SophieXin9636 / Convolution-Neural-Network-in-systemC

Convolution Neural Network implemented by C++ library "SystemC"

CNN using in systemC

Outline

[TOC]

Step

1. file input 28*28 pixel of image (need to divide to 256.0 to floating point).
   • load 5*5 kernel (or called filter) data from ROM.
2. First Convolution
   • Stride: 1
   • 28*28
   • Create 24*24(matrix) * 6 (channel) Convolved feature.
   • store(write) into RAM (index: 0~3455).
   • Adding bias.
   • Normalisation: using Activation function $ReLU(x) = max(x,0)$
3. First Pooling
   • Stride: 2
   • Load First Convulution layer data (24*24)*6 from RAM.
   • Create 12*12*6 filter and store(write) into RAM (index: 3456~4319).
4. Second Convolution
   • load 5*5*6(depth) kernel (or called filter) data from ROM.
   • Stride: 1
   • Depth: 6
   • Create 8*8 (matrix) * 16 (Depth or called channel) Convolved feature.
   • Store(write) into RAM (index: 4320~5343)
   • Adding bias
   • Normalisation: using Activation function $ReLU(x) = max(x,0)$
5. Second Pooling
   • Stride: 2
   • Load First Convulution layer data (8*8)*16 from RAM. (index: 4320~5343)
   • Create 4*4*16 filter and store(write) into RAM (index: 5344~5599).
6. First fully connected layer
   • Load Second Pooling layer data from RAM. (index: 5344~5599)
   • Write 128 neuron into RAM (index 5600~5727)
7. Second fully connected layer
   • Load First fully connected layer neuron data from RAM. (index: 5600~5727)
   • Write 84 neuron into RAM (index 5728~5811)
8. Third fully connected layer
   • Load Second fully connected layer neuron data from RAM. (index: 5728~5811)
   • write into Result to Monitor block, and it will output result.

System Block

執行結果

• MODE = 1 (floating), File = 1

• MODE = 1 (floating), File = 2

• MODE = 2 (floating), File = 1

• pct (PA)