Matlab深度学习上手初探
2024-05-21 06:07:12
文章目录
- 简介
- 实例
- 网络设计
- 代码
- 运行结果
- 参考文献
简介
Matlab降低了深度神经网络的开发难度，可以通过拖拽的模式设计网络，甚至训练的过程也是GUI操作。
实例
以高光谱图像分类为例，参考文献1 。构造一个卷积神经网络，输入为9×9×B9\times 9\times B9×9×B 的图像，其中BBB为波段数，类标为中心像素的标签。
网络设计
在Matlab的APPS中搜索Deep Network工具箱，打开后，选择New来创建网络，在弹出的界面中可以选择创建空白网络，也可以选择预训练的网络。
进入设计洁面后，从左侧拖拽相应的模块，命名-->设置参数-->连接不同模块，网络搭建完成后，可以选择Analyze来分析下网络，看看有没有错误，没有错误责可以导出代码。
代码
主训练文件"train_cnn.m"，主要完成加载数据、从图像中随机抽取小的图像块，构造训练集，验证集和测试集。注意，真值变量需要用categorical函数转换一下。
load('../data/WHU_Hi_HongHu_preprocessing_tensor_edgemap_7.mat')rng(2022);% In the experiments, the patch sizes of the three datasets were set as
% 9 × 9 × d, where d denotes the band number of the remote sensing image.
Ntrain = 1000;
Nvalid = 500;
Ntest = 200;
ptcsize = [9, 9];
M = ones(size(Label));
nclass = length(unique(Label));
[X, Y] = sample_patchs(X, Label, M, ptcsize, nclass, Ntrain+Nvalid+Ntest);
Xtrain = X(:, :, :, 1:Ntrain); Ytrain = Y(1:Ntrain);
Xvalid = X(:, :, :, 1:Nvalid); Yvalid = Y(1:Nvalid);
Xtest = X(:, :, :, 1:Ntest); Ytest = Y(1:Ntest);
Ytrain = categorical(Ytrain);
Yvalid = categorical(Yvalid);
Ytest = categorical(Ytest);layers = [imageInputLayer([9 9 270],"Name","imageinput")convolution2dLayer([3 3],128,"Name","conv1")batchNormalizationLayer("Name","batchnorm1")reluLayer("Name","relu1")convolution2dLayer([3 3],256,"Name","conv2")batchNormalizationLayer("Name","batchnorm2")reluLayer("Name","relu2")convolution2dLayer([3 3],256,"Name","conv3","Padding","same")batchNormalizationLayer("Name","batchnorm3")reluLayer("Name","relu3")convolution2dLayer([3 3],128,"Name","conv4")batchNormalizationLayer("Name","batchnorm4")reluLayer("Name","relu4")fullyConnectedLayer(128,"Name","fc1")batchNormalizationLayer("Name","batchnorm5")reluLayer("Name","relu5")fullyConnectedLayer(64,"Name","fc2")batchNormalizationLayer("Name","batchnorm6")reluLayer("Name","relu6")fullyConnectedLayer(nclass,"Name","fc3")softmaxLayer("Name","softmax")classificationLayer("Name", "classoutput")];% plot(layerGraph(layers));options = trainingOptions('adam', ...'ValidationData', {Xvalid, Yvalid}, ...'Plots', 'training-progress', ...'MaxEpochs', 100, ...'Shuffle', 'every-epoch', ...'InitialLearnRate', 1e-3, ...'LearnRateSchedule', 'piecewise', ...'LearnRateDropFactor', 0.1, ...'LearnRateDropPeriod', 20, ...'ExecutionEnvironment', 'gpu', ...'MiniBatchSize', 32);
net = trainNetwork(Xtrain, Ytrain, layers, options);Ptest = classify(net, Xtest);
precision = sum(Ptest==Ytest) / numel(Ptest);disp(precision)随机选图像块文件 “sample_patchs.m”
function [Xp, Yp] = sample_patchs(X, Y, M, ptcsize, nclass, nptcs)
% X: Data image
% Y: Label image
% M: mask: 1: candidate
% ptcsize: size (h, w) of patch
% nclass: number of classes
% nptcs: number of patchsif isempty(ptcsize) ptcsize = [9, 9];
end
if isempty(nptcs)nptcs = 100;
endpH = ptcsize(1);
pW = ptcsize(2);
pH2 = floor(pH / 2.);
pW2 = floor(pW / 2.);
[xH, xW, C] = size(X);M(1:pH2, :) = 0; % boundary
M(xH-pH2:xH, :) = 0; % boundary
M(:, 1:pW2) = 0; % boundary
M(:, xW-pW2:xW) = 0; % boundary[rows, cols] = find(M==1);
npixel = length(rows);idx = randi([1, npixel], nptcs, 1);
idxH = rows(idx);
idxW = cols(idx);Xp = zeros(ptcsize(1), ptcsize(2), C, nptcs);
Yp = zeros(nptcs, 1);
% Yp = zeros(nptcs, nclass); % one-hot
for i = 1:nptcsXp(:, :, :, i) = X(idxH(i) - pH2:idxH(i) + pH2, idxW(i) - pW2:idxW(i) + pW2, :);Yp(i, 1) = Y(idxH(i), idxW(i));% Yp(i, Y(idxH(i), idxW(i)) + 1) = 1; % one-hot
end运行结果
下图为训练过程的日志结果，图中曲线和一些统计信息是Matlab自动绘制的，不需要自己额外添加代码。
此外，Matlab命令窗口也有相应的信息，如下：
>> train_cnn
Initializing input data normalization.
