allow labelmap switching

benchmarks/cnn_imagenet_se_mcn.m

@@ -8,6 +8,7 @@
   opts.gpus = 3 ;
   opts.continue = 1 ;
   opts.batchSize = 256 ;
+  opts.labelFormat = 'v1' ;
   opts.model = 'SE-ResNet-50-mcn' ;
   opts.modelDir = fullfile(vl_rootnn, 'data/models-import') ;
   opts.dataDir = fullfile(vl_rootnn, 'data/datasets/ILSVRC2012') ;
@@ -38,7 +39,11 @@
   end
 
   % remap labels to match the order used in training
-  imdb = updateLabelMap(imdb, opts) ;
+  switch opts.labelFormat
+    case 'v1', imdb = updateLabelMap(imdb, opts) ;
+    case 'v2' % do nothing
+    otherwise, error('unrecognised label format %s', opts.labelFormat) ;
+  end
 
 % -------------------------------------------------------------------------
 %                                                             Prepare model
@@ -71,7 +76,12 @@
   end
 
   dag = dagnn.DagNN.loadobj(load(modelPath)) ;
-  dag.addLayer('softmax', dagnn.SoftMax(), dag.layers(end).outputs, 'prediction', {}) ;
+  if isa(dag.layers(end).block, 'dagnn.SoftMax')
+    dag.renameVar(dag.layers(end).outputs, 'prediction') ;
+  else % add softmax to logits if softmax isn't yet present
+    dag.addLayer('softmax', dagnn.SoftMax(), dag.layers(end).outputs, ...
+                                                   'prediction', {}) ;
+  end
   dag.addLayer('top1err', dagnn.Loss('loss', 'classerror'), ...
                {'prediction','label'}, 'top1err') ;
   dag.addLayer('top5err', dagnn.Loss('loss', 'topkerror', 'opts', {'topK',5}), ...
@@ -95,9 +105,9 @@
   end
 
 % ----------------------------------------
-function imdb = updateLabelMap(imdb, opts) 
+function imdb = updateLabelMap(imdb, opts)
 % ----------------------------------------
-  labelMap = importdata(opts.labelMap) ; 
+  labelMap = importdata(opts.labelMap) ;
   keep = imdb.images.label ~= 0 ;
   newLabels = labelMap(imdb.images.label(keep)) ;
   imdb.images.label(keep) = newLabels ;