torch7-autograd-optim.lua

require 'nn';
require 'nngraph';
require 'optim';
autograd = require 'autograd';

-- W0 = torch.ones(3, 1)
params = {W=torch.rand(3,1)}
optimState = {learningRate=0.001}
criterion = autograd.nn.MSECriterion()
function neuralNet(params, x, y)
    local out = x * params.W 
    return criterion(out, y)
end
dneuralNet = autograd(neuralNet)

batch_size = 5
input_dim = 3
function get_sample()
    local input = torch.rand(batch_size, input_dim)
    -- local label = torch.cmul(input, input)
    -- local label = torch.sum(label, 2)
    local label = torch.sum(input, 2)
    return input, label
end

local feval = function(params) 
    -- do something here
    dparams, loss = dneuralNet(params, input, label)
    return dparams, loss
end
local optimfn, states = autograd.optim.sgd(feval, optimState, params)

steps = 10000
log_per_steps = 1000
for i = 1,steps do
    input, label = get_sample()
    local grads, loss = optimfn(input, label)
    if i % log_per_steps == 0 then
        print(i, loss)
    end
end
print(params.W)