Mirror of forum.nim-lang.org

10196 :: slice 3D Tensor into 2D Tensor in arraymancer?

• oyster (orginal) [2023-05-14T16:16:48+02:00] view original

  In my following code, a colorful RGB image, which is a 3D Tensor, is converted into a gray image, which should be a 2D Tensor. In short

  
  GRAY_2D = 0.299*RED_2D + 0.587*GREEN_2D + 0.114*BLUE_2D
  Run

  However the result imgGray has a rank of 3. The question must root from imgIn[_, _, 0].map(i => float64(i)*0.299) which produces a tensor of 3-rank. So what is the way to fix the problem? Thanks

  
  import sugar
  
  import ./src/arraymancer
  
  
  var imgIn  = read_image("1.png").permute(2, 1, 0)  # chw2whc
  echo "imgIn.rank = ", imgIn.rank
  echo "imgIn.shape = ", imgIn.shape
  
  let
      M = imgIn.shape[0]
      N = imgIn.shape[1]
      C = imgIn.shape[2]
  
  var imgGray = zeros[uint8]([M, N])
  echo "zero imgGray.rank = ", imgGray.rank  # 2, is expected
  echo "zero imgGray.shape = ", imgGray.shape
  
  if imgIn.rank == 3:
      imgGray = (
                          imgIn[_, _, 0].map(i => float64(i)*0.299) +
                          imgIn[_, _, 1].map(i => float64(i)*0.587) +
                          imgIn[_, _, 2].map(i => float64(i)*0.114)
                      ).map(i => uint8(i))[_, _]
  else:
      imgGray = imgIn
  
  echo "new imgGray.rank = ", imgGray.rank   # 3, is unexpected
  echo "new imgGray.shape = ", imgGray.shape
  Run

---

• doofenstein (orginal) [2023-05-14T17:06:45+02:00] view original

  You can get rid of a remaining dimension which is just 1 by reshaping it.

  
  import sugar
  
  import arraymancer
  
  
  var imgIn  = read_image("narg.png").permute(2, 1, 0)  # chw2whc
  echo "imgIn.rank = ", imgIn.rank
  echo "imgIn.shape = ", imgIn.shape
  
  let
      M = imgIn.shape[0]
      N = imgIn.shape[1]
      C = imgIn.shape[2]
  
  var imgGray = zeros[uint8]([M, N])
  echo "zero imgGray.rank = ", imgGray.rank  # 2, is expected
  echo "zero imgGray.shape = ", imgGray.shape
  
  if imgIn.rank == 3:
      imgGray = (
              imgIn[_, _, 0].map(i => float64(i)*0.299) +
              imgIn[_, _, 1].map(i => float64(i)*0.587) +
              imgIn[_, _, 2].map(i => float64(i)*0.114)
          ).map(i => uint8(i))
      
      let reshapedImage = imgGray.reshape(imgIn.shape[0..1])
      # just a sample to show that they're identical:
      for i in 0..<100:
          assert reshapedImage[i,i] == imgGray[i,i,0]
      imgGray = reshapedImage
  else:
      imgGray = imgIn
  
  echo "new imgGray.rank = ", imgGray.rank   # 3, is unexpected
  echo "new imgGray.shape = ", imgGray.shape
  Run

---