Parallel Programming - Lenet 5 CUDA

Parallel Programming Final Project #HCMUS #CNN #LeNet-5

Member

Name	MSSV
Hoàng Hữu Minh An	20127102
Trần Tiến Hoàng	20127424
Nguyễn Tấn Phát	20127588

Introduction

In this final project, you will be implementing and optimizing the forward-pass of a convolutional layer using CUDA. Convolutional layers are the primary building blocks of convolutional neural networks (CNNs), which are used in many machine learning tasks like image classification, object detection, natural language processing, and recommendation systems. In general, CNNs work well on tasks where the data/input features have some level of spatial relationship.

Your optimized CUDA implementation of the convolutional layer will be used to perform inference for layers C1 and C3.

We can use mini-dnn-cpp (Mini-DNN) framework as a starting point to implement the modified version of LeNet-5.

We will use the Fashion MNIST dataset, where the inputs to the network will be a single channel images with dimensions of 28 x 28 pixels. The output layer consists of 10 nodes, where each node represents the likelihood of the input belonging to one of the 10 classes (T-shirt, dress, sneaker, boot etc.)

The overall learning objectives for this project are:

• Demonstrating command of CUDA and optimization approaches by designing and implementing an optimized neural-network convolutional layer forward pass

Test model

Video hướng dẫn chạy demo: Youtube

Linux

git clone https://github.com/hhman24/parallel-programming-leNet-5
cd parallel-programming-leNet-5
make clean
make setup
make custom0
make test_model

Train model

git clone https://github.com/hhman24/parallel-programming-leNet-5
cd parallel-programming-leNet-5
make clean
make setup
make custom0
make train_model

Background knowledge

1 . Convolution Neural Networks (CNNs)

• Standford cheatsheat

• Video CNNs: "How Convolutional Neural Networks work"

2 . Lenet-5

• A Complete Guide: here

Dataset

MNIST dataset, network will be a single channel images with dimensions of 28 x 28 pixels