Win 7下MatConvNet使用DAG网络方法记录

1、修改ResNet网络

function net = resnet_single() 
net = load('imagenet-resnet-50-dag.mat') ; 
net = dagnn.DagNN.loadobj(net) ;

net.removeLayer('fc1000') ; 
net.removeLayer('prob') ;

dropoutBlock = dagnn.DropOut('rate',0.9); 
net.addLayer('dropout',dropoutBlock,{'pool5'},{'pool5d'},{});

fc8Block = dagnn.Conv('size',[1 1 2048 8],'hasBias',true,'stride',[1,1],'pad',[0,0,0,0]); 
net.addLayer('fc8',fc8Block,{'pool5d'},{'prediction'},{'fc8f','fc8b'});

lossBlock = dagnn.Loss('loss', 'softmaxlog'); 
net.addLayer('softmaxloss',lossBlock,{'prediction','label'},'objective');

net.addLayer('top1err', dagnn.Loss('loss', 'classerror'), ... {'prediction','label'}, 'top1err') ;

2、构建Siamese网络

net1 = resnet_single();
net1 = net1.saveobj() ;

net2 = net1;
for i = 1 : length(net2.layers); 
    net2.layers(i).name = strcat(net2.layers(i).name,'_2'); 
    for j = 1 : length(net2.layers(i).inputs);
        net2.layers(i).inputs{j} = strcat(net2.layers(i).inputs{j},'_2'); 
    end;
    for j = 1 : length(net2.layers(i).outputs);
        net2.layers(i).outputs{j} = strcat(net2.layers(i).outputs{j},'_2');
    end;
end;
for i = 1 : length(net2.vars); 
    net2.vars(i).name = strcat(net2.vars(i).name,'_2'); 
end;

net1.vars = [net1.vars,net2.vars];
net1.layers = [net1.layers,net2.layers];

net = dagnn.DagNN.loadobj(net1) ;

% Siamese Network Loss
net.addLayer('SiameseLoss',dagnn.SiameseLoss(), {'pool5','pool5_2'}, {'objective_s'});

3、构建TripletLoss网络

net1 = resnet_single();
net1 = net1.saveobj() ;

net2 = net1;
for i = 1 : length(net2.layers); 
    net2.layers(i).name = strcat(net2.layers(i).name,'_2'); 
    for j = 1 : length(net2.layers(i).inputs);
        net2.layers(i).inputs{j} = strcat(net2.layers(i).inputs{j},'_2'); 
    end;
    for j = 1 : length(net2.layers(i).outputs);
        net2.layers(i).outputs{j} = strcat(net2.layers(i).outputs{j},'_2');
    end;
end;
for i = 1 : length(net2.vars); 
    net2.vars(i).name = strcat(net2.vars(i).name,'_2'); 
end;

net3 = net1;
for i = 1 : length(net3.layers); 
    net3.layers(i).name = strcat(net3.layers(i).name,'_3'); 
    for j = 1 : length(net3.layers(i).inputs);
        net3.layers(i).inputs{j} = strcat(net3.layers(i).inputs{j},'_3'); 
    end;
    for j = 1 : length(net3.layers(i).outputs);
        net3.layers(i).outputs{j} = strcat(net3.layers(i).outputs{j},'_3');
    end;
end;
for i = 1 : length(net3.vars); 
    net3.vars(i).name = strcat(net3.vars(i).name,'_3'); 
end;

net1.vars = [net1.vars,net2.vars,net3.vars];
net1.layers = [net1.layers,net2.layers,net3.layers];

net = dagnn.DagNN.loadobj(net1) ;

% Triplet Network Loss
net.addLayer('TripletLoss',dagnn.TripletLoss('margin',0.1),{'pool5','pool5_2','pool5_3'},{'objective_triplet'});