|======================================================================================================================|
| Epoch | Iteration | Time Elapsed | Mini-batch | Validation | Mini-batch | Validation | Base Learning |
| | | (hh:mm:ss) | Accuracy | Accuracy | Loss | Loss | Rate |
|======================================================================================================================|
| 1 | 1 | 00:00:08 | 0.00% | 30.60% | 3.6983 | 2.8994 | 0.0010 |
| 2 | 50 | 00:00:11 | 75.00% | 64.80% | 1.0919 | 1.2310 | 0.0010 |
| 4 | 100 | 00:00:14 | 65.62% | 65.20% | 0.9713 | 1.0583 | 0.0010 |
| 5 | 150 | 00:00:17 | 62.50% | 74.80% | 1.1589 | 0.8747 | 0.0010 |
| 7 | 200 | 00:00:19 | 62.50% | 73.00% | 0.9210 | 0.8468 | 0.0010 |
| 9 | 250 | 00:00:23 | 78.12% | 76.60% | 0.6505 | 0.7860 | 0.0010 |
| 10 | 300 | 00:00:25 | 78.12% | 77.80% | 0.7985 | 0.7317 | 0.0010 |
| 12 | 350 | 00:00:28 | 81.25% | 80.00% | 0.6691 | 0.6691 | 0.0010 |
| 13 | 400 | 00:00:31 | 71.88% | 80.20% | 0.9969 | 0.6473 | 0.0010 |
| 15 | 450 | 00:00:34 | 87.50% | 80.20% | 0.4374 | 0.6442 | 0.0010 |
| 17 | 500 | 00:00:37 | 84.38% | 81.20% | 0.4327 | 0.6272 | 0.0010 |
| 18 | 550 | 00:00:39 | 84.38% | 83.80% | 0.3872 | 0.5438 | 0.0010 |
| 20 | 600 | 00:00:42 | 81.25% | 83.00% | 0.6669 | 0.5028 | 0.0010 |
| 21 | 650 | 00:00:45 | 81.25% | 86.40% | 0.4656 | 0.4147 | 0.0001 |
| 23 | 700 | 00:00:48 | 78.12% | 88.00% | 0.6784 | 0.3880 | 0.0001 |
| 25 | 750 | 00:00:51 | 96.88% | 88.40% | 0.2379 | 0.3900 | 0.0001 |
| 26 | 800 | 00:00:53 | 93.75% | 88.20% | 0.3173 | 0.4199 | 0.0001 |
| 28 | 850 | 00:00:56 | 87.50% | 89.00% | 0.3716 | 0.3864 | 0.0001 |
| 30 | 900 | 00:00:59 | 87.50% | 89.20% | 0.3112 | 0.3499 | 0.0001 |
| 31 | 950 | 00:01:01 | 81.25% | 90.60% | 0.4589 | 0.3472 | 0.0001 |
| 33 | 1000 | 00:01:04 | 90.62% | 90.20% | 0.2410 | 0.3030 | 0.0001 |
| 34 | 1050 | 00:01:07 | 96.88% | 91.00% | 0.2589 | 0.3052 | 0.0001 |
| 36 | 1100 | 00:01:10 | 84.38% | 92.00% | 0.5322 | 0.2920 | 0.0001 |
| 38 | 1150 | 00:01:12 | 96.88% | 91.20% | 0.2072 | 0.2998 | 0.0001 |
| 39 | 1200 | 00:01:15 | 90.62% | 92.20% | 0.2447 | 0.2759 | 0.0001 |
| 41 | 1250 | 00:01:18 | 93.75% | 92.00% | 0.1627 | 0.2724 | 1.0000e-05 |
| 42 | 1300 | 00:01:20 | 96.88% | 92.40% | 0.1265 | 0.2751 | 1.0000e-05 |
| 44 | 1350 | 00:01:23 | 93.75% | 90.80% | 0.1679 | 0.3054 | 1.0000e-05 |
| 46 | 1400 | 00:01:26 | 96.88% | 93.40% | 0.1650 | 0.2544 | 1.0000e-05 |
| 47 | 1450 | 00:01:29 | 93.75% | 92.20% | 0.2000 | 0.2709 | 1.0000e-05 |
| 49 | 1500 | 00:01:32 | 93.75% | 92.40% | 0.1877 | 0.2520 | 1.0000e-05 |
| 50 | 1550 | 00:01:34 | 93.75% | 92.20% | 0.1618 | 0.2842 | 1.0000e-05 |
| 52 | 1600 | 00:01:37 | 93.75% | 91.80% | 0.3416 | 0.2809 | 1.0000e-05 |
