kkanshul / hide-and-seek Goto Github PK

Hi, I'm a novice on Torch7. I'm sorry for my confusion about how to get the prediction of an image. I wrote the following code:

-- ....

local modelCache   = paths.concat('./cache', 'model_googlenet_has_32.t7')
local model = torch.load(modelCache) 
model:evaluate()
model:cuda()
-- print(model)

for line in io.lines(paths.concat(DATA_ROOT, '../images_1-25000.txt')) do
  line = split(line, " ")
  local img = testHook({loadSize}, paths.concat(DATA_ROOT, line[1]))

  if img:dim() == 3 then
    img = img:view(1, img:size(1), img:size(2), img:size(3))
  end

  local predictions = model:forward(img:cuda())
  local pred_idx, pred_sorted = predictions:sort(2, true)
  print(line[1])
  print(pred_sorted[1][1]-1)
end

Where 'testHook' is extracted from https://github.com/soumith/imagenet-multiGPU.torch. But I got hardly 0 top-1 accuracy on 'ILSVRC2012_img_val'. What's wrong with my code?

Looking forward to your reply :)

Hi, Krishna

Thank you for sharing the source cod for the awesome paper!
I was reading your paper to embed HaS into my pytorch model.
You mentioned the hidden pixel values are supposed to set to the mean of all the pixel values in the entire training. Does that mean I just need to change 0 to the mean of all the pixels at line 30 in hide_patch.py? (

For us to freely use the code.

What's the advantage of HaS over the random cropping that is commonly used for data augmentation?