----------------------------------------------------------------------

-- Main script for training a model for semantic segmentation

--

-- Abhishek Chaurasia, Eugenio Culurciello

-- Sangpil Kim, Adam Paszke

-- Edited by Eren Golge

----------------------------------------------------------------------

require 'pl'

require 'nn'

require 'cudnn'

require 'cunn'

local opts = require 'opts'

local DataLoader = require 'data/dataloader'

torch.setdefaulttensortype('torch.FloatTensor')

----------------------------------------------------------------------

-- Get the input arguments parsed and stored in opt

opt = opts.parse(arg)

print(opt)

--print(cutorch.getDeviceProperties(opt.devid))

cutorch.setDevice(opt.devid)

print("Folder created at " .. opt.save)

os.execute('mkdir -p ' .. opt.save)

----------------------------------------------------------------------

print '==> load modules'

local data, chunks, ft

-- data loading

local trainLoader, valLoader = DataLoader.create(opt)

opt.classes = trainLoader.classes

-- save opt to file

print 'saving opt as txt and t7'

local filename = paths.concat(opt.save,'opt.txt')

local file = io.open(filename, 'w')

for i,v in pairs(opt) do

file:write(tostring(i)..' : '..tostring(v)..'\n')

end

file:close()

torch.save(path.join(opt.save,'opt.t7'),opt)

----------------------------------------------------------------------

print '==> training!'

local epoch = 1

-- create model

t = paths.dofile("models/"..opt.model..".lua")

local train = require 'train'

local test = require 'test'

local besterr = 9999999999

while epoch < opt.maxepoch do

print("----- epoch # " .. epoch)

local trainConf, model, loss = train(trainLoader, epoch)

besterr = test(valLoader, epoch, trainConf, model, loss, besterr )

-- trainConf = nil

collectgarbage()

epoch = epoch + 1

end