Created by Hang Zhang

HzProc is a torch data augmentation toolbox for deep learning. The name of "HzProc" means high speed image processing, where "Hz" (hertz) is the unit of frequency and "Proc" is abbreviation of processing. HZ is also the initial of the author. It contains the functions within the following subcategories:

Quick Start Example

• Online
  • Crop and Flip
  • Affine Transformation
  • Get Affine Matrix
• Offline
  • Remap
  • Get Lookup Table
  • Combine Transforms

Deep learning is known as one of the most popular methods in Computer Vision area today. It has achieved superior result in many competitions such as ImageNet and Coco. Deep Neural Network optimizes millions of the parameters during the training. However, its application on real life problem is typically limited by the dataset size. A sophisticated data augmentation strategy can be the key to improve the results and avoid overfitting. But bad implementation can easily become the computational bottleneck.

HzProc addresses this issue, which transfers different data augmentation strategies into two simple sub-problems: affine transformation and lookup table remapping. Therefore, we can generate affine transformation matrix or lookup table offine (every epoch), and augment the image data online (taking constant time). HzProc is an open source toolbox, which has a highly optimized CUDA backend.

This package relies on torch7 and cutorch. Please note the package also relies on a NVIDIA GPU compitable with CUDA 6.5 or higher version.

git clone https://github.com/zhanghang1989/hzproc
cd hzproc
luarocks make hzproc-scm-1.rockspec

Demo The demo script relies on qtlua and image package to load and display the images. The test script is also a good usage example. The documentations can be found in the link.

qlua test/demo.lua

Fast Examples

• online example:

function RandomCrop(size, pad)
  return function(input)
		-- avoid unnecessary opterations
    local w, h = input:size(3), input:size(2)
    if w == size and h == size then
      return input
    end
    local x1, y1 = torch.random(0, w + 2*pad - size), 
									torch.random(0, h + 2*pad - size)
    local out = hzproc.Crop.Pad(input, x1, y1, size, size, pad)
    assert(out:size(2) == size and out:size(3) == size, 'wrong crop size')
    return out
  end
end

• offline example:

function Resize(inw, inh, ow, oh)
	-- generating the lookup map only once during init
	local map = hzproc.Table.Resize(inw, inh, ow, oh)
  return function(input)
		-- avoid unnecessary opterations
	  local w, h = input:size(3), input:size(2)
    if (w == ow and h == oh) then
      return input
    end	
		-- mapping
		return hzproc.Remap.Bilinear(input, map)
  end
end