Merge pull request #15 from szagoruyko/master

imagenet classification demo

Author: soumith

7_imagenet_classification/classify.lua

@@ -0,0 +1,86 @@
+-- Imagenet classification with Torch7 demo
+require 'loadcaffe'
+require 'image'
+
+-- Helper functions
+
+-- Loads the mapping from net outputs to human readable labels
+function load_synset()
+  local file = io.open 'synset_words.txt'
+  local list = {}
+  while true do
+    local line = file:read()
+    if not line then break end
+    table.insert(list, string.sub(line,11))
+  end
+  return list
+end
+
+
+-- Converts an image from RGB to BGR format and subtracts mean
+function preprocess(im, img_mean)
+  -- rescale the image
+  local im3 = image.scale(im,227,227,'bilinear')*255
+  -- RGB2BGR
+  local im4 = im3:clone()
+  im4[{1,{},{}}] = im3[{3,{},{}}]
+  im4[{3,{},{}}] = im3[{1,{},{}}]
+
+  -- subtract imagenet mean
+  return im4 - image.scale(img_mean, 227, 227, 'bilinear')
+end
+
+
+
+-- Setting up networks and downloading stuff if needed
+proto_name = 'deploy.prototxt'
+model_name = 'bvlc_reference_caffenet.caffemodel'
+img_mean_name = 'ilsvrc_2012_mean.t7'
+image_name = 'Goldfish3.jpg'
+
+prototxt_url = 'https://raw.githubusercontent.com/BVLC/caffe/master/models/bvlc_reference_caffenet/'..proto_name
+model_url = 'http://dl.caffe.berkeleyvision.org/'..model_name
+img_mean_url = 'https://www.dropbox.com/s/p33rheie3xjx6eu/'..img_mean_name
+image_url = 'http://upload.wikimedia.org/wikipedia/commons/e/e9/Goldfish3.jpg'
+
+if not paths.filep(proto_name) then os.execute('wget '..prototxt_url) end
+if not paths.filep(model_name) then os.execute('wget '..model_url)    end
+if not paths.filep(img_mean_name) then os.execute('wget '..img_mean_url) end
+if not paths.filep(image_name) then os.execute('wget '..image_url)   end
+
+
+
+print '==> Loading network'
+-- we'll use the fastest CUDA ConvNet implementation available, cuda-convnet2
+-- this loads the network in Caffe format and returns in Torch format, ready to use!
+net = loadcaffe.load(proto_name, model_name, 'ccn2')
+
+-- as we want to classify, let's disable dropouts by enabling evaluation mode
+net:evaluate()
+
+print '==> Loading synsets'
+synset_words = load_synset()
+
+print '==> Loading image and imagenet mean'
+im = image.load(image_name)
+img_mean = torch.load(img_mean_name).img_mean:transpose(3,1)
+
+print '==> Preprocessing'
+-- Have to resize and convert from RGB to BGR and subtract mean
+I = preprocess(im, img_mean)
+
+-- cuda-convnet2 implementation support only batched routines, so
+-- we have to allocate memory for 32 inputs and then put crops to 10 of them.
+-- let's however use just one image for simplicity.
+-- note that for other networks that use cunn ore cudnn that might not be needed
+batch = torch.CudaTensor(32,3,227,227)
+batch[1]:copy(I)
+
+print '==> Propagating through the network'
+net:forward(batch)
+
+-- for the outputs of SoftMax layer sort them in decreasing order
+_,classes = net:get(25).output[{1,{}}]:float():sort(true)
+for i=1,5 do
+  print('predicted class '..tostring(i)..': ', synset_words[classes[i]])
+end