Torch7 教程 Supervised Learning CNN              

分类：                           2014-08-08 15:59     1426人阅读     (0)              

全部代码放在：https://github.com/guoyilin/CNN_Torch7

在搭建好Torch7之后，我们开始进行监督式Supervised Learning for CNN, Torch7提供了代码和一些说明文件：

http://code.madbits.com/wiki/doku.php?id=tutorial_supervised_1_data　和http://torch.cogbits.com/doc/tutorials_supervised/说的比较详细。

结合http://ufldl.stanford.edu/wiki/index.php/Feature_extraction_using_convolution了解ＣＮＮ的做法，最关键的是要熟悉http://ufldl.stanford.edu/wiki/index.php/Backpropagation_Algorithm　算法的主要做法。ｂｐ算法的目的是为了一次性计算所有的参数导数，该算法利用了ｃｈａｉｎ　ｒｕｌｅ进行ｅｒｒｏｒ的后向传播。这篇文章写了bp算法: http://neuralnetworksanddeeplearning.com/chap2.html,  写的比较详细。

如果背景不熟悉，可以看看Linear Classification, Neutral Network, SGD算法。

由于该教程使用了torch自己的数据格式，因此如果你要使用自己的数据，需要预先转换下。这里我训练的是图像分类，因此可以使用

https://github.com/clementfarabet/graphicsmagick 进行数据的加载。

如下是加载图像的代码：

height = 200
width = 200
--see if the file exists
function file_exists(file)
  local f = io.open(file, "rb")
  if f then f:close() end
  return f ~= nil
end

function read_file (file)
  if not file_exists(file) then return {} end
  lines = {}
  for line in io.lines(file) do
    lines[#lines + 1] = line
  end
  return lines
end

-- read all label name. hash them to id.
labels_id = {}
label_lines = read_file('labels.txt')
for i = 1, #label_lines do
  labels_id[label_lines[i]] = i
end

-- read train data. iterate train.txt

local train_lines = read_file("train.txt")
local train_features = torch.Tensor(#train_lines, 3, height, width) -- dimension: sample number, YUV, height, width
local train_labels = torch.Tensor(#train_lines) -- dimension: sample number

for i = 1, #train_lines do
  local image = gm.Image("/train_images/" .. train_lines[i])
  image:size(width, height)
  img_yuv = image:toTensor('float', 'YUV', 'DHW')
  --print(img_yuv:size())
  --print(img_yuv:size())
  train_features[i] = img_yuv
  local label_name = train_lines[i]:match("([^,]+)/([^,]+)")
  train_labels[i] = labels_id[label_name]
  --print(train_labels[i])
  if(i % 100 == 0) then
    print("train data: " .. i)
  end
end

trainData = {
  data = train_features:transpose(3,4),
  labels = train_labels,
  --size = function() return #train_lines end
  size = function() return #train_lines end
}

-- read test data. iterate test.txt
local test_lines = read_file("test.txt")

local test_features = torch.Tensor(#test_lines, 3, height, width) -- dimension: sample number, YUV, height, width
local test_labels = torch.Tensor(#test_lines) -- dimension: sample number

for i = 1, #test_lines do
  -- if image size is zero, gm.Imge may throw error, we need to dispose it later.
  local image = gm.Image("test_images/" .. test_lines[i])
  --print(test_lines[i])

  image:size(width, height)
  local img_yuv = image:toTensor('float', 'YUV', 'DHW')
  --print(img_yuv:size())
  test_features[i] = img_yuv
  local label_name = test_lines[i]:match("([^,]+)/([^,]+)")
  test_labels[i] = labels_id[label_name]
  --print(test_labels[i])
  if(i % 100 == 0) then
    print("test data: " .. i)
  end
end

testData = {
  data = test_features:transpose(3,4),
  labels = test_labels,
  --size = function() return #test_lines end
  size = function() return #test_lines end
}
trsize = #train_lines
tesize = #test_lines

height = 200width = 200--see if the file existsfunction file_exists(file)  local f = io.open(file, "rb")  if f then f:close() end  return f ~= nilendfunction read_file (file)  if not file_exists(file) then return {} end  lines = {}  for line in io.lines(file) do    lines[#lines + 1] = line  end  return linesend-- read all label name. hash them to id.labels_id = {}label_lines = read_file('labels.txt')for i = 1, #label_lines do  labels_id[label_lines[i]] = iend-- read train data. iterate train.txtlocal train_lines = read_file("train.txt")local train_features = torch.Tensor(#train_lines, 3, height, width) -- dimension: sample number, YUV, height, widthlocal train_labels = torch.Tensor(#train_lines) -- dimension: sample numberfor i = 1, #train_lines do  local image = gm.Image("/train_images/" .. train_lines[i])  image:size(width, height)  img_yuv = image:toTensor('float', 'YUV', 'DHW')  --print(img_yuv:size())  --print(img_yuv:size())  train_features[i] = img_yuv  local label_name = train_lines[i]:match("([^,]+)/([^,]+)")  train_labels[i] = labels_id[label_name]  --print(train_labels[i])  if(i % 100 == 0) then    print("train data: " .. i)  endendtrainData = {  data = train_features:transpose(3,4),  labels = train_labels,  --size = function() return #train_lines end  size = function() return #train_lines end}-- read test data. iterate test.txtlocal test_lines = read_file("test.txt")local test_features = torch.Tensor(#test_lines, 3, height, width) -- dimension: sample number, YUV, height, widthlocal test_labels = torch.Tensor(#test_lines) -- dimension: sample numberfor i = 1, #test_lines do  -- if image size is zero, gm.Imge may throw error, we need to dispose it later.  local image = gm.Image("test_images/" .. test_lines[i])  --print(test_lines[i])  image:size(width, height)  local img_yuv = image:toTensor('float', 'YUV', 'DHW')  --print(img_yuv:size())  test_features[i] = img_yuv  local label_name = test_lines[i]:match("([^,]+)/([^,]+)")  test_labels[i] = labels_id[label_name]  --print(test_labels[i])  if(i % 100 == 0) then    print("test data: " .. i)  endendtestData = {  data = test_features:transpose(3,4),  labels = test_labels,  --size = function() return #test_lines end  size = function() return #test_lines end}trsize = #train_linestesize = #test_lines

由于图像的大小从32*32变成了200*200， 因此需要修改相应的model中的每一层的大小。

假定其他层没有变化，最后一层需要修改：

-- stage 3 : standard 2-layer neural network
 model:add(nn.Reshape(nstates[2]*47*47))
 model:add(nn.Linear(nstates[2]*47*47, nstates[3]))
 model:add(nn.Tanh())
 model:add(nn.Linear(nstates[3], noutputs))