simple_DNN

Introduction

This project aims to realize a simple 3-layer DNN with C++ and recognize the handwritten digits. What's more, I wanna implement codes with Vivado HLS and optimize it for specified targets. I'll use FPGA xc7z010clg400-1 provided by Xilinx.

Realize a simple 3-layer DNN with C++ and recognize the handwritten digits;
Implement codes with Vivado HLS and optimize it for specified targets.

Targets

•Use xc7z010clg400-1.
Accuracy 90%
Clock Period < 7 ns
Resource utilization DSP48E ≤ 80
Throughput Interval < 150

Structure

dnn.cpp: include the DNN code;
dnn_test.cpp: include the test bench, which is used for reading data, calling the main function and calculating the accuracy;
dnn.h: header file included by dnn.cpp and dnn_test.cpp;
W1.txt, Wout.txt: the weights for DNN, data type is float;
b1.txt, bout.txt: the bias for DNN, data type is float;
testImage folder: include 100 14*14 images in txt format, the label.txt includes the correct labels for these images.

Optimize

<ap_fixed> library provided by HLS can quantize the data so as to reduce the resource utilization and interval;
- typedef ap_fixed<10,5> FIXED;
- The top function dnn set as HLS DATAFLOW;
All inputs such as input_image, w_i and b_i, all set as HLS ARRAY_PARTITION complete dim = i, i=1,2;
Intermediate arrays such as hidden, output and test also set as HLS ARRAY_PARTITION complete dim = 1, and initialization all set as HLS_UNROLL;
Loops with multiply set as HLS PIPELINE enable_flush rewind and HLS_UNROLL factor = 2;
Loops without multiply set as HLS_UNROLL.

Results

Accuracy 93%
Clock Period 6.88 ns
Resource utilization DSP48E 74
Throughput Interval 98