diff --git a/get_cnn_activations.m b/cnn_shape_get_activations.m
similarity index 74%
rename from get_cnn_activations.m
rename to cnn_shape_get_activations.m
index 5251acd..5b5e9ac 100644
--- a/get_cnn_activations.m
+++ b/cnn_shape_get_activations.m
@@ -1,23 +1,27 @@
-function [ feat ] = get_cnn_activations( im, net, subWins, layers, varargin)
-%GET_CNN_FEATURE Get CNN activation responses for im
+function [ feat ] = cnn_shape_get_activations(im, net, layers, subWins, varargin)
+%CNN_SHAPE_GET_ACTIVATIONS Get CNN activation responses for im
 %   im:: 
 %       image matrix, #channels (size(im,3)) must be compatible with net
 %       0~255
 %   net::
 %       cnn model structure 
+%   layers:: {'relu7'}
+%       can be either a structure (.name, .sizes, .index) or string array 
 %   subWins:: [0; 1; 0; 1; 0]
 %       see get_augmentation_matrix.m for details 
-%   layers:: {'fc7'}
-%       can be either a structure (.name, .sizes, .index) or string array 
 %   `gpuMode`:: false 
 %       set to true to compute on GPU 
 
-if nargin<4 || isempty(layers), 
-    layers = {'fc7'};
+if ~exist('layers','var') || isempty(layers), 
+    layers = {'relu7'};
 end
-if nargin<3 || isempty(subWins), 
+if ~exist('subWins','var') || isempty(subWins), 
     subWins = get_augmentation_matrix('none');
 end
+
+opts.gpuMode = false;
+opts = vl_argparse(opts,varargin);
+
 nSubWins = size(subWins,2);
 if isfield(net.layers{1},'weights'), 
   nChannels = size(net.layers{1}.weights{1},3);
@@ -27,8 +31,8 @@
 
 if iscell(im), 
     imCell = im;
-    im = zeros(net.normalization.imageSize(1), ...
-        net.normalization.imageSize(2), ...
+    im = zeros(net.meta.normalization.imageSize(1), ...
+        net.meta.normalization.imageSize(2), ...
         nChannels, ...
         numel(imCell));
     for i=1:numel(imCell), 
@@ -36,7 +40,7 @@
             error('image (%d channels) is not compatible with net (%d channels)', ...
                 size(imCell{i},3), nChannels);
         end
-        im(:,:,:,i) = imresize(imCell{i}, net.normalization.imageSize(1:2));
+        im(:,:,:,i) = imresize(imCell{i}, net.meta.normalization.imageSize(1:2));
     end
 elseif size(im,3) ~= nChannels, 
     error('image (%d channels) is not compatible with net (%d channels)', ...
@@ -49,14 +53,11 @@
     if numel(viewpoolIdx)>1, 
         error('More than one viewpool layers found!'); 
     end
-    nViews = net.layers{viewpoolIdx}.stride;
+    nViews = net.layers{viewpoolIdx}.vstride;
 else
     nViews = 1;
 end
 
-opts.gpuMode = false;
-opts = vl_argparse(opts,varargin);
-
 if iscell(layers),
     layersName = layers;
     layers = struct;
@@ -67,8 +68,8 @@
     [~,layers.index] = ismember(layers.name,allLayersName);
     layers.index = layers.index + 1;
     % sizes
-    im0 = zeros(net.normalization.imageSize(1), ...
-        net.normalization.imageSize(2), nChannels, nViews, 'single') * 255;
+    im0 = zeros(net.meta.normalization.imageSize(1), ...
+        net.meta.normalization.imageSize(2), nChannels, nViews, 'single') * 255;
     if opts.gpuMode, im0 = gpuArray(im0); end
     res = vl_simplenn(net,im0);
     layers.sizes = zeros(3,numel(layers.name));
@@ -84,6 +85,10 @@
 end
 
 im = single(im);
+averageImage = net.meta.normalization.averageImage;
+if numel(averageImage)==nChannels, 
+    averageImage = reshape(averageImage, [1 1 nChannels]); 
+end
 
 for ri = 1:nSubWins,
     r = subWins(1:4,ri).*[size(im,2);size(im,2);size(im,1);size(im,1)];
@@ -91,8 +96,8 @@
     im_ = im(max(1,r(3)):min(size(im,1),r(3)+r(4)),...
         max(1,r(1)):min(size(im,2),r(1)+r(2)),:,:);
     if subWins(5,ri), im_ = flipdim(im_,2); end
-    im_ = bsxfun(@minus, imresize(im_, net.normalization.imageSize(1:2)), ...
-        net.normalization.averageImage);
+    im_ = bsxfun(@minus, imresize(im_, net.meta.normalization.imageSize(1:2)), ...
