Cifar 10 numbers reported use data augmentation about densenet HOT 3 CLOSED

@FredericGodin The paper does use standard data augmentation techniques. See Table 2, and its description underneath. "+" next to a dataset means it went through standard data augmentation, which is what I am comparing my results against.

from densenet.

Let me rephrase that. I meant that the data augmentation used is different from the one mentioned in the paper.

generator = ImageDataGenerator(rotation_range=15,
                               width_shift_range=5./32,
                               height_shift_range=5./32)

The standard data augmentation is mirroring and using a random crop I believe?
You use rotation?

From the Lua code:

function CifarDataset:preprocess()
   if self.split == 'train' then
      return t.Compose{
         t.ColorNormalize(meanstd),
         t.HorizontalFlip(0.5),
         t.RandomCrop(32, 4),
      }
   elseif self.split == 'val' then
      return t.ColorNormalize(meanstd)
   else
      error('invalid split: ' .. self.split)
   end
end

from densenet.

Oh that's what you meant. Yes, I did not use horizontal mirroring even though it was there because it was causing poor results compared to not using it.

Center cropping isn't available yet in image preprocessing so I could not implement that.

There are several other papers which use width and height shift as standard augmentation for Cifar 10 and 100, though they also suggest horizontal mirroring.

from densenet.