example.lua

#!/usr/bin/env th
require('torch')
require('nn')
require('image')
require('paths')
torch.setdefaulttensortype('torch.FloatTensor')


local eye = 231              -- small net requires 231x231
local label_nb = 630         -- label of 'bee'
local mean = 118.380948/255  -- global mean used to train overfeat
local std = 61.896913/255    -- global std used to train overfeat
local intensity = 1          -- pixel intensity for gradient sign

local path_img = 'bee.jpg'
local path_model = 'model.net'
-- this script requires pretrained overfeat model from the repo
-- (https://github.com/jhjin/overfeat-torch)
-- it will fetch the repo and create the model if it does not exist


-- get dependent files
if not paths.filep('model.net') then
   os.execute([[
      git clone https://github.com/jhjin/overfeat-torch
      cd overfeat-torch
      echo "torch.save('model.net', net)" >> run.lua
      . install.sh && th run.lua && mv model.net .. && cd ..
   ]])
end
if not paths.filep('bee.jpg') then
   os.execute('wget https://raw.githubusercontent.com/sermanet/OverFeat/master/samples/bee.jpg')
end
if not paths.filep('overfeat_label.lua') then
   os.execute('wget https://raw.githubusercontent.com/jhjin/overfeat-torch/master/overfeat_label.lua')
end
local label = require('overfeat_label')


-- resize input/label
local img = image.scale(image.load(path_img), '^'..eye)
local tx = math.floor((img:size(3)-eye)/2) + 1
local ly = math.floor((img:size(2)-eye)/2) + 1
img = img[{{},{ly,ly+eye-1},{tx,tx+eye-1}}]
img:add(-mean):div(std)

-- get trained model (switch softmax to logsoftmax)
local model = torch.load(path_model)
model.modules[#model.modules] = nn.LogSoftMax()

-- set loss function
local loss = nn.ClassNLLCriterion()

-- generate adversarial examples
local img_adv = require('adversarial-fast')(model, loss, img:clone(), label_nb, std, intensity)

model.modules[#model.modules] = nn.SoftMax()
-- check prediction results
local pred = model:forward(img)
local val, idx = pred:max(pred:dim())
print('==> original:', label[ idx[1] ], 'confidence:', val[1])

local pred = model:forward(img_adv)
local val, idx = pred:max(pred:dim())
print('==> adversarial:', label[ idx[1] ], 'confidence:', val[1])

local img_diff = torch.add(img, -img_adv)
print('==> mean absolute diff between the original and adversarial images[min/max]:', torch.abs(img_diff):mean())

image.save('img.png', img:mul(std):add(mean):clamp(0,255))
image.save('img_adv.png', img_adv:mul(std):add(mean):clamp(0,255))

if pcall(require,'qt') then
  local img_cat = torch.cat(torch.cat(img, img_adv, 3), img_diff:mul(127), 3)
  image.display(img_cat)
end