+        averageImage);
     if opts.gpuMode,
         im_ = gpuArray(im_);
     end
diff --git a/evaluate_classification.m b/evaluate_classification.m
deleted file mode 100644
index f6ab01a..0000000
--- a/evaluate_classification.m
+++ /dev/null
@@ -1,224 +0,0 @@
-function evaluate_classification(feat, varargin)
-%EVALUATE_CLASSIFICATION Evaluate CNN features for classification 
-%
-%   feat::
-%       a structure containing cnn feature
-%   `imdb`:: []
-%       optional, usually a field in feat
-%   `cv`:: 5
-%       #folds in cross validation (-v)
-%   `log2c`:: [-4:2:4]
-%       tunable liblinear svm parameter (-c) 
-%   `quiet`:: true
-%       liblinear parameter (-q)
-%   `multiview`:: true
-%       set to false to evaluate on every single views of same instances 
-%   `method`:: 'avgdesc'
-%       used only if multiview is true; other choices: 'maxdesc','avgsvmscore'
-%   `logPath`:: 'log/eval.txt'
-%       place to save log information
-%   `predPath`:: 'data/pred.mat'
-%       place to save prediction results
-%   `confusionPath`:: 'data/confusion.pdf' 
-%       place to save confusion matrix plot 
-%   
-% NOTE: assume all classes in imdb appear in training set 
-
-% default options 
-opts.imdb = [];
-opts.cv = 5;
-opts.log2c = [-4:2:4];
-opts.quiet = true;
-opts.multiview = true;
-opts.method = 'avgsdesc';
-opts.logPath = fullfile('log','eval.txt');
-opts.predPath = fullfile('data','pred.mat');
-[opts, varargin] = vl_argparse(opts, varargin) ;
-
-opts.confusionPath = fullfile(fileparts(opts.predPath), 'confusion.pdf');
-[opts, varargin] = vl_argparse(opts, varargin) ;
-
-if ~exist(fileparts(opts.logPath),'dir'), 
-    vl_xmkdir(fileparts(opts.logPath)); 
-end
-if ~exist(fileparts(opts.predPath),'dir'), 
-    vl_xmkdir(fileparts(opts.predPath)); 
-end
-if ischar(feat), 
-    feat = load(feat);
-end
-
-% -------------------------------------------------------------------------
-%                       sort imdb.images & feat.x w.r.t (sid,view) or (id)
-% -------------------------------------------------------------------------
-if isfield(feat,'imdb'), 
-    imdb = feat.imdb; 
-else
-    imdb = opts.imdb;
-end
-if ~isfield(imdb.images,'sid'), 
-    nInstances = numel(imdb.images.name); 
-else
-    nInstances = length(unique(imdb.images.sid));
-end
-
-% sort imdb.images wrt id
-[imdb.images.id,I] = sort(imdb.images.id);
-imdb.images.name = imdb.images.name(I);
-imdb.images.class = imdb.images.class(I);
-imdb.images.set = imdb.images.set(I);
-if isfield(imdb.images,'sid'), imdb.images.sid = imdb.images.sid(I); end
-
-% sort feat.x wrt id/sid
-if isfield(feat, 'sid'), 
-    pooledFeat = true;
-    opts.multiview = false;
-    [feat.sid,I] = sort(feat.sid);
-    feat.x = feat.x(I,:);
-else
-    pooledFeat = false;
-    [feat.id,I] = sort(feat.id);
-    feat.x = feat.x(I,:);
-end
-
-% sort imdb.images wrt sid
-if isfield(imdb.images,'sid'),
-    [imdb.images.sid, I] = sort(imdb.images.sid);
-    imdb.images.name = imdb.images.name(I);
-    imdb.images.class = imdb.images.class(I);
-    imdb.images.set = imdb.images.set(I);
-    imdb.images.id = imdb.images.id(I);
-    if ~pooledFeat, feat.x = feat.x(I,:); end
-end
-
-% -------------------------------------------------------------------------
-%                                                      feature descriptors
-% -------------------------------------------------------------------------
-nViews = length(imdb.images.name)/nInstances;
-nDescPerShape = size(feat.x,1)/nInstances;
-shapeGtClasses = imdb.images.class(1:nViews:end);
-shapeSets = imdb.images.set(1:nViews:end);
-nDims = size(feat.x,2);
-
-% train & val
-trainSets = {'train','val'};
-testSets = {'test'};
-[~,I] = ismember(trainSets,imdb.meta.sets);
-trainSids = find(ismember(shapeSets, I));
-tmp = zeros(nDescPerShape, nInstances);
-tmp(:,trainSids) = 1;
-trainFeat = feat.x(find(tmp)',:);
-trainLabel = shapeGtClasses(trainSids)';
-nTrainShapes = length(trainLabel); 
-
-% test 
-[~,I] = ismember(testSets,imdb.meta.sets);
-testSids = find(ismember(shapeSets, I));
-tmp = zeros(nDescPerShape, nInstances);
-tmp(:,testSids) = 1;
-testFeat = feat.x(find(tmp)',:);
-testLabel = shapeGtClasses(testSids)';
-nTestShapes = length(testLabel);
-
-if ~pooledFeat, 
-    if opts.multiview && strcmp(opts.method,'avgdesc'), 
-        % average descriptor across views 
-        trainFeat = reshape(mean(reshape(trainFeat, ...
