最近在學習基於torch的文字識別。用的是VGG模型。

程式碼資源：https://github.com/bgshih/crnn

但是改模型是在GPU上訓練的，我沒有GPU所以要改到CPU上執行。這無疑是一個難題。在網上看到了一個改變的例子，借鑑過來參考。

原始碼連結：https://github.com/vic-w/torch-practice/tree/master/mnist

require 'torch'
require 'nn'
require 'optim'
require 'cunn'
require 'cutorch'
mnist = require 'mnist'

print('Read data set')
fullset = mnist.traindataset()
testset = mnist.testdataset()

print('Split validation set')
trainset = {
    size = 50000,
    data = fullset.data[{{1,50000}}]:double(),
    label = fullset.label[{{1,50000}}]
}

validationset = {
    size = 10000,
    data = fullset.data[{{50001,60000}}]:double(),
    label = fullset.label[{{50001,60000}}]
}

print('Normalize')
trainset.data = trainset.data - trainset.data:mean()
validationset.data = validationset.data - validationset.data:mean()


model = nn.Sequential()
model:add(nn.Reshape(1, 28, 28))
model:add(nn.MulConstant(1/256.0*3.2))
model:add(nn.SpatialConvolutionMM(1, 20, 5, 5, 1, 1, 0, 0))
model:add(nn.SpatialMaxPooling(2, 2 , 2, 2, 0, 0))
model:add(nn.SpatialConvolutionMM(20, 50, 5, 5, 1, 1, 0, 0))
model:add(nn.SpatialMaxPooling(2, 2 , 2, 2, 0, 0))
model:add(nn.Reshape(4*4*50))
model:add(nn.Linear(4*4*50, 500))
model:add(nn.ReLU())
model:add(nn.Linear(500, 10))
model:add(nn.LogSoftMax())

model = require('weight-init')(model, 'xavier')

criterion = nn.ClassNLLCriterion()

model = model:cuda()
criterion = criterion:cuda()
trainset.data = trainset.data:cuda()
trainset.label = trainset.label:cuda()
validationset.data = validationset.data:cuda()
validationset.label = validationset.label:cuda()

sgd_params = {
   learningRate = 1e-2,
   learningRateDecay = 1e-4,
   weightDecay = 1e-3,
   momentum = 1e-4
}

x, dl_dx = model:getParameters()

step = function(batch_size)
    local current_loss = 0
    local count = 0
    local shuffle = torch.randperm(trainset.size)
    batch_size = batch_size or 200
    for t = 1,trainset.size,batch_size do
        -- setup inputs and targets for this mini-batch
        local size = math.min(t + batch_size - 1, trainset.size) - t
        local inputs = torch.Tensor(size, 28, 28):cuda()
        local targets = torch.Tensor(size):cuda()
        for i = 1,size do
            local input = trainset.data[shuffle[i+t]]
            local target = trainset.label[shuffle[i+t]]
            -- if target == 0 then target = 10 end
            inputs[i] = input
            targets[i] = target
        end
        targets:add(1)
        local feval = function(x_new)
            -- reset data
            if x ~= x_new then x:copy(x_new) end
            dl_dx:zero()

            -- perform mini-batch gradient descent
            local loss = criterion:forward(model:forward(inputs), targets)
            model:backward(inputs, criterion:backward(model.output, targets))
            return loss, dl_dx
        end
        _, fs = optim.sgd(feval, x, sgd_params)
        -- fs is a table containing value of the loss function
        -- (just 1 value for the SGD optimization)
        count = count + 1
        current_loss = current_loss + fs[1]
    end
    -- normalize loss
    return current_loss / count
end
eval = function(dataset, batch_size)
    local count = 0
    batch_size = batch_size or 200
    for i = 1,dataset.size,batch_size do
        local size = math.min(i + batch_size - 1, dataset.size) - i
        local inputs = dataset.data[{{i,i+size-1}}]:cuda()
        local targets = dataset.label[{{i,i+size-1}}]:cuda()
        local outputs = model:forward(inputs)
        local _, indices = torch.max(outputs, 2)
        indices:add(-1)
        indices = indices:cuda()
        local guessed_right = indices:eq(targets):sum()
        count = count + guessed_right
    end
    return count / dataset.size
end
max_iters = 30
print('Start training')
do
    local last_accuracy = 0
    local decreasing = 0
    local threshold = 1 -- how many deacreasing epochs we allow
    for i = 1,max_iters do
        local loss = step()
        print(string.format('Epoch: %d Current loss: %4f', i, loss))
        local accuracy = eval(validationset)
        print(string.format('Accuracy on the validation set: %4f', accuracy))
        if accuracy < last_accuracy then
            if decreasing > threshold then break end
            decreasing = decreasing + 1
        else
            decreasing = 0
        end
        last_accuracy = accuracy
    end
end
testset.data = testset.data:double()
eval(testset)
改過之後

require 'torch'  
require 'nn'  
require 'optim'  
--require 'cunn'  
--require 'cutorch'  
mnist = require 'mnist'  
  --刪了一行print('Read data set')
fullset = mnist.traindataset()  
testset = mnist.testdataset()  
  --刪了一行print

