main_snapshot_ensemble.lua

--
--  Copyright (c) 2016, Facebook, Inc.
--  All rights reserved.
--
--  This source code is licensed under the BSD-style license found in the
--  LICENSE file in the root directory of this source tree. An additional grant
--  of patent rights can be found in the PATENTS file in the same directory.
--
--  Code modified for Snapshot Ensembles by Gao Huang

require 'torch'
require 'paths'
require 'optim'
require 'nn'

local DataLoader = require 'dataloader'
local models = require 'models/init'
local Trainer = require 'train'
local EnsembleTester = require 'test_ensemble'
local opts = require 'opts'
local checkpoints = require 'checkpoints'


torch.setdefaulttensortype('torch.FloatTensor')
torch.setnumthreads(1)

local opt = opts.parse(arg)
torch.manualSeed(opt.manualSeed)
cutorch.manualSeedAll(opt.manualSeed)

-- Load previous checkpoint, if it exists
local checkpoint, optimState = checkpoints.latest(opt)

-- Create model
local model, criterion = models.setup(opt, checkpoint)

-- Data loading
local trainLoader, valLoader = DataLoader.create(opt)

-- The trainer handles the training loop and evaluation on validation set
local trainer = Trainer(model, criterion, opt, optimState)

if opt.testOnly then
   local top1Err, top5Err = trainer:test(0, valLoader)
   print(string.format(' * Results top1: %6.3f  top5: %6.3f', top1Err, top5Err))
   return
end

local startEpoch = checkpoint and checkpoint.epoch + 1 or opt.epochNumber
local bestTop1 = math.huge
local bestTop5 = math.huge
for epoch = startEpoch, opt.nEpochs do
   -- Train for a single epoch
   local trainTop1, trainTop5, trainLoss = trainer:train(epoch, trainLoader)

   -- Run model on validation set
   local testTop1, testTop5 = trainer:test(epoch, valLoader)

   local bestModel = false
   if testTop1 < bestTop1 then
      bestModel = true
      bestTop1 = testTop1
      bestTop5 = testTop5
      print(' * Best model ', testTop1, testTop5)
   end

   checkpoints.save(epoch, model, trainer.optimState, bestModel, opt)
end

-- Test Snapshot Ensembles
local model_ensemble_fw = {}
local model_ensemble_bw = {}
for cycle = 1, opt.nCycles do
   local idx = (cycle == opt.nCycles) and opt.nEpochs or cycle * torch.floor(opt.nEpochs / opt.nCycles)
   model_ensemble_fw[#model_ensemble_fw+1] = torch.load(opt.save..'/model_'..idx..'.t7')
   model_ensemble_bw[opt.nCycles-cycle+1] = model_ensemble_fw[#model_ensemble_fw] 
end
local tester_fw = EnsembleTester(model_ensemble_fw, opt)
local SingleTop1_fw, SingleTop5_fw, EnsembleTop1Evolve_fw, EnsembleTop5Evolove_fw = tester_fw:test(valLoader)
local tester_bw = EnsembleTester(model_ensemble_bw, opt)
local SingleTop1_bw, SingleTop5_bw, EnsembleTop1Evolve_bw, EnsembleTop5Evolove_bw = tester_bw:test(valLoader)