-            [nDescPerShape nTrainShapes*nDims]),1),[nTrainShapes nDims]);
-        testFeat = reshape(mean(reshape(testFeat, ...
-            [nDescPerShape nTestShapes*nDims]),1),[nTestShapes nDims]);
-    elseif opts.multiview && strcmp(opts.method,'maxdesc'), 
-        % max descriptor across views 
-        trainFeat = reshape(max(reshape(trainFeat, ...
-            [nDescPerShape nTrainShapes*nDims]),[],1),[nTrainShapes nDims]);
-        testFeat = reshape(max(reshape(testFeat, ...
-            [nDescPerShape nTestShapes*nDims]),[],1),[nTestShapes nDims]);
-    else
-        % expand labels across views 
-        trainLabel = reshape(repmat(trainLabel',[nDescPerShape 1]),...
-            [nDescPerShape*nTrainShapes 1]);
-        testLabel = reshape(repmat(testLabel',[nDescPerShape 1]),...
-            [nDescPerShape*nTestShapes 1]);
-    end
-end   
-trainFeat = sparse(double(trainFeat));
-testFeat = sparse(double(testFeat)); 
-
-
-% -------------------------------------------------------------------------
-%                                                   train and evaluate SVM
-% -------------------------------------------------------------------------
-if exist(opts.predPath, 'file'), 
-    load(opts.predPath); 
-    fprintf('SVM model and predictions loaded from %s. \n', opts.predPath);
-else
-    fprintf('Evaluating ... \n');
-    
-    bestcv = 0;
-    for log2c = opts.log2c,
-        cmd = ['-v ', num2str(opts.cv) ,' -c ', num2str(2^log2c)];
-        if opts.quiet, cmd = [cmd ' -q']; end;
-        cv = liblinear_train(trainLabel,trainFeat,cmd);
-        if (cv >= bestcv),
-            bestcv = cv; bestc = 2^log2c;
-        end
-        fprintf('%g %g (best c=%g, rate=%g)\n', log2c, cv, bestc, bestcv);
-    end
-    
-    cmd = ['-c ', num2str(bestc)];
-    if opts.quiet, cmd = [cmd ' -q']; end;
-    model = liblinear_train(trainLabel,trainFeat,cmd);
-    
-    cmd = [''];
-    if opts.quiet, cmd = [cmd ' -q']; end;
-    [~,accuTrain,~] = liblinear_predict(trainLabel,trainFeat,model,cmd);
-    [predTest,accuTest,decTest] = liblinear_predict(testLabel,testFeat,...
-        model,cmd);
-    [~,I] = sort(model.Label);
-    decTest = decTest(:,I);
-    accuTest = accuTest(1)/100;
-    accuTrain = accuTrain(1)/100;
-    
-    if opts.multiview && strcmp(opts.method,'avgsvmscore'),
-        testLabel = testLabel(1:nDescPerShape:end);
-        nClasses = length(model.Label);
-        decTest = reshape(mean(reshape(decTest, ...
-            [nDescPerShape nTestShapes*nClasses]),1),[nTestShapes nClasses]);
-        [~,predTest] = max(decTest,[],2);
-        accuTest = sum(predTest==testLabel)/length(predTest);
-    end
-
-    save(opts.predPath,'bestc','accuTrain','predTest','accuTest','decTest',...
-        'model','opts');
-end
-
-% -------------------------------------------------------------------------
-%                                                   write & output results 
-% -------------------------------------------------------------------------
-
-% confusion matrix
-plot_confusionmat(testLabel, predTest, imdb.meta.classes, ...
-    opts.confusionPath, [imdb.imageDir ' : ' feat.modelName]);
-
-STR_01 = {'false', 'true'};
-fprintf('Evaluation finished! \n');
-fprintf('\tc: %g (cv=%d)\n', bestc, opts.cv);
-fprintf('\tdataset: %s\n', imdb.imageDir);
-fprintf('\tmodel: %s\n',feat.modelName);
-fprintf('\tlayer: %s\n',feat.layerName);
-fprintf('\tmultiview: %s', STR_01{(opts.multiview~=0)+1});
-if opts.multiview, fprintf(' (%s)', opts.method); end; fprintf('\n');
-fprintf('\taccuracy (train): %g%%\n',accuTrain*100);
-fprintf('\taccuracy (test): %g%%\n',accuTest*100);
-
-fid = fopen(opts.logPath,'a+');
-fprintf(fid, '(%s) -- Classification\n', datestr(now));
-fprintf(fid, '\tc: %g (cv=%d)\n', bestc, opts.cv);
-fprintf(fid, '\tdataset: %s\n', imdb.imageDir);
-fprintf(fid, '\tmodel: %s\n',feat.modelName);
-fprintf(fid, '\tlayer: %s\n',feat.layerName);
-fprintf(fid, '\tmultiview: %s', STR_01{(opts.multiview~=0)+1});
-if opts.multiview, fprintf(fid,' (%s)', opts.method); end; fprintf(fid,'\n');
-fprintf(fid, '\taccuracy (train): %g%%\n',accuTrain*100);
-fprintf(fid, '\taccuracy (test): %g%%\n',accuTest*100);
-fclose(fid);
diff --git a/imdb_compute_cnn_features.m b/imdb_compute_cnn_features.m
index 1c1c09a..9b8c560 100644
--- a/imdb_compute_cnn_features.m
+++ b/imdb_compute_cnn_features.m
@@ -222,7 +222,7 @@
         end
     end
     
-    feat = get_cnn_activations( im, net, subWins, layers, ...
