First steps with torch -- logistic regression

logreg_adv_index.lua

require 'torch'
require 'nn'
require 'optim'

trainset = torch.load("mnist.t7/train_32x32.t7", "ascii")
trainset.data = trainset.data:resize(60000, 32*32)
trainset.data = trainset.data:double()

test = torch.load("mnist.t7/test_32x32.t7", "ascii")
test.data = test.data:resize(10000, 32*32)
test.data = test.data:double()

function trainset:size() return trainset.data:size(1) end
function test:size() return test.data:size(1) end

sz = trainset.data:size(2)
model = nn.Sequential()
model:add(nn.Linear(sz, 10))
model:add(nn.LogSoftMax())


criterion = nn.ClassNLLCriterion()
criterion.sizeAverage = false

params, grad_params = model:getParameters()
params:zero()

config = {learningRate = 1e-5, momentum = 0.9}
btchsz = 100
epoch = 1
function train()
    epoch = epoch or 1
    print("Epoch:", epoch)

    for t = 1, trainset:size(), btchsz do
        
        local feval = function(x)

            if x ~= params then
                params:copy(x)
            end

            grad_params:zero()
            
            local inputs = trainset.data[{{t,t+btchsz-1},{}}]
            local targets = trainset.labels[{{t,t+btchsz-1}}]
            
            local output = model:forward(inputs)
            local err = criterion:forward(output, targets)

            local dfdo = criterion:backward(output, targets)
            
            model:backward(inputs, dfdo)
            grad_params:div(inputs:size(1))

            return err, grad_params
        end

        optim.sgd(feval, params, config)
    end
    tst(trainset)
    print()
    epoch = epoch + 1
end


function tst(test)
    local output = model:forward(test.data)
    local err = criterion:forward(output, test.labels)
    max, idx = output:max(2)
    local zo = torch.ne(idx:byte(), test.labels):sum()
    print("NLL per sample:", err/test:size())
    print("Wrong classifications:", zo)
end

for j = 1,10 do
    train()
end

print("Test set evaluation")
tst(test)

logreg_for_index.lua

require 'torch'
require 'nn'
require 'optim'

trainset = torch.load("mnist.t7/train_32x32.t7", "ascii")
trainset.data = trainset.data:resize(60000, 32*32)
trainset.data = trainset.data:double()

test = torch.load("mnist.t7/test_32x32.t7", "ascii")
test.data = test.data:resize(10000, 32*32)
test.data = test.data:double()

function trainset:size() return trainset.data:size(1) end
function test:size() return test.data:size(1) end

sz = trainset.data:size(2)
model = nn.Sequential()
model:add(nn.Linear(sz, 10))
model:add(nn.LogSoftMax())


criterion = nn.ClassNLLCriterion()
criterion.sizeAverage = false

params, grad_params = model:getParameters()
params:zero()

config = {learningRate = 1e-5, momentum = 0.9}
btchsz = 100
epoch = 1
function train()
    epoch = epoch or 1
    print("Epoch:", epoch)

    for t = 1, trainset:size(), btchsz do

        local inputs = {}
        local targets = {}

        for i = t, math.min(t+btchsz-1, trainset:size()) do
            local input = trainset.data[i]
            local target = trainset.labels[i]
            table.insert(inputs, input)
            table.insert(targets, target)
        end
        
        local feval = function(x)
            
            if x ~= params then
                params:copy(x)
            end

            grad_params:zero()


            local f = 0
            for i = 1,#inputs do
                local output = model:forward(inputs[i])
                local err = criterion:forward(output, targets[i])
                f = f + err

                local df_do = criterion:backward(output, targets[i])
                model:backward(inputs[i], df_do)
            end

            grad_params:div(#inputs)
            return f, grad_params
        end

        optim.sgd(feval, params, config)
    end
    tst(trainset)
    epoch = epoch + 1
    print()
end

function tst(test)
    local output = model:forward(test.data)
    local err = criterion:forward(output, test.labels)
    max, idx = output:max(2)
    local zo = torch.ne(idx:byte(), test.labels):sum()
    print("NLL per sample:", err/test:size())
    print("Wrong classifications:", zo)
end

for j = 1,10 do
    train()
end
print("Test set evaluation")
tst(test)