| 54 | 1650 | 00:01:40 | 96.88% | 91.60% | 0.1159 | 0.2628 | 1.0000e-05 |
| 55 | 1700 | 00:01:43 | 90.62% | 94.00% | 0.2882 | 0.2346 | 1.0000e-05 |
| 57 | 1750 | 00:01:46 | 93.75% | 93.00% | 0.1924 | 0.2571 | 1.0000e-05 |
| 59 | 1800 | 00:01:48 | 100.00% | 94.40% | 0.0592 | 0.2273 | 1.0000e-05 |
| 60 | 1850 | 00:01:51 | 93.75% | 91.40% | 0.1993 | 0.2669 | 1.0000e-05 |
| 62 | 1900 | 00:01:54 | 87.50% | 91.00% | 0.3692 | 0.2943 | 1.0000e-06 |
| 63 | 1950 | 00:01:57 | 96.88% | 92.80% | 0.2041 | 0.2607 | 1.0000e-06 |
| 65 | 2000 | 00:02:00 | 93.75% | 91.60% | 0.2100 | 0.2653 | 1.0000e-06 |
| 67 | 2050 | 00:02:03 | 87.50% | 92.60% | 0.3792 | 0.2715 | 1.0000e-06 |
| 68 | 2100 | 00:02:06 | 93.75% | 91.80% | 0.1791 | 0.2868 | 1.0000e-06 |
| 70 | 2150 | 00:02:08 | 96.88% | 92.60% | 0.2040 | 0.2728 | 1.0000e-06 |
| 71 | 2200 | 00:02:11 | 90.62% | 93.20% | 0.2053 | 0.2353 | 1.0000e-06 |
| 73 | 2250 | 00:02:14 | 93.75% | 93.60% | 0.2120 | 0.2299 | 1.0000e-06 |
| 75 | 2300 | 00:02:17 | 90.62% | 93.20% | 0.2796 | 0.2405 | 1.0000e-06 |
| 76 | 2350 | 00:02:19 | 93.75% | 92.60% | 0.2731 | 0.2586 | 1.0000e-06 |
| 78 | 2400 | 00:02:22 | 93.75% | 91.80% | 0.1932 | 0.2732 | 1.0000e-06 |
| 80 | 2450 | 00:02:25 | 96.88% | 92.80% | 0.1315 | 0.2484 | 1.0000e-06 |
| 81 | 2500 | 00:02:28 | 93.75% | 93.60% | 0.2221 | 0.2730 | 1.0000e-07 |
| 83 | 2550 | 00:02:31 | 93.75% | 92.20% | 0.1957 | 0.2558 | 1.0000e-07 |
| 84 | 2600 | 00:02:34 | 96.88% | 91.80% | 0.1457 | 0.2807 | 1.0000e-07 |
| 86 | 2650 | 00:02:36 | 87.50% | 93.20% | 0.4540 | 0.2724 | 1.0000e-07 |
| 88 | 2700 | 00:02:39 | 93.75% | 93.40% | 0.2235 | 0.2315 | 1.0000e-07 |
| 89 | 2750 | 00:02:42 | 100.00% | 93.40% | 0.0892 | 0.2506 | 1.0000e-07 |
| 91 | 2800 | 00:02:45 | 93.75% | 92.00% | 0.2005 | 0.2666 | 1.0000e-07 |
| 92 | 2850 | 00:02:48 | 100.00% | 91.20% | 0.1301 | 0.2748 | 1.0000e-07 |
| 94 | 2900 | 00:02:51 | 96.88% | 92.20% | 0.1594 | 0.2691 | 1.0000e-07 |
| 96 | 2950 | 00:02:53 | 93.75% | 93.00% | 0.1665 | 0.2548 | 1.0000e-07 |
| 97 | 3000 | 00:02:56 | 93.75% | 94.00% | 0.2878 | 0.2366 | 1.0000e-07 |
| 99 | 3050 | 00:02:59 | 90.62% | 92.00% | 0.1891 | 0.2761 | 1.0000e-07 |
| 100 | 3100 | 00:03:02 | 93.75% | 92.00% | 0.1937 | 0.2665 | 1.0000e-07 |
|======================================================================================================================|0.9500参考文献
WHU-Hi: UAV-borne hyperspectral with high spatial resolution (H2) benchmark datasets and classifier for precise crop identification based on deep convolutional neural network with CRF ↩︎
Matlab深度学习上手初探相关推荐
- 【MATLAB深度学习工具箱】学习笔记--体脂估计算例再分析：拟合神经网络fitnet里面的数据结构】
原文链接如下 [MATLAB深度学习工具箱]学习笔记--体脂估计Body Fat Estimation_bear_miao的博客-CSDN博客介绍本示例展示一个函数拟合神经网络如何根据解剖学测量结果估 ...
