请问利用label和高斯分布生成一个 label distribution的方法都在哪里？matlab深度学习代码太不熟悉了

[Sean]

请问利用label和高斯分布生成一个 label distribution的方法都在哪里？matlab深度学习代码太不熟悉了

[Bin-Bin Gao]

谢谢关注我们的工作。
年龄估计：

DLDL/Age-Pose/dldl_age.m

Lines 142 to 147 in 0f53197

 case 'MORPH_Album2' 
     img_label  = 1./(sqrt(2*pi)*opts.sigma).*exp(-dif_age.^2./(2*opts.sigma.^2)); 
 case 'ChaLearn' 
     sigma = label(2,:); 
     img_label =1./repmat(sqrt(2*pi)*sigma,ld_num,1).*... 
         exp(-(bsxfun(@minus,opts.label_set',repmat(label(1,:),ld_num,1))).^2./repmat(2*sigma.^2,ld_num,1)); 

头部姿态：

DLDL/Age-Pose/dldl_pose.m

Lines 159 to 187 in 0f53197

 case {'point04','bjut-3d'} 
     labels = []; 
     for i =1 : numel(imdb.images.label) 
          
         dif_cor = bsxfun(@minus, true_label(i,:) , cordin); 
         t = zeros(93,1); 
         if sigma ~= 0 
             t = exp(-0.5*(dif_cor(:,1).*dif_cor(:,1)+dif_cor(:,2).*dif_cor(:,2))./sigma^2); 
         else 
             t(imdb.images.label(i)) = 1; 
         end 
         labels(:,i) = t./sum(t); 
         imdb.images.label =  max(min(labels,1-10^-15),10^-15); 
     end 
 case 'aflw_det' 
     cordin_yaw = repmat(-90:3:90,61,1); 
     cordin_pitch = repmat([90:-3:-90]',1,61); 
     cordin = [reshape(cordin_yaw,[],1),reshape(cordin_pitch,[],1)];   
      
     for i =1 : numel(true_label(1,:)) 
         dif_cor = bsxfun(@minus, true_label(:,i)' ,cordin); 
         if sigma ~= 0 
             %sigma = 2; 
             t = exp(-0.5*(dif_cor(:,1).*dif_cor(:,1)./(2*sigma^2)+dif_cor(:,2).*dif_cor(:,2)./sigma^2)); 
         else 
             t(imdb.images.label(i)) = 1; 
         end 
         labels(:,i) = t./sum(t); 
     end 

多标签：

DLDL/utils/IFDLDLNet.m

Lines 187 to 192 in 0f53197

 ind = find(labels_ld ==0); 
 labels_ld(ind) = opts.PD; 
 ind = find(labels_ld ==-1); 
 labels_ld(ind) = 0; 
 labels_ld = bsxfun(@times,labels_ld,1./sum(labels_ld)); 
 imdb.images.labels_ld = (1-0.01).*labels_ld + 0.01/20; 

语义分割：

DLDL/Segmentation/dldl_getBatch.m

Lines 54 to 81 in 0f53197

 rgbPath = sprintf(imdb.paths.image, imdb.images.name{images(i)}) ; 
 labelsPath = sprintf(imdb.paths.classSegmentation, imdb.images.name{images(i)}) ; 
  
 rgb = vl_imreadjpeg({rgbPath}) ; 
 rgb = rgb{1} ; 
 if opts.useGpu 
    anno = gpuArray(single(imread(labelsPath))); 
 end 
 h = size(anno,1); 
 w = size(anno,2); 
  
 conv_ld = gpuArray(zeros(h,w,21)); 
 for ci = 0:20 
     feat =single( anno==ci); 
     x = vl_nnconv(feat, kernel, [], ... 
         'pad',2, ... 
         'stride', 1,... 
         'CuDNN' ) ; 
     x (1:2,:) = feat(1:2,:); 
     x(end-1:end,:) = feat(end-1:end,:); 
     x(:,1:2) = feat(:,1:2); 
     x(:,end-1:end) = feat(:,end-1:end); 
     conv_ld(:,:,ci+1) = x; 
 end 
 conv_ld = bsxfun(@times,conv_ld,1./sum(conv_ld,3)); 
 conv_ld(isnan(conv_ld)) = 0; 
 anno_ld = gather(conv_ld); 
 anno_ld(:,:,22) = gather(anno); 

[Sean]

谢谢回复！但是现在我们有了新的处理方案，十分感谢！   离殇 ***@***.***  

…

------------------ 原始邮件 ------------------ 发件人: "Bin-Bin ***@***.***>; 发送时间: 2023年2月22日(星期三) 中午12:30 收件人: ***@***.***>; 抄送: ***@***.***>; ***@***.***>; 主题: Re: [gaobb/DLDL] 请问利用label和高斯分布生成一个 label distribution的方法都在哪里？matlab深度学习代码太不熟悉了 (Issue #2) 谢谢关注我们的工作。 年龄估计：https://github.com/gaobb/DLDL/blob/0f531971dc81d95c3f5e675051ef8d10caa44308/Age-Pose/dldl_age.m#L142-L147 头部姿态：https://github.com/gaobb/DLDL/blob/0f531971dc81d95c3f5e675051ef8d10caa44308/Age-Pose/dldl_pose.m#L159-L187 多标签：https://github.com/gaobb/DLDL/blob/0f531971dc81d95c3f5e675051ef8d10caa44308/utils/IFDLDLNet.m#L187-L192 语义分割： https://github.com/gaobb/DLDL/blob/0f531971dc81d95c3f5e675051ef8d10caa44308/Segmentation/dldl_getBatch.m#L54-L81 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>