+    feat = cnn_shape_get_activations( im, net, layers, subWins, ...
         'gpuMode', opts.gpuMode);
     parsave(fullfile(cacheDir, [num2str(i) '.mat']),feat);
     
diff --git a/imdb_render_views.m b/imdb_render_views.m
deleted file mode 100644
index 79652d1..0000000
--- a/imdb_render_views.m
+++ /dev/null
@@ -1,58 +0,0 @@
-function imdb_render_views( imdb, saveDir, varargin )
-%IMDB_RENDER_VIEWS render 3d shapes in database
-%   imdb::
-%       structure containing info about all 3d shape objects
-%   saveDir::
-%       place to save rendered images 
-%   `az`:: [0:30:330]
-%       horizontal viewing angles
-%   `el`:: 30
-%       vertical elevation
-%   `colorMode`:: 'gray'
-%       color mode of output images (only 'gray' is supported now)
-%   `outputSize`:: 224
-%       output image size (both dimensions)
-%   `minMargin`:: 0.1
-%       minimun margin ratio in output images
-%   `maxArea`:: 0.3
-%       maximun area ratio in output images 
-%   `figureStartIdx`:: floor(rand()*1e8)
-%       used to avoid conflict
-
-opts.az = [0:30:330];
-opts.el = 30;
-opts.imageExt = '.jpg'; 
-opts.colorMode = 'gray';
-opts.outputSize = 224;
-opts.minMargin = 0.1;
-opts.maxArea = 0.3;
-opts.figureStartIdx = floor(rand()*1e8);
-opts = vl_argparse(opts,varargin);
-renderOpts = rmfield(opts,'figureStartIdx');
-
-%{
-poolObj = gcp('nocreate');
-if isempty(poolObj), 
-    poolSize = 1;
-else
-    poolSize = poolObj.NumWorkers;
-end
-parfor (i=1:numel(imdb.images.name),poolSize), 
-%}
-for i=1:numel(imdb.images.name), 
-    shapePath = fullfile(imdb.imageDir,imdb.images.name{i});
-    fprintf('%d %s\n',i,shapePath);
-    fh = figure(i+opts.figureStartIdx-1);
-    ims = render_views(shapePath,'figHandle',fh,renderOpts);
-    [pathstr,namestr,extstr] = fileparts(imdb.images.name{i});
-    savePathPrefix = fullfile(saveDir,pathstr);
-    vl_xmkdir(savePathPrefix);
-    for j = 1:numel(ims), 
-        savePath = fullfile(savePathPrefix,sprintf('%s_%03d%s',namestr,j,opts.imageExt));
-        imwrite(ims{j},savePath);
-    end
-    close(fh);
-end
-
-end
-
diff --git a/retrieve_shapes_cnn.m b/retrieve_shapes_cnn.m
deleted file mode 100644
index 4d77616..0000000
--- a/retrieve_shapes_cnn.m
+++ /dev/null
@@ -1,419 +0,0 @@
-function [ results,info ] = retrieve_shapes_cnn( shape, feat, varargin )
-%RETRIEVE_SHAPES_CNN Retrieve 3d shapes using CNN features
-%
-%   shape::
-%       Can be either
-%         cell array containing projected images or paths to them
-%       OR
-%         [] (evaluation within dataset)
-%   feat::
-%       structure containing features from reference images and meta data
-%       .x
-%       .id
-%       .imdb
-%       .modelName
-%       .layerName
-%       .pcaMean
-%       .pcaCoeff
-%       .powerTrans
-%   `method`:: 'avgdesc'
-%       other choices: 'maxdesc','avgdist','mindist','avgmindist','minavgdist'
-%   `net`::[]
-%       preloaded cnn model (required when feat.modelName is not available)
-%   `gpuMode`:: false
-%       set to true to compute on GPU
-%   `numWorkers`:: 1
-%       number of workers used to compute pairwise 
-%   `nTop`:: Inf
-%       number of images in results
-%   `querySets`:: {'test'}
-%       set of query shapes (used only when shape==[])
-%   `refSets`:: {'test'}
-%       set of reference images
-%   `metric`:: 'L2'
-%       other choices: 'LINF', 'L1', 'L0', 'CHI2', 'HELL'
-%   `multiview`:: true
-%       set to false to evaluate on single views of each instance 
-%   `logPath:: []
-%       place to save log information
-%
-% Hang Su
-
-% default options
-opts.method = 'avgdesc';
-opts.net = [];
-opts.gpuMode = false;
-opts.numWorkers = 1;
-opts.nTop = Inf;
-opts.querySets = {'test'};
-opts.refSets = {'test'};
-opts.metric = 'L2';
-opts.multiview = true;
-opts.logPath = [];
-[opts, varargin] = vl_argparse(opts,varargin);
-
-if opts.numWorkers>1, 
