MEDOFLOW: AN EDUCATIONAL SOFTWARE FRAMEWORK FOR DEEP LEARNING

Bai, Hao

Permalink

https://hdl.handle.net/2142/124859 Copy

Description

Title

MEDOFLOW: AN EDUCATIONAL SOFTWARE FRAMEWORK FOR DEEP LEARNING

Author(s)

Bai, Hao

Issue Date

2023-05-01

Keyword(s)

Machine Learning Systems; Software Frameworks for Machine Learning; MLSys; Educational Software Systems.

Abstract

This work proposes MedoFlow: an educational software framework for deep learning. MedoFlow is small in size (2k lines) and completely self-contained, so a user can easily understand how the underlying system works, add components to the existing framework, and modify the source code as they want. MedoFlow utilizes Python as the framework frontend and TVM as the framework backend. The frontend implements the computational graph, auto gradient calculation (autograd) system, shape inference system, and operator abstractions. The backend implements the TVM implementation of a variety of operators such as Matrix Multiplication (MatMul), Convolutional-2D (Conv2D) and Max Pooling-2D (Maxpool2D), to facilitate parallel computation on a set of compilers such as LLVM and CUDA. This framework facilitates to implement complex applications such as LeNet-5 and the Recurrent Neural Network (RNN). To illustrate MedoFlow is comparable to mainstream frameworks, this work conducts comprehensive experiments on a variety of models like Multilayer Perceptron (MLP), LeNet-5 and RNN using PyTorch, TensorFlow and MedoFlow to compare their accuracy and time efficiency. Experimental results show that MedoFlow reaches the same level of accuracy as the other frameworks, and its time efficiency can be significantly improved by implementing small TVM optimizations. The code is publicly available at https://github.com/BiEchi/MedoFlow.

Type of Resource

text

Language

eng

Owning Collections