test_sharedclone.lua

require 'nn'
require 'rnn'
require 'image'
require 'optim'

require 'loader'
require 'ctc_log'
require 'utils/decoder'

local threads = require 'threads'


-- initialize
torch.setdefaulttensortype('torch.FloatTensor')
torch.manualSeed(450)

-- debug switch
DEBUG = false

-- timer initialize
base = 0
timer = torch.Timer()

function show_log(log)
	local now = timer:time().real
	local cost = now - base
	base = now
	print(string.format("[%.4f][%.4f]%s", now, cost, log))
end 

-- settings

DROPOUT_RATE = 0.4
GPU_ENABLED = false
local input_size = 32
local hidden_size = 200
clamp_size = 5

-- configuration
training_list_file = "1.txt"

-- GPU

if GPU_ENABLED then
	require 'cutorch'
	require 'cunn'
end

-- load samples

show_log("Loading samples...")

local loader = Loader()
loader:load(training_list_file)
local codec = loader:codec()

show_log(string.format("Loading finished. Got %d samples, %d classes of characters.", #loader.samples, codec.codec_size))

local class_num = codec.codec_size

-- build network

show_log("Building networks...")

local net

net = nn.Sequential()

net:add(nn.Dropout(DROPOUT_RATE))
net:add(nn.SplitTable(1))
net:add(nn.BiSequencer(nn.FastLSTM(input_size, hidden_size)))

output = nn.Sequential()
output:add(nn.Linear(hidden_size * 2, class_num + 1))
output:add(nn.SoftMax())
net:add(nn.Sequencer(output))
net:float()

-- prepare prarmeters and training method

local params, grad_params

params, grad_params = net:getParameters()

n = net:sharedClone(true, false)

p, gd = n:getParameters()

local sample = loader:pick()
local im = sample.img
local target = codec:encode(sample.gt)

outputTable = n:forward(im)
loss, grad = ctc.getCTCCostAndGrad(outputTable, target)

n:backward(im, grad)