-1

I'm learning LSTM networks and decided to try synthetic test. I want LSTM network fed by some points (x,y) to distinguish between three basic functions:

  • line: y = k*x + b
  • parabola: y = k*x^2 + b
  • sqrt: y = k*sqrt(x) + b

I'm using lua + torch.

Dataset is totally virtual - it is created on-the-fly at the 'dataset' object. When training cycle asks for another minibatch of samples, function mt.__index returns sample, created dynamically. It randomly selects on of the three described functions and picks some random points for them.

Idea is that LSTM network would learn some features to recognize what kind of a function do last points belong to.

Full yet simple source script included:

require "torch"
require "nn"
require "rnn"

-- hyper-parameters 
batchSize = 8
rho = 5 -- sequence length
hiddenSize = 100
outputSize = 3
lr = 0.001

-- Initialize synthetic dataset
-- dataset[index] returns table of the form: {inputs, targets}
-- where inputs is a set of points (x,y) of a randomly selected function: line, parabola, sqrt
-- and targets is a set of corresponding class of a function (1=line, 2=parabola, 3=sqrt)
local dataset = {}
dataset.size = function (self)
  return 1000
end
local mt = {}
mt.__index = function (self, i)
  local class = math.random(3)

  local t = torch.Tensor(3):zero()
  t[class] = 1
  local targets = {}
  for i = 1,batchSize do table.insert(targets, class) end

  local inputs = {}
  local k = math.random()
  local b = math.random()*5

  -- Line
  if class == 1 then
    for i = 1,batchSize do
      local x = math.random()*10 + 5
      local y = k*x + b
      input = torch.Tensor(2)
      input[1] = x
      input[2] = y
      table.insert(inputs, input)
    end

  -- Parabola
  elseif class == 2 then
    for i = 1,batchSize do
      local x = math.random()*10 + 5
      local y = k*x*x + b
      input = torch.Tensor(2)
      input[1] = x
      input[2] = y
      table.insert(inputs, input)
    end

  -- Sqrt
  else
    for i = 1,batchSize do
      local x = math.random()*5 + 5
      local y = k*math.sqrt(x) + b
      input = torch.Tensor(2)
      input[1] = x
      input[2] = y
      table.insert(inputs, input)
    end
  end

  return { inputs, targets }
end -- dataset.__index meta function
setmetatable(dataset, mt)

-- Initialize random number generator
math.randomseed( os.time() )

-- build simple recurrent neural network
local model = nn.Sequencer(
  nn.Sequential()
    :add( nn.LSTM(2, hiddenSize, rho) )
    :add( nn.Linear(hiddenSize, outputSize) )
    :add( nn.LogSoftMax() )
)

print(model)

-- build criterion
local criterion = nn.SequencerCriterion( nn.ClassNLLCriterion() )

-- training
model:training()

local epoch = 1
while true do

  print ("Epoch "..tostring(epoch).." started")

  for iteration = 1, dataset:size() do
    -- 1. Load minibatch of samples
    local sample = dataset[iteration] -- pick random sample (dataset always returns random set)
    local inputs = sample[1]
    local targets = sample[2]

    -- 2. Perform forward run and calculate error
    local outputs = model:forward(inputs)
    local err = criterion:forward(outputs, targets)

    print(string.format("Epoch %d Iteration %d Error = %f", epoch, iteration, err))

    -- 3. Backward sequence through model(i.e. backprop through time)
    local gradOutputs = criterion:backward(outputs, targets)
    -- Sequencer handles the backwardThroughTime internally
    model:backward(inputs, gradOutputs)
    model:updateParameters(lr)
    model:zeroGradParameters()     

  end -- for dataset

  epoch = epoch + 1
end -- while epoch

The problem is: network does not converge. Could you share any ideas what I'm doing wrong?

4

2 回答 2

2

This approach is completely wrong. Learning LSTM in this way will not learn it what you want because of many reasons. I will state two of them :

  1. Let's assume that you draw your x uniformly from (-1, 1). Then functions |x| and 0.5x + 0.5 will give you exactly the same distribiutions of y. This shows you that the method which you use is not the best for function recognition.

  2. What is crucial in LSTM it is it's memory which allow you to store informations between inputs. It is exactly oposite from drawing the sequence of points independently (what you do in your script). Every memory correlation which would be learnt in your method could be only spurious.

于 2016-04-26T22:58:28.937 回答
0

I decided to post my own answer since I solved the problem and received good results.

First about applicability of LSTM to this kind of task. As stated, LSTM is good to deal with time series. You may also think of line, parabola and sqrt as a kind of a time function. So LSTM is totally applicable here. Say you're receiving experimental results, one vector at a moment, and you want to find out what kind of function could describe your series?

One may argue that in the code above we always get feed NN with a fixed number of points (i.e. batch_size). So why use LSTM? Maybe try to use instead some Linear or Convolution Network?