('.......set')
trainset = {  
    size = 50000,  
    data = fullset.data[{{1,50000}}]:double(),  
    label = fullset.label[{{1,50000}}]  
}  
 
validationset = {  
    size = 10000,  
    data = fullset.data[{{50001,60000}}]:double(),  
    label = fullset.label[{{50001,60000}}]  
}  
  --刪除print

('Normalize')
trainset.data = trainset.data - trainset.data:mean()  
validationset.data = validationset.data - validationset.data:mean()  
 
 
model = nn.Sequential()  
model:add(nn.Reshape(1, 28, 28))  
model:add(nn.MulConstant(1/256.0*3.2))  
model:add(nn.SpatialConvolutionMM(1, 20, 5, 5, 1, 1, 0, 0))  
model:add(nn.SpatialMaxPooling(2, 2 , 2, 2, 0, 0))  
model:add(nn.SpatialConvolutionMM(20, 50, 5, 5, 1, 1, 0, 0))  
model:add(nn.SpatialMaxPooling(2, 2 , 2, 2, 0, 0))  
model:add(nn.Reshape(4*4*50))  
model:add(nn.Linear(4*4*50, 500))  
model:add(nn.ReLU())  
model:add(nn.Linear(500, 10))  
model:add(nn.LogSoftMax())  
 
model = require('weight-init')(model, 'xavier')  
 
criterion = nn.ClassNLLCriterion()  
 
--model = model:cuda()  
--criterion = criterion:cuda()  
--trainset.data = trainset.data:cuda()  
--trainset.label = trainset.label:cuda()  
--validationset.data = validationset.data:cuda()  
--validationset.label = validationset.label:cuda()  
 
sgd_params = {  
   learningRate = 1e-2,  
   learningRateDecay = 1e-4,  
   weightDecay = 1e-3,  
   momentum = 1e-4  
}  
 
x, dl_dx = model:getParameters()  
 
step = function(batch_size)  
    local current_loss = 0  
    local count = 0  
    local shuffle = torch.randperm(trainset.size)  
    batch_size = batch_size or 200  
    for t = 1,trainset.size,batch_size do  
        -- setup inputs and targets for this mini-batch  
        local size = math.min(t + batch_size - 1, trainset.size) - t  
        local inputs = torch.Tensor(size, 28, 28) --:cuda()  
        local targets = torch.Tensor(size) --:cuda()  
        for i = 1,size do  
            local input = trainset.data[shuffle[i+t]]  
            local target = trainset.label[shuffle[i+t]]  
            -- if target == 0 then target = 10 end  
            inputs[i] = input  
            targets[i] = target  
        end  
        targets:add(1)  
        local feval = function(x_new)  
            -- reset data  
            if x ~= x_new then x:copy(x_new) end  
            dl_dx:zero()  
 
            -- perform mini-batch gradient descent  
            local loss = criterion:forward(model:forward(inputs), targets)  
            model:backward(inputs, criterion:backward(model.output, targets))  
 
            return loss, dl_dx  
        end  
 
        _, fs = optim.sgd(feval, x, sgd_params)  
 
        -- fs is a table containing value of the loss function  
        -- (just 1 value for the SGD optimization)  
        count = count + 1  
        current_loss = current_loss + fs[1]  
    end  
 
    -- normalize loss  
    return current_loss / count  
end  
 
eval = function(dataset, batch_size)  
    local count = 0  
    batch_size = batch_size or 200  
      
    for i = 1,dataset.size,batch_size do  
        local size = math.min(i + batch_size - 1, dataset.size) - i  
        local inputs = dataset.data[{{i,i+size-1}}] --:cuda()  
        local targets = dataset.label[{{i,i+size-1}}]:long() --:cuda()  
        local outputs = model:forward(inputs)  
        local _, indices = torch.max(outputs, 2)  
        indices:add(-1)  
        local guessed_right = indices:eq(targets):sum()  
        count = count + guessed_right  
    end  
 
    return count / dataset.size  
end  
 
max_iters = 30  
 
do  
    local last_accuracy = 0  
    local decreasing = 0  
    local threshold = 1 -- how many deacreasing epochs we allow  
    for i = 1,max_iters do  
        local loss = step()  
        print(string.format('Epoch: %d Current loss: %4f', i, loss))  
        local accuracy = eval(validationset)  
        print(string.format('Accuracy on the validation set: %4f', accuracy))  
        if accuracy < last_accuracy then  
            if decreasing > threshold then break end  
            decreasing = decreasing + 1  
        else  
            decreasing = 0  
        end  
        last_accuracy = accuracy  
    end  
end  
 
testset.data = testset.data:double()  
eval(testset)