lenet and caffe-lenet

方法1：代码测试

下面的代码测试caffe-lenet中权重w,b以及bottom,up数据大小变化的情况。理解caffe-lenet中的具体实现。并对mnist数据集进行了测试。对于caffe生成的mean.binaryproto文件其在matlab中同样遵循[width * height *c]以及bgr模式。

实现需要的文件:

mnist数据集的 ：mean.binaryproto文件,mnist_test 和mnist_test_label。

clearvars;close all;if exist('../+caffe', 'dir')    addpath('..');else    error('Please run this demo from caffe/matlab/demo');end%% load mnist_test_data and mean_file% width * height * c  and bgrmean_file = caffe.io.read_mean('mean.binaryproto');load mnist_test.mat; %  mnist_testload mnist_test_labels.mat; % mnist_test_labelstnum = 6; % test_numtrue_label = mnist_test_labels(1:tnum)';test_img = reshape(mnist_test,[size(mnist_test,1) 28 28]);test_img = permute(test_img,[3 2 1]); %交互三个维度。test_img = test_img(:,:,1:tnum); % 可显示化的,height *　width * nsamples% convert [h * w * n] to [w * h * n]test_img = permute(test_img,[2 1 3]);% 减去均值文件test_img = bsxfun(@minus,test_img,mean_file);% 尺度化test_img = test_img * 0.00390625; % 除以 255%convert [w * h * n] to [w * h * c * n]test_img = reshape(test_img,[size(test_img,1) size(test_img,2) 1 size(test_img,3)]);%% 预测net_model = 'lenet.prototxt';net_weights = '1lenet_iter_10000.caffemodel';phase = 'test';net = caffe.Net(net_model, net_weights, phase);% input_feature must be a cellscores = net.forward({test_img});scores = scores{1};[~,max_label] = max(scores);pre_label = max_label - 1; % 0 1 2 3 4 5 6 7 8 9disp('pre_label:');disp(pre_label);disp('true_label:');disp(true_label);%% 将层的权重和偏移，以及输入，输出数据的大小打印出来。% conv1 ==> bottom: "data"  top: "conv1"% pool1 ==> bottom: "conv1" top: "pool1"% conv2 ==> bottom: "pool1" top: "conv2"% pool2 ==> bottom: "conv2" top: "pool2"% ip1   ==> bottom: "pool2" top: "ip1"% relu1 ==> bottom: "ip1"   top: "ip1% ip2   ==> bottom: "ip1"   top: "ip2"% prob  ==> bottom: "ip2"   top: "prob"params = cell(8,3);params{1,1} = 'conv1'; params{1,2} = 'data';  params{1,3} = 'conv1';params{2,1} = 'pool1'; params{2,2} = 'conv1'; params{2,3} = 'pool1';params{3,1} = 'conv2'; params{3,2} = 'pool1'; params{3,3} = 'conv2';params{4,1} = 'pool2'; params{4,2} = 'conv2'; params{4,3} = 'pool2';params{5,1} = 'ip1';   params{5,2} = 'pool2'; params{5,3} = 'ip1';params{6,1} = 'relu1'; params{6,2} = 'ip1';   params{6,3} = 'ip1';params{7,1} = 'ip2';   params{7,2} = 'ip1';   params{7,3} = 'ip2';params{8,1} = 'prob';  params{8,2} = 'ip2';   params{8,3} = 'prob';matlab_show = true;% matlab中 权重w,偏置b，bottom和top数据格式如下,group = 1(default):% w = [width * height * c * num_output1]  [width * height * num_output1 * num_output2] ...% b = [num_output 1];% bottom and top : [width * height * c * nsamples] [width * height *　num_output1 * nsamples]...disp('------------------------------------------------');for i = 1:size(params,1)    if strcmp(params{i,1}(1:end-1),'conv')  ||  strcmp(params{i,1}(1:end-1),'ip')        conv1_layer_w = net.layers(params{i,1}).params(1).get_data();        conv1_layer_b = net.layers(params{i,1}).params(2).get_data();        if matlab_show == false            % convert [width * height * c * n] to [n * c * height * witdh]            conv1_layer_w = permute(conv1_layer_w,[4 3 2 1]);        end    elseif strcmp(params{i,1}(1:end-1),'pool')        %conv1_layer_w = net.layers(params{i,1}).params(1).get_data();    end    bottom  = net.blobs(params{i,2}).get_data();    top = net.blobs(params{i,3}).get_data();    if matlab_show == false        % convert [width * height * c * n] to [n * c * height * witdh]        bottom = permute(bottom,[4 3 2 1]);        top = permute(top,[4 3 2 1]);    end    if strcmp(params{i,1}(1:end-1),'conv')  || strcmp(params{i,1}(1:end-1),'ip')        disp([params{i,1} '_layer: w: [' num2str(size(conv1_layer_w)) '] ;b: [' num2str(size(conv1_layer_b)) '] ']);    end    disp([ params{i,2} ': [' num2str(size(bottom)) '] ;' params{i,3} ': [' num2str(size(top)) ']']);    disp('------------------------------------------------');end

