory named data and copy the data sets there. Then build the mnist_mlp example implementing a simple multi-layer perceptron network for MNIST data set. Use the following commands (Linux is assumed, commands for Windows are similar).

For the CUDA backend:

mkdir -p bin
mkdir -p mnist_mlp_cuda
go build -o bin/mnist_mlp fragata/arhat/examples/mnist_mlp
bin/mnist_mlp -o mnist_mlp_cuda

For the CPU backend:

mkdir -p bin
mkdir -p mnist_mlp_cpu
go build -o bin/mnist_mlp fragata/arhat/examples/mnist_mlp
bin/mnist_mlp -b cpu -o mnist_mlp_cpu

The generated code will be placed in subdirectories mnist_mlp_cuda and mnist_mlp_cpu respectively. Separately for each directory, compile all the contained source files and link resulting object files with the runtime libraries. Run the resulting executables from your current directory to train and evaluate the neural network.

License

We are releasing Arhat Prequel under an open source Apache 2.0 License.