load_copy.lua

require 'nn'
require 'ntm'
require 'gnuplot'
local tasks = require 'tasks'

local ntm_params = {
    input_size = 9,
    output_size = 9,
    mem_locations = 128,
    mem_location_size = 20,
    hidden_state_size = 100,
    allowed_shifts = {-1,0,1}
}

local ntm = nn.NTM(ntm_params)

-- Parameters of a model trained with sequences of length 25.
-- Trained with zeros as targets for input phase :
-- input size : 9, output size : 9, n memeory slots : 128
local loaded_params = torch.load('parameters/copy/copy_force/24.01.2017_14:26:31_len=1-20_lr=0.0001/25000-0.00001.params')

-- Trained without specific output target for input phase :
-- input size : 9, output size : 9, n memeory slots : 128
-- local loaded_params = torch.load('parameters/copy/copy_no_force/24.01.2017_12:32:00_len=1-20_lr=0.0001/25000-0.00002.params')

-- Smaller model : input size : 7, output size : 7, n memeory slots : 10
-- local loaded_params = torch.load('parameters/copy/toy/copy_force=false_seed=1/29.01.2017_10:58:15_len=1-7_lr=0.0001/25000-0.00002.params')


local ntm_p, ntm_g = ntm:getParameters()
ntm_p:copy(loaded_params)


local min_seq_len = 37
local max_seq_len = 37

local crit = nn.BCECriterion()

for i = min_seq_len, max_seq_len do
    local seq_len = i
    local inputs, targets, exp_out  = tasks.generate_copy_sequence(seq_len, ntm_params.input_size, false)

    local out = torch.Tensor(targets:size())

    local err = 0
    local n_out = 0
    for j=1,inputs:size(1) do
        out[j] = ntm:forward(inputs[j])
        if exp_out[j] then
            err = err + crit:forward(out[j], targets[j])
            n_out = n_out + 1
        end
    end

    err = err / n_out

    io.write('Input : \n')
    print(inputs)


    io.write('Output : \n')
    print(out)

    local diff = (targets - out):abs() 
    io.write('Error : \n')
    print (diff)

    gnuplot.figure(1)
    gnuplot.imagesc(inputs:t(),'color')

    gnuplot.figure(2)
    gnuplot.imagesc(out:t(),'color')

    gnuplot.figure(3)
    gnuplot.imagesc(diff:t(),'color')



    io.write('Sequence length\t\tError\n')
    local str_format = '%d\t\t\t%f\n'
    io.write(str_format:format(seq_len, err))
    io.flush()
    ntm:new_sequence()
end