-    pool = gcp('nocreate');
-    if isempty(pool) || pool.NumWorkers<opts.numWorkers, 
-        if ~isempty(pool), delete(pool); end
-        pool = parpool(opts.numWorkers);
-    end
-end
-
-if isequal(opts.querySets,opts.refSets), 
-    isSelfRef = true;
-else
-    isSelfRef = false;
-end
-
-% -------------------------------------------------------------------------
-%                       sort imdb.images & feat.x w.r.t (sid,view) or (id)
-% -------------------------------------------------------------------------
-imdb = feat.imdb;
-
-% sort imdb.images wrt id
-[imdb.images.id, I] = sort(imdb.images.id);
-imdb.images.name = imdb.images.name(I);
-imdb.images.class = imdb.images.class(I);
-imdb.images.set = imdb.images.set(I);
-if isfield(imdb.images,'sid'), imdb.images.sid = imdb.images.sid(I); end
-
-% sort feat.x wrt id/sid
-if isfield(feat, 'sid'), 
-    pooledFeat = true; 
-    [feat.sid,I] = sort(feat.sid);
-    feat.x = feat.x(I,:);
-else
-    pooledFeat = false; 
-    [feat.id,I] = sort(feat.id);
-    feat.x = feat.x(I,:);
-end
-
-% sort imdb.images wrt sid
-if isfield(imdb.images,'sid'),
-    [imdb.images.sid, I] = sort(imdb.images.sid);
-    imdb.images.name = imdb.images.name(I);
-    imdb.images.class = imdb.images.class(I);
-    imdb.images.set = imdb.images.set(I);
-    imdb.images.id = imdb.images.id(I);
-    if ~pooledFeat, feat.x = feat.x(I,:); end
-end
-
-% -------------------------------------------------------------------------
-%                                                      feature descriptors
-% -------------------------------------------------------------------------
-if isfield(imdb.images,'sid'), 
-    shapeStartIdx = find(imdb.images.sid-circshift(imdb.images.sid,[0 1]));
-    shapeIds = imdb.images.sid(shapeStartIdx);
-else
-    shapeStartIdx = 1:length(imdb.images.id);
-    shapeIds = imdb.images.id;
-end
-shapeGtClasses = imdb.images.class(shapeStartIdx);
-shapeSets = imdb.images.set(shapeStartIdx);
-
-% refX
-[~,I] = ismember(opts.refSets,imdb.meta.sets);
-shapeRefInds = ismember(shapeSets,I);
-refShapeIds = shapeIds(shapeRefInds);
-nRefShapes = numel(refShapeIds);
-if ~pooledFeat && isfield(imdb.images,'sid'), 
-    nRefDescPerShape = sum(imdb.images.sid==refShapeIds(1));
-else
-    nRefDescPerShape = 1;
-end
-if isfield(imdb.images,'sid'), 
-    if pooledFeat, 
-        descRefInds = shapeRefInds;
-    else
-        descRefInds = ismember(imdb.images.sid,refShapeIds); 
-    end
-else
-    descRefInds = ismember(imdb.images.id,refShapeIds); 
-end
-refX = feat.x(descRefInds, :);
-nDims = size(refX,2);
-
-% queryX
-if ~isempty(shape), % retrieval given a query shape
-    % load model
-    net = opts.net;
-    if isempty(net),
-        netFilePath = fullfile('data','models', [feat.modelName '.mat']);
-        % download model if not found
-        if ~exist(netFilePath,'file'),
-            fprintf('Downloading model (%s) ...', feat.modelName) ;
-            vl_xmkdir(fullfile('data','models')) ;
-            urlwrite(fullfile('http://maxwell.cs.umass.edu/deep-shape-data/models', ...
-                [feat.modelName '.mat']), netFilePath) ;
-            fprintf(' done!\n');
-        end
-        net = load(netFilePath);
-	net = convert_net_format(net,'old');
-    end  
-    % get raw responses
-    nViews = numel(shape);
-    for i=1:numel(shape), 
-        if ischar(shape{i}), shape{i} = imread(strrep(shape{i},'\',filesep)); end
-    end
-    if pooledFeat, 
-        queryX = get_cnn_activations(shape, net, [], {feat.layerName}, ...
-            'gpuMode', opts.gpuMode);
-        queryX = squeeze(queryX.(feat.layerName))';
-        nDescPerShape = size(queryX,1);
-    else
-        nDescPerShape = nViews;
-        queryX = zeros(nDescPerShape, nDims);
-        for i=1:nDescPerShape,
-            desc0 = get_cnn_activations(shape{i}, net, [], {feat.layerName}, ...
