cleanup

benchmarks/run_se_benchmarks.m

@@ -9,12 +9,13 @@
 useCached = 1 ; % load results from cache if available
 
 importedModels = {
-'SE-ResNet-152-mcn', ...
-'SENet-mcn', ...
-'SE-ResNeXt-101-32x4d-mcn', ...
-'SE-ResNeXt-50-32x4d-mcn', ...
+'SE-BN-Inception-mcn', ...
 'SE-ResNet-50-mcn', ...
 'SE-ResNet-101-mcn', ...
+'SE-ResNet-152-mcn', ...
+'SE-ResNeXt-50-32x4d-mcn', ...
+'SE-ResNeXt-101-32x4d-mcn', ...
+'SENet-mcn', ...
 } ;
 
 for ii = 1:numel(importedModels)

check.m

@@ -1,11 +1,15 @@
 %net = load('data/models-import/SE-ResNeXt-50-mcn.mat') ;
 %net = load('data/models-import/SE-ResNet-50-mcn.mat') ;
-net = load('data/models-import/SE-ResNet-101-mcn.mat') ;
+%net = load('data/models-import/SE-ResNet-101-mcn.mat') ;
 %net = load('data/models-import/imagenet-resnet-50-dag.mat') ;
-%net = load('data/models-import/SE-BN-Inception-mcn.mat') ;
+net = load('data/models-import/SE-BN-Inception-mcn.mat') ;
 dag = dagnn.DagNN.loadobj(net) ;
 
-imPath = '000017.jpg' ;
+labelMapPath = fullfile(vl_rootnn, 'contrib/mcnSENets/misc/label_map.txt') ;
+labelMap = importdata(labelMapPath) ; 
+
+%imPath = '000017.jpg' ;
+imPath = 'peppers.png' ;
 im = single(imresize(imread(imPath), [224 224])) ;
 RGB = [123, 117, 104] ;
 rgb = permute(RGB, [1 3 2]) ;
@@ -14,7 +18,10 @@
 dag.eval({'data', im}) ;
 preds = dag.vars(dag.getVarIndex('prob')).value ; 
 
-[bestScore, best] = max(preds) ;
-%figure(1) ; clf ; imagesc(im) ; zs_dispFig ;
+[bestScore, best_] = max(preds) ;
+%best = best_ ;
+best = find(labelMap == best_) ; % remap label
+%best = labelMap(best_) ;
+figure(1) ; clf ; imagesc(im) ; zs_dispFig ;
 fprintf('%s (%d), score %.3f\n',...
 dag.meta.classes.description{best}, best, bestScore) ;

matlab/vl_nnreshape.m

@@ -24,23 +24,24 @@
 %   block projected onto DZDY. DZDX and DZDY have the same dimensions
 %   as X and Y respectively.
 %
-% Copyright (C) 2017 Samuel Albanie and Andrea Vedaldi.
+% Copyright (C) 2017 Samuel Albanie and Andrea Vedaldi
 % Licensed under The MIT License [see LICENSE.md for details]
 
-[~, dzdy] = vl_argparsepos(struct(), varargin) ;
+  [~, dzdy] = vl_argparsepos(struct(), varargin) ;
 
-if isnumeric(shape) % apply caffe style conventions if needed
-  shape_ = num2cell(shape) ;
-  k = find(shape == -1) ; if k, shape_{k} = [] ; end
-  k = find(shape == 0) ;
-  if k, rep = arrayfun(@(i) {size(x,i)}, k) ; shape_(k) = rep ; end
-  shape = shape_ ;
-end
+  if isnumeric(shape) % apply caffe style conventions if needed
+    shape_ = num2cell(shape) ;
+    if numel(shape_) == 2, shape_{3} = [] ; end
+    k = find(shape == -1) ; if k, shape_{k} = [] ; end
+    k = find(shape == 0) ;
+    if k, rep = arrayfun(@(i) {size(x,i)}, k) ; shape_(k) = rep ; end
+    shape = shape_ ;
+  end
 
-batchSize = size(x, 4);
+  batchSize = size(x, 4);
 
-if isempty(dzdy)
-  y = reshape(x, shape{1}, shape{2}, shape{3}, batchSize) ;
-else
-  y = reshape(dzdy{1}, size(x)) ;
-end
+  if isempty(dzdy)
+    y = reshape(x, shape{1}, shape{2}, shape{3}, batchSize) ;
+  else
+    y = reshape(dzdy{1}, size(x)) ;
+  end

misc/fix_se_imports.m

@@ -42,12 +42,7 @@ function fix_se_imports(varargin)
     % fix meta 
     fprintf('adding info to %s (%d/%d)\n', modelPath, mm, numel(modelNames)) ;
     net.meta.classes = imdb.classes ;
-    if contains(modelPath, 'Inception')
-      imSize = [299 299 3] ;
-    else
-      imSize = [224 224 3] ;
-    end
-    net.meta.normalization.imageSize = imSize ;
+    net.meta.normalization.imageSize = [224 224 3] ;
     net = dagnn.DagNN.loadobj(net) ; 
     net = net.saveobj() ; save(modelPath, '-struct', 'net') ; %#ok
   end