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
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
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。
-- 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))