-                'gpuMode', opts.gpuMode);
-            desc0 = squeeze(desc0.(feat.layerName))';
-        end
-        queryX(i,:) = desc0;
-    end    
-    % normalization & projection
-    if isfield(feat,'pcaMean'),
-        queryX = bsxfun(@rdivide,queryX,sqrt(sum(queryX.^2,2)));
-        queryX = bsxfun(@minus,queryX,feat.pcaMean) * feat.pcaCoeff;
-        queryX = bsxfun(@rdivide,queryX,sqrt(sum(queryX.^2,2)));
-        if feat.powerTrans~=1,
-            queryX = (abs(queryX).^feat.powerTrans).*sign(queryX);
-        end
-    end
-else 
-    [~,I] = ismember(opts.querySets,imdb.meta.sets);
-    shapeQueryInds = ismember(shapeSets,I);
-    queryShapeIds = shapeIds(shapeQueryInds);
-    nQueryShapes = numel(queryShapeIds);
-    if ~pooledFeat && isfield(imdb.images,'sid'), 
-        nDescPerShape = sum(imdb.images.sid==queryShapeIds(1));
-    else
-        nDescPerShape = 1;
-    end
-    if isfield(imdb.images,'sid'), 
-        if pooledFeat, 
-            descQueryInds = shapeQueryInds;
-        else
-            descQueryInds = ismember(imdb.images.sid,queryShapeIds); 
-        end
-    else
-        descQueryInds = ismember(imdb.images.id,queryShapeIds); 
-    end
-    queryX = feat.x(descQueryInds, :);
-    
-    nDescPerShapeOri = nDescPerShape;
-    nQueryShapesOri = nQueryShapes;
-    if ~opts.multiview,
-        nDescPerShape = 1;
-        nQueryShapes = nQueryShapes * nDescPerShapeOri;
-    end
-    
-end
-
-% -------------------------------------------------------------------------
-%                                            retrieve by sorting distances
-% -------------------------------------------------------------------------
-switch opts.method,
-    case 'mindist',
-        dists0 = par_alldist(queryX',refX',...
-            'measure',opts.metric,'numWorkers',opts.numWorkers);
-        dists = -1*vl_nnpool(-1*single(dists0),[nDescPerShape nRefDescPerShape], ...
-            'stride', [nDescPerShape nRefDescPerShape], ...
-            'method', 'max');
-    case 'avgdist',
-        dists0 = par_alldist(queryX',refX',...
-            'measure',opts.metric,'numWorkers',opts.numWorkers);
-        dists = vl_nnpool(single(dists0),[nDescPerShape nRefDescPerShape], ...
-            'stride', [nDescPerShape nRefDescPerShape], ...
-            'method', 'avg');
-    case 'avgmindist',
-        dists0 = par_alldist(queryX',refX',...
-            'measure',opts.metric,'numWorkers',opts.numWorkers);
-        dists_1 = -1*vl_nnpool(-1*single(dists0),[nDescPerShape 1], ...
-            'stride',[nDescPerShape 1], 'method','max');
-        dists_1 = vl_nnpool(dists_1,[1 nRefDescPerShape], ...
-            'stride',[1 nRefDescPerShape], 'method','avg');
-        dists_2 = -1*vl_nnpool(-1*single(dists0),[1 nRefDescPerShape], ...
-            'stride',[1 nRefDescPerShape], 'method','max');
-        dists_2 = vl_nnpool(dists_2,[nDescPerShape 1], ...
-            'stride',[nDescPerShape 1], 'method','avg');
-        dists = 0.5*(dists_1+dists_2);
-    case 'minavgdist',
-        dists0 = par_alldist(queryX',refX',...
-            'measure',opts.metric,'numWorkers',opts.numWorkers);
-        dists_1 = vl_nnpool(single(dists0),[nDescPerShape 1], ...
-            'stride',[nDescPerShape 1], 'method','avg');
-        dists_1 = -1*vl_nnpool(-1*dists_1,[1 nRefDescPerShape], ...
-            'stride',[1 nRefDescPerShape], 'method','max');
-        dists_2 = vl_nnpool(single(dists0),[1 nRefDescPerShape], ...
-            'stride',[1 nRefDescPerShape], 'method','avg');
-        dists_2 = -1*vl_nnpool(-1*dists_2,[nDescPerShape 1], ...
-            'stride',[nDescPerShape 1], 'method','max');
-        dists = 0.5*(dists_1+dists_2);
-    case 'avgdesc',
-        desc = reshape(mean(reshape(queryX, ...
-            [nDescPerShape nQueryShapes*nDims]),1),[nQueryShapes nDims]);
-        descs = reshape(mean(reshape(refX, ...
-            [nRefDescPerShape nRefShapes*nDims]),1),[nRefShapes nDims]);
-        dists0 = par_alldist(desc',descs',...
-            'measure',opts.metric,'numWorkers',opts.numWorkers);
-        dists = dists0;
-    case 'maxdesc',
-        desc = reshape(max(reshape(queryX, ...
-            [nDescPerShape nQueryShapes*nDims]),[],1),[nQueryShapes nDims]);
-        descs = reshape(max(reshape(refX, ...
-            [nRefDescPerShape nRefShapes*nDims]),[],1),[nRefShapes nDims]);
-        dists0 = par_alldist(desc',descs',...
