基于深度信念网络的硬件模拟器研究（Matlab代码实现）

 👨‍🎓个人主页：研学社的博客 

💥💥💞💞欢迎来到本博客❤️❤️💥💥

🏆博主优势：🌞🌞🌞博客内容尽量做到思维缜密，逻辑清晰，为了方便读者。

⛳️座右铭：行百里者，半于九十。

📋📋📋本文目录如下：🎁🎁🎁

目录

💥1 概述

📚2 运行结果

🎉3 参考文献

🌈4 Matlab代码实现

---

💥1 概述

这是一个在 MATLAB 中实现的寄存器传输级 （RTL） 硬件模拟器，用于模拟深度置信网络 （DBN） 的专用硬件系统。在机器学习中，DBN是一种生成图形模型，或者是一种深度神经网络，由多层受限玻尔兹曼机组成。
该硬件系统基于我们所说的神经元机（NM）硬件架构，该架构可以专门用于神经网络系统。通过这个模拟，以及其他模型，我们试图解释我们的系统如何能够实现比CPU和GPU高几个数量级的性能资源比，这意味着更小的芯片面积和功耗，以及使用相同的资源更好的计算速度。尽管源代码是用顺序语言编码的，但它是周期精确的，并以与硬件相同的方式模拟各种 RTL 组件，如寄存器、算术运算符和存储器。读者可以即时执行模拟器并监控其工作情况。

📚2 运行结果

部分代码：

clear all;clf;
% read an image in MATLAB (therefore no train files required)
I = imadjust(imresize(rgb2gray(imread('blason.jpg')),1.15),[0 0.8],[]);
% extract characters from the image
I = 1 - (double(I) / 255);  % inverse image
I3 = I(51:78, 71:98+28*7);  % character area
for i=0:7                   % extract images for 8 characters (AZERGUES)
  DB(i+1,:,1) = reshape(I3(:,(1:28)+28*i-i+(i>4)*4),1,[]);
end;

% uncommenting the following line turns this code for NMIST training.
% it requires additional files from Hinton's site [5].
% converter; makebatches; DB = batchdata; % uncomment this line for MNIST

% set network parameters
Nn = [784 500 500 2000];  % define network size
[Nc Nn(1) Nb] = size(DB); % batch size, pixels, and number of batches
ew = 0.1; eb = 0.1;       % learning rates for weight and bias

% define ob-chip block memories
MW = zeros(60000,64);     % weights
MD = zeros(10000,64);     % weight differences
MM = MW;                  % network topology information (forward)
MR = MW;                  % network topology information (reverse)
MN = MW + 1;              % null connection indicator
MX = zeros(10000,64);     % MX
MB = zeros(1,20000);      % bias
MA = MB;                  % bias differences

% initialize SOT and memories
SOT = zeros(11,12);
moff = 0; woff = 0; roff = 0; % memory offset variables
SOT(1,[1,5]) = [1,Nn(1)];     % set for first stage (training data input)
oi = 2;                       % starting SOT index
for l=1:3     % for each RBM
  hb = ceil(Nn(l) / 64);      % bpn for hidden layer
  hNb = hb * Nn(l+1);         % Nb for hisdden layser
  vb = ceil(Nn(l+1) / 64);    % bpn for visible layer
  vNb = vb * Nn(l);           % Nb for visible layer
  for j = 0:Nn(l+1)-1         % for each visible-hidden neuron pair
    for i=0:Nn(l)-1
      hinx = mod(i+j,hb*64);  % position in the connection space (forward)
      hbi = floor(hinx/64);   % SB index for current neuron (forward)
      snu_i = mod(hinx,64);   % SNU number
      ma = j * hb + hbi + moff + 1;   % MM address
      wa = j * hb + hbi + woff + 1;   % MW address
      MW(wa,snu_i+1) = 0.1 * randn(); % set weight
      MM(ma,snu_i+1) = i;             % set neuron pointer
      MN(ma,snu_i+1) = 0;             % clear null
      vinx = mod(i+j,vb*64);  % position in the connection space (reverse)
      vbi = floor(vinx/64);   % SB index for current neuron (reverse)
      ma = i * vb + vbi + hNb + moff + 1; % MM address
      ra = i * vb + vbi + roff + 1;       % MR address
      MR(ra,snu_i+1) = j * hb + hbi;      % set MR forward weight pointer
      MM(ma,snu_i+1) = j;             % set neuron pointer
      MN(ma,snu_i+1) = 0;             % clear null
    end; 
  end;
  k2 = max(Nn);
  k4 = max(Nn) * 2;
  of1 = mod(l,2) * k2;    % MX read offset 2
  of2 = mod(l+1,2) * k2;  % MX read offset 1
  bo = (l - 1) * k4;      % MBias offset
  % fields in SOT record
  % --------------------
  % 1: stage type 
  % 2: N
  % 3: bpn
  % 4: Nb
  % 5: input (0 = no input)
  % 6: MM offset
  % 7: MX read offset
  % 8: MX wrire offset
  % 9: MR offset
  %10: MW offset
  %11: MBias offset
  %12: MA offset
  % --------------------
  SOT(oi+0,:) = [2,Nn(l+1),hb,hNb,0,    moff,of2,of1,   0,woff,   bo+k2,l*k2];
  SOT(oi+1,:) = [3,  Nn(l),vb,vNb,0,moff+hNb,of1, k4,roff,woff,bo,(l-1)*k2];
  SOT(oi+2,:) = [4,Nn(l+1),hb,hNb,0,    moff, k4,  0,   0,woff,   bo+k2,l*k2];
  oi = oi + 3;
  moff = moff + hNb + vNb;  % adjust offsets
  woff = woff + hNb;
  roff = roff + vNb;
end
SOT(oi-1,5) = Nn(1);  % training data input
SOT(oi,1:2) = [9,2];  % set for last stage (go to)

🎉3 参考文献

部分理论来源于网络，如有侵权请联系删除。

[1]Byungik Ahn (2023). Hardware simulator for a deep belief netrowk system .

🌈4 Matlab代码实现