输出：

输出结果给出了测试的6个样本的预测标签和正确标签。以及各个层的参数大小和数据的大小。打印的方式采用matlab数据的存储方式。我们可以看到经过池化层后的数据为：pool2 : [4 4 50 6]，紧接着是全连接层ip1_layer: w :[800 500]，这里在prototxt中设置的num_output为500。我们可以看到： 800 = 4 * 4 * 50 ，因此是所有神经元都进行了连接，而没有像很多博客给的文中全连接层是通过进一步的卷积核(例如5*5（当然这里肯定不行)或者其他大小的卷积核)卷积实现的。

方法二：可视化工具

将下面的test协议复制到在线可视化工具，可以看到每个blob的大小。

name: "LeNet"layer {  name: "data"  type: "Input"  top: "data"  input_param { shape: { dim: 6 dim: 1 dim: 28 dim: 28 } }}layer {  name: "conv1"  type: "Convolution"  bottom: "data"  top: "conv1"  param {    lr_mult: 1  }  param {    lr_mult: 2  }  convolution_param {    num_output: 20    kernel_size: 5    stride: 1    weight_filler {      type: "xavier"    }    bias_filler {      type: "constant"    }  }}layer {  name: "pool1"  type: "Pooling"  bottom: "conv1"  top: "pool1"  pooling_param {    pool: MAX    kernel_size: 2    stride: 2  }}layer {  name: "conv2"  type: "Convolution"  bottom: "pool1"  top: "conv2"  param {    lr_mult: 1  }  param {    lr_mult: 2  }  convolution_param {    num_output: 50    kernel_size: 5    stride: 1    weight_filler {      type: "xavier"    }    bias_filler {      type: "constant"    }  }}layer {  name: "pool2"  type: "Pooling"  bottom: "conv2"  top: "pool2"  pooling_param {    pool: MAX    kernel_size: 2    stride: 2  }}layer {  name: "ip1"  type: "InnerProduct"  bottom: "pool2"  top: "ip1"  param {    lr_mult: 1  }  param {    lr_mult: 2  }  inner_product_param {    num_output: 500    weight_filler {      type: "xavier"    }    bias_filler {      type: "constant"    }  }}layer {  name: "relu1"  type: "ReLU"  bottom: "ip1"  top: "ip1"}layer {  name: "ip2"  type: "InnerProduct"  bottom: "ip1"  top: "ip2"  param {    lr_mult: 1  }  param {    lr_mult: 2  }  inner_product_param {    num_output: 10    weight_filler {      type: "xavier"    }    bias_filler {      type: "constant"    }  }}layer {  name: "prob"  type: "Softmax"  bottom: "ip2"  top: "prob"}