-            'measure',opts.metric,'numWorkers',opts.numWorkers);
-        dists = dists0;
-    otherwise,
-        error('Unknown retrieval method: %s', opts.method);
-end
-
-% -------------------------------------------------------------------------
-%                                               prepare and return results
-% -------------------------------------------------------------------------
-if ~isempty(shape), % retrieval given a query shape
-    [~,I] = sort(dists,'ascend');
-    results = refShapeIds(I(1:min(opts.nTop,nRefShapes)));
-    info = [];
-else                % no query given, evaluation within dataset
-    % pr curves
-    recall = zeros(nQueryShapes, nRefShapes+1);
-    precision = zeros(nQueryShapes, nRefShapes+1);
-    ap = zeros(nQueryShapes, 1);
-    auc = zeros(nQueryShapes, 1);
-    % interpolated pr curves
-    recall_i = zeros(nQueryShapes, nRefShapes+1);
-    precision_i = zeros(nQueryShapes, nRefShapes+1);
-    ap_i = zeros(nQueryShapes, 1);
-    auc_i = zeros(nQueryShapes, 1);
-    
-    refClass = shapeGtClasses(refShapeIds);
-    refClass = repmat(refClass, [nQueryShapes, 1]);
-    queryClass = shapeGtClasses(queryShapeIds);
-    if ~opts.multiview, 
-        queryClass = reshape(repmat(queryClass,[nDescPerShapeOri 1]), ...
-            [1 nQueryShapes]);
-    end
-    qrDists = dists;
-    
-    if isSelfRef, 
-        assert(nRefShapes==nQueryShapesOri);
-        recall(:,end) = [];
-        precision(:,end) = [];
-        recall_i(:,end) = [];
-        precision_i(:,end) = [];
-        refClass = rmDiagOnRow(refClass);
-        qrDists = rmDiagOnRow(qrDists);
-    end
-    if opts.numWorkers>1,
-        parfor q = 1:nQueryShapes,
-            [r, p, info] = vl_pr(...
-                (refClass(q,:)==queryClass(q))-0.5, ... % LABELS
-                -1*qrDists(q,:), ... % SCORES
-                'interpolate', false);
-            recall(q,:) = r;
-            precision(q,:) = p;
-            ap(q) = info.ap;
-            auc(q) = info.auc;
-            % interpolated
-            [r, p, info] = vl_pr(...
-                (refClass(q,:)==queryClass(q))-0.5, ... % LABELS
-                -1*qrDists(q,:), ... % SCORES
-                'interpolate', true);
-            recall_i(q,:) = r;
-            precision_i(q,:) = p;
-            ap_i(q) = info.ap;
-            auc_i(q) = info.auc;
-        end
-    else
-        for q = 1:nQueryShapes,
-            [r, p, info] = vl_pr(...
-                (refClass(q,:)==queryClass(q))-0.5, ... % LABELS
-                -1*qrDists(q,:), ... % SCORES
-                'interpolate', false);
-            recall(q,:) = r;
-            precision(q,:) = p;
-            ap(q) = info.ap;
-            auc(q) = info.auc;
-            % interpolated
-            [r, p, info] = vl_pr(...
-                (refClass(q,:)==queryClass(q))-0.5, ... % LABELS
-                -1*qrDists(q,:), ... % SCORES
-                'interpolate', true);
-            recall_i(q,:) = r;
-            precision_i(q,:) = p;
-            ap_i(q) = info.ap;
-            auc_i(q) = info.auc;
-        end
-    end
-    info = [];
-    info.recall = recall;
-    info.precision = precision;
-    info.ap = ap;
-    info.auc = auc;
-    info.recall_i = recall_i;
-    info.precision_i = precision_i;
-    info.ap_i = ap_i;
-    info.auc_i = auc_i;
-    clear results;
-    results.dists = dists; 
-    results.dists0 = dists0;
-    [~,I] = sort(dists,2,'ascend');
-    results.rankings = refShapeIds(I(:,1:min(opts.nTop,nRefShapes)));
-    
-    % output to log
-    if ~isempty(opts.logPath), 
-        fprintf('Evaluation finished! \n');
-        fprintf('\tdataset: %s\n', imdb.imageDir);
-        fprintf('\tmodel: %s\n',feat.modelName);
-        fprintf('\tlayer: %s\n',feat.layerName);
-        fprintf('\tmethod: %s\n', opts.method);
-        fprintf('\tmAP: %g%%\n',mean(info.ap)*100);
-        fprintf('\tAUC: %g%%\n',mean(info.auc)*100);
-        fprintf('\tmAP (interpolated): %g%%\n',mean(info.ap_i)*100);
-        fprintf('\tAUC (interpolated): %g%%\n',mean(info.auc_i)*100);
-        
-        fid = fopen(opts.logPath,'a+');