- 【MATLAB深度学习工具箱】学习笔记--体脂估计算例再分析：拟合神经网络fitnet里面的函数】
介绍 上一篇 [MATLAB深度学习工具箱]学习笔记--体脂估计算例再分析:拟合神经网络fitnet里面的数据结构]_bear_miao的博客-CSDN博客原文链接如下[MATLAB深度学习工具箱]学 ...
- 第 09 章 基于特征匹配的英文印刷字符识别 MATLAB深度学习实战案例
基于特征匹配的英文印刷字符识别 MATLAB深度学习实战 话不多讲,直接开撸代码 MainForm函数 function MainForm global bw; global bl; global b ...
- 基于Matlab深度学习Yolov4-tiny的交通标志识别道路标志识别检测
交通标志检测是辅助驾驶.自动驾驶系统中的重要组成部分,针对交通标志检测任务中复杂环境下的小目标检测精度低的问题,提出一种基于YOLOv4-tiny的交通标志检测方法. 基于Matlab深度学习的道路标 ...
- MATLAB深度学习（1） --- 想要做好深度学习？数据集是第一步
MATLAB深度学习(1) --- 想要做好深度学习?数据集是第一步 创作目的 项目简介 本期重点---数据集构建 本文所使用数据集简介 用table来搭建训练集 总结 创作目的 大家好,这里是微信公 ...
- MATLAB深度学习之LSTM
MATLAB深度学习之LSTM 深度学习工具箱 net = trainNetwork(sequences,Y,layers,options) clc clear %% 训练深度学习 LSTM 网络,进 ...
- MATLAB 深度学习部署到树莓派问题解决方法
1.直接按照树莓派 Matlab 深度学习_哔哩哔哩_bilibili部署会出现 arm_compute/runtime/NEON/NEFunctions.h: No such file or dir ...
- (转)Matlab深度学习工具试玩手册一：基本操作与迁移学习
原贴博客:https://blog.csdn.net/zfrycw/article/details/80633979 目录 前言 一.利用现成网络进行分类 二.构建简单的分类网络 三.迁移学习 前言 ...
- 【第 07 章 基于主成分分析的人脸二维码识别MATLAB深度学习实战案例】
基于主成分分析的人脸二维码识别MATLAB深度学习实战案例 人脸库 全套文件资料目录下载链接–>传送门 本文全文源码下载[链接–>传送门] 如下分析: 主文件 function varar ...
- 第 12 章 基于块匹配的全景图像拼接--Matlab深度学习实战图像处理应用
第 12 章 基于块匹配的全景图像拼接–Matlab深度学习实战图像处理应用GUI实现 效果如图所示 完整案例 主函数文件 Gui_Main.m文件 function varargout = Gui_ ...
最新文章
- python浪漫表白代码
- multisim变压器反馈式_变压器的分类及特点
- R语言dplyr包distinct函数去除重复数据行实战
- VS2005工程的Device右边内容为空问题
- 成功解决You are using pip version 9.0.1, however version 9.0.3 is available. You should consider upgra
- erlang精要(1)-四则算术运算
- （一）python3 只需3小时带你轻松入门—— 编程尝试
- asp.net使用post方式action到另一个页面，在另一个页面接受form表单的值！（报错，已解决！）...
- linux用户及权限详解（20170425）
- 苹果6s强制删除id锁_回收宝：苹果6S手机从细节查真假
- python多重继承super父类参数_python – 多重继承如何与super()和不同的__init __()参数一起使用？...
- Springboot记录
- systemtap的一些总结
- 互联网基础知识_互联网广告前景怎么样？
- html查重报告转换,知网查重报告网页版如何转换成PDF和WORD?
- 非零段划分（python）
- android系统更新原理简介
- MATLAB系列笔记：三维绘图（一）
- OCJP 考试题之九
- UE5笔记【十】第一个蓝图项目：bluePrint。