深度学习框架 Torch 7 问题笔记

  1 --    We now have 5 steps left to do in training our first torch neural network
  2 --    1. Load and normalize data
  3 --    2. Define Neural Network
  4 --    3. Define Loss function
  5 --    4. Train network on training data
  6 --    5. Test network on test data.
  7 
  8 
  9 
 10 
 11 --    1. Load and normalize data
 12 require ‘paths‘
 13 require ‘image‘;
 14 if (not paths.filep("cifar10torchsmall.zip")) then
 15     os.execute(‘wget -c https://s3.amazonaws.com/torch7/data/cifar10torchsmall.zip‘)
 16     os.execute(‘unzip cifar10torchsmall.zip‘)
 17 end
 18 trainset = torch.load(‘cifar10-train.t7‘)
 19 testset = torch.load(‘cifar10-test.t7‘)
 20 classes = {‘airplane‘, ‘automobile‘, ‘bird‘, ‘cat‘,
 21            ‘deer‘, ‘dog‘, ‘frog‘, ‘horse‘, ‘ship‘, ‘truck‘}
 22 
 23 print(trainset)
 24 print(#trainset.data)
 25 
 26 itorch.image(trainset.data[100]) -- display the 100-th image in dataset
 27 print(classes[trainset.label[100]])
 28 
 29 -- ignore setmetatable for now, it is a feature beyond the scope of this tutorial.
 30 -- It sets the index operator 
 31 setmetatable(trainset,
 32     {__index = function(t, i)
 33                     return {t.data[i], t.label[i]}
 34                 end}
 35 );
 36 trainset.data = trainset.data:double()  -- convert the data from a ByteTensor to a DoubleTensor.
 37 
 38 function trainset:size()
 39     return self.data:size(1)
 40 end
 41 
 42 print(trainset:size())
 43 print(trainset[33])
 44 itorch.image(trainset[33][11])
 45 
 46 redChannel = trainset.data[{ {}, {1}, {}, {} }] -- this pick {all images, 1st channel, all vertical pixels, all horizontal pixels}
 47 print(#redChannel)
 48 
 49 -- TODO:fill
 50 mean = {}
 51 stdv = {}
 52 for i = 1,3 do 
 53     mean[i] = trainset.data[{ {}, {i}, {}, {} }]:mean()  -- mean estimation 
 54     print(‘Channel ‘ .. i .. ‘ , Mean: ‘ .. mean[i])
 55     trainset.data[{ {}, {i}, {}, {} }]:add(-mean[i]) -- mean subtraction 
 56 
 57     stdv[i] = trainset.data[ { {}, {i}, {}, {} }]:std()  -- std estimation 
 58     print(‘Channel ‘ .. i .. ‘ , Standard Deviation: ‘ .. stdv[i])
 59     trainset.data[{ {}, {i}, {}, {} }]:div(stdv[i])  -- std scaling 
 60 end 
 61 
 62 
 63 
 64 --    2. Define Neural Network
 65 net = nn.Sequential()
 66 net:add(nn.SpatialConvolution(3, 6, 5, 5)) -- 3 input image channels, 6 output channels, 5x5 convolution kernel
 67 net:add(nn.ReLU())                       -- non-linearity 
 68 net:add(nn.SpatialMaxPooling(2,2,2,2))     -- A max-pooling operation that looks at 2x2 windows and finds the max.
 69 net:add(nn.SpatialConvolution(6, 16, 5, 5))
 70 net:add(nn.ReLU())                       -- non-linearity 
 71 net:add(nn.SpatialMaxPooling(2,2,2,2))
 72 net:add(nn.View(16*5*5))                    -- reshapes from a 3D tensor of 16x5x5 into 1D tensor of 16*5*5
 73 net:add(nn.Linear(16*5*5, 120))             -- fully connected layer (matrix multiplication between input and weights)
 74 net:add(nn.ReLU())                       -- non-linearity 
 75 net:add(nn.Linear(120, 84))
 76 net:add(nn.ReLU())                       -- non-linearity 
 77 net:add(nn.Linear(84, 10))                   -- 10 is the number of outputs of the network (in this case, 10 digits)
 78 net:add(nn.LogSoftMax())                     -- converts the output to a log-probability. Useful for classification problems
 79 
 80 
 81 -- 3. Let us difine the Loss function 
 82 criterion = nn.ClassNLLCriterion()
 83 
 84 
 85 
 86 -- 4. Train the neural network 
 87 trainer = nn.StochasticGradient(net, criterion)
 88 trainer.learningRate = 0.001
 89 trainer.maxIteration = 5 -- just do 5 epochs of training. 
 90 trainer:train(trainset)
 91 
 92 
 93 
 94 -- 5. Test the network, print accuracy
 95 print(classes[testset.label[100]])
 96 itorch.image(testset.data[100])
 97 
 98 testset.data = testset.data:double()   -- convert from Byte tensor to Double tensor
 99 for i=1,3 do -- over each image channel
100     testset.data[{ {}, {i}, {}, {}  }]:add(-mean[i]) -- mean subtraction    
101     testset.data[{ {}, {i}, {}, {}  }]:div(stdv[i]) -- std scaling
102 end
103 
104 -- for fun, print the mean and standard-deviation of example-100
105 horse = testset.data[100]
106 print(horse:mean(), horse:std())
107 
108 print(classes[testset.label[100]])
109 itorch.image(testset.data[100])
110 predicted = net:forward(testset.data[100])
111  
112 -- the output of the network is Log-Probabilities. To convert them to probabilities, you have to take e^x 
113 print(predicted:exp())
114 
115 
116 for i=1,predicted:size(1) do
117     print(classes[i], predicted[i])
118 end
119 
120 
121 -- test the accuracy 
122 correct = 0
123 for i=1,10000 do
124     local groundtruth = testset.label[i]
125     local prediction = net:forward(testset.data[i])
126     local confidences, indices = torch.sort(prediction, true)  -- true means sort in descending order
127     if groundtruth == indices[1] then
128         correct = correct + 1
129     end
130 end
131 
132 
133 print(correct, 100*correct/10000 .. ‘ % ‘)
134 
135 class_performance = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0}
136 for i=1,10000 do
137     local groundtruth = testset.label[i]
138     local prediction = net:forward(testset.data[i])
139     local confidences, indices = torch.sort(prediction, true)  -- true means sort in descending order
140     if groundtruth == indices[1] then
141         class_performance[groundtruth] = class_performance[groundtruth] + 1
142     end
143 end
144 
145 
146 for i=1,#classes do
147     print(classes[i], 100*class_performance[i]/1000 .. ‘ %‘)
148 end
149 
150 require ‘cunn‘;
151 net = net:cuda()
152 criterion = criterion:cuda()
153 trainset.data = trainset.data:cuda()
154 trainset.label = trainset.label:cuda()
155 
156 trainer = nn.StochasticGradient(net, criterion)
157 trainer.learningRate = 0.001
158 trainer.maxIteration = 5 -- just do 5 epochs of training.
159 
160 
161 trainer:train(trainset)