This repository is about our undergraduate graduation project, implementing LeNet-5 by using Vivado High Level Synthesis 2016.4 & Vivado SDSoC 2016.4
You can test the accelerator by your own handwritten digits image.
If you want to test the app, follow these instruction
username@Zedboard:~# ifconfig
username@Zedboard:~# lenet5_test.elf 5555
I did not put a zoom in/out function to the app, so please suit the image size.
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
We used Zedboard(Zynq 7z020) for testing.
HW Functions : CONVOLUTION LAYER 1, CONVOLUTION LAYER 2, and CONVOLUTION LAYER 3, Clk freq set as 100MHz.
SW accuracy : 98.63% (single precision fp)
HW accuracy : 98.63% (single precision fp)
# of images : 10,000, batch size : 1
SW runtime : 59.4456 seconds
HW runtime : 16.3954 seconds
speedup : 3.63x faster