GitPlanet

Cheap and reliable Node.js hosting starts at $3/month, and $1/month static HTML hosting

Created with love in Canada, visit hostnodejs.com today

Feel like to post an Ad? Learn Details
All Projects → changwoolee → Lenet5_hls

changwoolee / Lenet5_hls

Licence: mit
FPGA Accelerator for CNN using Vivado HLS

Labels

cnnhlsaccelerator

Projects that are alternatives of or similar to Lenet5 hls

Tensorflow template application
TensorFlow template application for deep learning
Stars: ✭ 1,851 (+1008.38%)
Mutual labels:  cnn
Weixinxiaochengxu
小帅一点资讯微信小程序图像识别源码，微信小程序百度AI接口源码，微信小程序图片上传显示缩放缩略图,人工智能，图像识别，人脸颜值分析，植物、动物、车型、LOGO、食材、手写文字识别等
Stars: ✭ 162 (-2.99%)
Mutual labels:  cnn
Cnn captcha
use cnn recognize captcha by tensorflow. 本项目针对字符型图片验证码，使用tensorflow实现卷积神经网络，进行验证码识别。
Stars: ✭ 2,273 (+1261.08%)
Mutual labels:  cnn
P2p Cdn Sdk Javascript
Free p2p cdn github javascript sdk to reduce video streaming costs of live and on demand video using webrtc by upto 90% and improve scalability by 6x - 🚀 Vadootv 🚀
Stars: ✭ 158 (-5.39%)
Mutual labels:  hls
Anime Face Gan Keras
A DCGAN to generate anime faces using custom mined dataset
Stars: ✭ 161 (-3.59%)
Mutual labels:  cnn
Iresnet
Improved Residual Networks (https://arxiv.org/pdf/2004.04989.pdf)
Stars: ✭ 163 (-2.4%)
Mutual labels:  cnn
Documentclassification
This code implements a simple CNN model for document classification with tensorflow.
Stars: ✭ 151 (-9.58%)
Mutual labels:  cnn
Eeg Dl
A Deep Learning library for EEG Tasks (Signals) Classification, based on TensorFlow.
Stars: ✭ 165 (-1.2%)
Mutual labels:  cnn
Tf Adnet Tracking
Deep Object Tracking Implementation in Tensorflow for 'Action-Decision Networks for Visual Tracking with Deep Reinforcement Learning(CVPR 2017)'
Stars: ✭ 162 (-2.99%)
Mutual labels:  cnn
Person Reid Triplet Loss
Person re-ID baseline with triplet loss
Stars: ✭ 165 (-1.2%)
Mutual labels:  cnn
Simple cnn
Simple Convolutional Neural Network Library
Stars: ✭ 158 (-5.39%)
Mutual labels:  cnn
Pytorch Receptive Field
Compute CNN receptive field size in pytorch in one line
Stars: ✭ 160 (-4.19%)
Mutual labels:  cnn
Fpga readings
Recipe for FPGA cooking
Stars: ✭ 164 (-1.8%)
Mutual labels:  hls
Shaka Player Embedded
Shaka Player in a C++ Framework
Stars: ✭ 153 (-8.38%)
Mutual labels:  hls
Handpose
A python program to detect and classify hand pose using deep learning techniques
Stars: ✭ 168 (+0.6%)
Mutual labels:  cnn
Mamba
Mamba is a Swift iOS, tvOS and macOS framework to parse, validate and write HTTP Live Streaming (HLS) data.
Stars: ✭ 151 (-9.58%)
Mutual labels:  hls
Mstar deeplearning project
Radar target classification, detection and recognition using deeplearning methods on MSTAR dataset
Stars: ✭ 163 (-2.4%)
Mutual labels:  cnn
Kaggle Competition Favorita
5th place solution for Kaggle competition Favorita Grocery Sales Forecasting
Stars: ✭ 169 (+1.2%)
Mutual labels:  cnn
Cnnimageretrieval
CNN Image Retrieval in MatConvNet: Training and evaluating CNNs for Image Retrieval in MatConvNet
Stars: ✭ 168 (+0.6%)
Mutual labels:  cnn
Mediasdk
The library is working for downloading video while playing the video, the video contains M3U8/MP4
Stars: ✭ 164 (-1.8%)
Mutual labels:  hls
View All Similar Projects ➔

LeNet-5 in HLS

This repository is about my graduate report, implementing LeNet-5 in Vivado High Level Synthesis 2016.4 & Vivado SDSoC 2016.4

lenet5

Win 10 Test App

You can test the accelerator by your own handwritten digits image.

Youtube Video

Youtube Video Here

If you want to test the app, follow these instruction

  1. Configure the IP address of Zedboard.
	[email protected]:~# ifconfig
  1. Start .elf file with port name argument (in here, 5555 is port name)
	[email protected]:~# lenet5_test.elf 5555
  1. Start the win 10 test application and input the IP address & port name.
  2. Press connect
  3. Open image file

I did not put a zoom in/out function to the app, so please suit the image size.

Model description

Used model is LeNet5-Like Deep CNN
Input : -1.0 to 1.0
Conv1 : 1x32x32 -> 6x28x28, ksize = 1x6x5x5, stride = 1
Pool1 : 6x28x28 -> 6x14x14, average pooling, window size = 2x2, stride = 2
Conv2 : 6x14x14 -> 16x10x10, ksize = 6x16x25, stride = 1
Pool2 : 16x10x10 -> 16x5x5, average pooling, window size = 2x2, stride = 2
Conv3 : 16x5x5 -> 120x1x1, ksize = 16x120x25, stride = 1
FC1 : 120x84
FC2 : 84x10

Environments

I used Zedboard(Zynq 7z020) for testing.

HW Functions : CONVOLUTION_ LAYER_ 1, CONVOLUTION_ LAYER_ 2, and CONVOLUTION_ LAYER_ 3, Clk freq set as 100MHz.

Accuracy

SW accuracy : 98.63% (single precision fp)    
HW accuracy : 98.63% (single precision fp)

Runtime

# of images : 10,000, batch size : 1  

SW runtime  : 59.4456 seconds  
HW runtime  : 16.3954 seconds  

speedup : x3.63 faster

Contributors

  • Changwoo Lee (Hanyang University, Seoul, South Korea)
  • Jeonghyun Woo (Hanyang University, Seoul, South Korea)
Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].