Well, don't forget - this is a synthetic test. In a real life application you may feed NN with some significant amount of data points and expect it to recognize the form of function.

For instance in the code below we train NN with 8 points at once (batch_size), but when we test NN we use only 4 points (test_size).

And we get pretty good results: after about 1000 iterations NN gives ~99% of correct answers.

But one-layer NN is not a magician. It can't learn any features if we change the form of functions on each iterations. I.e. in the original code k and b are changed at every request to dataset. What we should do is to generate them at startup and do not change.

So the working code below:

require "torch"
require "nn"
require "rnn"

-- Initialize random number generator
math.randomseed( os.time() )

-- hyper-parameters 
batch_size = 8
test_size = 4
rho = 5 -- sequence length
hidden_size = 100
output_size = 3
learning_rate = 0.001

-- Initialize synthetic dataset
-- dataset[index] returns table of the form: {inputs, targets}
-- where inputs is a set of points (x,y) of a randomly selected function: line, parabola, sqrt
-- and targets is a set of corresponding class of a function (1=line, 2=parabola, 3=sqrt)
local dataset = {}
dataset.k = math.random()
dataset.b = math.random()*5
dataset.size = function (self)
  return 1000
end
local mt = {}
mt.__index = function (self, i)
  local class = math.random(3)

  local t = torch.Tensor(3):zero()
  t[class] = 1
  local targets = {}
  for i = 1,batch_size do table.insert(targets, class) end

  local inputs = {}
  local k = self.k
  local b = self.b

  -- Line
  if class == 1 then
    for i = 1,batch_size do
      local x = math.random()*10 + 5
      local y = k*x + b
      input = torch.Tensor(2)
      input[1] = x
      input[2] = y
      table.insert(inputs, input)
    end

  -- Parabola
  elseif class == 2 then
    for i = 1,batch_size do
      local x = math.random()*10 + 5
      local y = k*x*x + b
      input = torch.Tensor(2)
      input[1] = x
      input[2] = y
      table.insert(inputs, input)
    end

  -- Sqrt
  else
    for i = 1,batch_size do
      local x = math.random()*5 + 5
      local y = k*math.sqrt(x) + b
      input = torch.Tensor(2)
      input[1] = x
      input[2] = y
      table.insert(inputs, input)
    end
  end

  return { inputs, targets }
end -- dataset.__index meta function
setmetatable(dataset, mt)


-- build simple recurrent neural network
local model = nn.Sequencer(
  nn.Sequential()
    :add( nn.LSTM(2, hidden_size, rho) )
    :add( nn.Linear(hidden_size, output_size) )
    :add( nn.LogSoftMax() )
)

print(model)

-- build criterion
local criterion = nn.SequencerCriterion( nn.ClassNLLCriterion() )


local epoch = 1
local err = 0
local pos = 0
local N = math.floor( dataset:size() * 0.1 )

while true do

  print ("Epoch "..tostring(epoch).." started")

  -- training
  model:training()
  for iteration = 1, dataset:size() do
    -- 1. Load minibatch of samples
    local sample = dataset[iteration] -- pick random sample (dataset always returns random set)
    local inputs = sample[1]
    local targets = sample[2]

    -- 2. Perform forward run and calculate error
    local outputs = model:forward(inputs)
    local _err = criterion:forward(outputs, targets)

    print(string.format("Epoch %d (pos=%f) Iteration %d Error = %f", epoch, pos, iteration, _err))

    -- 3. Backward sequence through model(i.e. backprop through time)
    local gradOutputs = criterion:backward(outputs, targets)
    -- Sequencer handles the backwardThroughTime internally
    model:backward(inputs, gradOutputs)
    model:updateParameters(learning_rate)
    model:zeroGradParameters()     

  end -- for training

  -- Testing
  model:evaluate()
  err = 0
  pos = 0
  for iteration = 1, N do
    -- 1. Load minibatch of samples
    local sample = dataset[ math.random(dataset:size()) ]
    local inputs = sample[1]
    local targets = sample[2]
    -- Drop last points to reduce to test_size
    for i = #inputs, test_size, -1 do
      inputs[i] = nil
      targets[i] = nil
    end

    -- 2. Perform forward run and calculate error
    local outputs = model:forward(inputs)
    err = err + criterion:forward(outputs, targets)

    local p = 0
    for i = 1, #outputs do
      local _, oi = torch.max(outputs[i], 1)
      if oi[1] == targets[i] then p = p + 1 end
    end
    pos = pos + p/#outputs

  end -- for testing
  err = err / N
  pos = pos / N
  print(string.format("Epoch %d testing results: pos=%f err=%f", epoch, pos, err))

  if (pos > 0.95) then break end

  epoch = epoch + 1
end -- while epoch
于 2016-04-27T21:24:22.710 回答