ethanyhzhang/hzproc

torch data augmentation toolbox (supports affine transform)

HzProc

Created by Hang Zhang

HzProc is a fast data augmentation toolbox supporting affine transformation with GPU backend. (PyTorch version is also provide.) 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

Install

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.

On Linux

luarocks install https://raw.githubusercontent.com/zhanghang1989/hzproc/master/hzproc-scm-1.rockspec

On OSX

CC=clang CXX=clang++ luarocks install https://raw.githubusercontent.com/zhanghang1989/hzproc/master/hzproc-scm-1.rockspec

Quick Demo

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.

git clone https://github.com/zhanghang1989/hzproc
cd hzproc
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