Add SpatialDropout, BatchNormalization, input centering and scaling

Inputs are now normalized to have std deviation 1 and mean 0.

BatchIterator.lua

@@ -143,6 +143,21 @@ function BatchIterator:processImage(img, rois)
     end
   end
 
+  if cfg.normalization.centering then
+    for i = 1,3 do
+      img[i] = img[i]:add(-img[i]:mean())
+    end
+  end
+
+  if cfg.normalization.scaling then
+  for i = 1,3 do
+      local s = img[i]:std()
+      if s > 1e-8 then
+        img[i] = img[i]:div(s)
+      end
+    end
+  end
+
   img[1] = self.normalization:forward(img[{{1}}])   -- normalize luminance channel img
 
   return img, rois

config/duplo.lua

@@ -3,7 +3,7 @@
   target_smaller_side = 450,
   scales = { 32, 64, 128, 256 },
   max_pixel_size = 1000,
-  normalization = { method = 'contrastive', width = 7 },
+  normalization = { method = 'contrastive', width = 7, centering = true, scaling = true },
   augmentation = { vflip = 0.5, hflip = 0.5, random_scaling = 0.0, aspect_jitter = 0.0 },
   color_space = 'yuv',
   roi_pooling = { kw = 6, kh = 6 },

config/imagenet.lua

@@ -3,7 +3,7 @@
   target_smaller_side = 480,
   scales = { 48, 96, 192, 384 },
   max_pixel_size = 1000,
-  normalization = { method = 'contrastive', width = 7 },
+  normalization = { method = 'contrastive', width = 7, centering = true, scaling = true },
   augmentation = { vflip = 0, hflip = 0.25, random_scaling = 0, aspect_jitter = 0 },
   color_space = 'yuv',
   roi_pooling = { kw = 6, kh = 6 },

main.lua

@@ -215,6 +215,6 @@ function evaluation_demo(cfg, model_path, training_data_filename, network_filena
 
 end
 
---graph_training(cfg, opt.model, opt.name, opt.train, opt.restore)
-evaluation_demo(cfg, opt.model, opt.train, opt.restore)
+graph_training(cfg, opt.model, opt.name, opt.train, opt.restore)
+--evaluation_demo(cfg, opt.model, opt.train, opt.restore)
 

models/model_utilities.lua

@@ -8,7 +8,7 @@ function create_proposal_net(layers, anchor_nets)
     container:add(nn.SpatialConvolution(nInputPlane, nOutputPlane, kW,kH, 1,1, padW,padH))
     container:add(nn.PReLU())
     if dropout and dropout > 0 then
-      container:add(nn.Dropout(dropout))
+      container:add(nn.SpatialDropout(dropout))
     end
     return container
   end
@@ -80,6 +80,9 @@ function create_classification_net(inputs, class_count, class_layers)
   local prev_input_count = inputs
   for i,l in ipairs(class_layers) do
     net:add(nn.Linear(prev_input_count, l.n))
+    if l.batch_norm then
+      net:add(nn.BatchNormalization(l.n))
+    end
     net:add(nn.PReLU())
     if l.dropout and l.dropout > 0 then
       net:add(nn.Dropout(l.dropout))

models/vgg_large.lua

@@ -17,8 +17,8 @@ function vgg_large(cfg)
   }
 
   local class_layers =  {
-    { n=1024, dropout=0.5 },
-    { n=1024, dropout=0.5 },
+    { n=1024, dropout=0.5, batch_norm=true },
+    { n=512, dropout=0.5 },
   }
 
   return create_model(cfg, layers, anchor_nets, class_layers)

models/vgg_small.lua

@@ -17,8 +17,8 @@ function vgg_small(cfg)
   }
 
   local class_layers =  {
-    { n=1024, dropout=0.5 },
-    { n=1024, dropout=0.5 },
+    { n=1024, dropout=0.5, batch_norm=true },
+    { n=512, dropout=0.5 },
   }
 
   return create_model(cfg, layers, anchor_nets, class_layers)