-        fprintf(fid, '(%s) -- Retrieval\n', datestr(now));
-        fprintf(fid, '\tdataset: %s\n', imdb.imageDir);
-        fprintf(fid, '\tmodel: %s\n',feat.modelName);
-        fprintf(fid, '\tlayer: %s\n',feat.layerName);
-        fprintf(fid, '\tmethod: %s\n', opts.method);
-        fprintf(fid, '\tmAP: %g%%\n',mean(info.ap)*100);
-        fprintf(fid, '\tAUC: %g%%\n',mean(info.auc)*100);
-        fprintf(fid, '\tmAP (interpolated): %g%%\n',mean(info.ap_i)*100);
-        fprintf(fid, '\tAUC (interpolated): %g%%\n',mean(info.auc_i)*100);
-        fclose(fid);
-    end
-    
-end
-
-end
-
-function M = rmDiagOnRow(M)
-    assert(ismatrix(M));
-    [sz1, sz2] = size(M);
-    if mod(sz1,sz2)==0, 
-        sy = sz1/sz2;
-        sx = 1;
-    elseif mod(sz2,sz1)==0, 
-        sy = 1;
-        sx = sz2/sz1;
-    elseif sz1==sz2,
-        u = triu(M,1);
-        l = tril(M,-1);
-        M = u(:,2:end) + l(:,1:end-1);
-        return; 
-    else
-        error('Wrong matrix size: [%d %d]',sz1, sz2);
-    end
-    ix = 1;
-    M0 = M;
-    M = zeros(sz1, sz2-sx);
-    for iy=1:sy:sz1, 
-        M(iy:iy+sy-1,:) = M0(iy:iy+sy-1,[1:ix-1 ix+sx:sz2]);
-        ix = ix + sx;
-    end
-end
diff --git a/shape_compute_descriptor.m b/shape_compute_descriptor.m
index 7c36f96..e6b84da 100644
--- a/shape_compute_descriptor.m
+++ b/shape_compute_descriptor.m
@@ -27,15 +27,11 @@
 %       'metric-relu7-v2.mat'
 %   `gpuMode`:: (default) false
 %       set to true to compute on GPU
-%   `numWorkers`:: (default) 12
-%       number of CPU workers, only in use when gpuMode is false
 
-addpath(genpath('utils'));
-run dependencies/vlfeat/toolbox/vl_setup.m
-run dependencies/matconvnet/matlab/vl_setupnn.m
+setup;
 
 if nargin<1 || isempty(path_to_shape),
-    imdbName = 'data/';
+    path_to_shape = 'data/';
 end
 
 % default options
@@ -43,7 +39,6 @@
 opts.post_process_desriptor_metric = true;
 opts.metric_model = 'metric-relu7-v1.mat';
 opts.gpuMode = false;
-opts.numWorkers = 12;
 opts = vl_argparse(opts,varargin);
 
 if exist(opts.cnn_model, 'file')
@@ -88,7 +83,11 @@
     if numel(viewpoolIdx)>1,
         error('More than one viewpool layers found!');
     end
-    num_views = cnn.layers{viewpoolIdx}.stride;
+    if ~isfield(cnn.layers{viewpoolIdx},'vstride'), 
+        num_views = cnn.layers{viewpoolIdx}.stride; % old format
+    else
+        num_views = cnn.layers{viewpoolIdx}.vstride;
+    end
     fprintf('CNN model is based on %d views. Will process %d views per mesh.\n', num_views, num_views);
 else
     error('Computing a descriptor per shape requires a multi-view CNN.');
@@ -122,7 +121,7 @@
     else
         ims = render_views(mesh, 'use_dodecahedron_views', true, 'figHandle', fig);
     end
-    outs = get_cnn_activations(ims, cnn, [], {'relu7','prob'});
+    outs = cnn_shape_get_activations(ims, cnn, {'relu7','prob'}, [], 'gpuMode', opts.gpuMode);
     out = outs.relu7(:);
     if opts.post_process_desriptor_metric
         out = double(modelDimRedFV.W*out);
diff --git a/utils/render_views_of_all_meshes_in_a_folder.m b/utils/render_views_of_all_meshes_in_a_folder.m
index 7860ea3..5abc926 100644
--- a/utils/render_views_of_all_meshes_in_a_folder.m
+++ b/utils/render_views_of_all_meshes_in_a_folder.m
@@ -1,7 +1,7 @@
 function render_views_of_all_meshes_in_a_folder(folder, varargin)
 % calls 'render_views' for every shape found in the given folder
 
-opts.ext = '.png';          % extension of target files
+opts.ext = '.jpg';          % extension of target files
 opts.range = [];            % if empty, all found shapes will be rendered, while a range [X:Y] will render shapes in the given range
 opts.useUprightAssumption = true; % if true, 12 views will be used to render meshes, otherwise 80 views based on a dodecahedron
 opts = vl_argparse(opts, varargin);
@@ -30,7 +30,7 @@ function render_views_of_all_meshes_in_a_folder(folder, varargin)
     end
     
     for ij=1:length(ims)
-        imwrite( ims{ij}, sprintf('%s_%03d.png', mesh_filenames(fi).name(1:end-4), ij) );
+        imwrite( ims{ij}, sprintf('%s_%03d%s', mesh_filenames(fi).name(1:end-4), ij, opts.ext) );
     end
 end
 close(fig);