Move all calculations to float16

daltonize/daltonize.py

@@ -36,7 +36,7 @@ def transform_colorspace(img, mat):
  out : array of shape (M, N, 3)
  """
  # Fast element (=pixel) wise matrix multiplication
- return np.einsum("ij, ...j", mat, img)
+ return np.einsum("ij, ...j", mat, img, dtype=np.float16, casting="same_kind")
 
 
 def simulate(rgb, color_deficit="d"):
@@ -58,17 +58,17 @@ def simulate(rgb, color_deficit="d"):
  """
  # Colorspace transformation matrices
  cb_matrices = {
- "d": np.array([[1, 0, 0], [1.10104433, 0, -0.00901975], [0, 0, 1]]),
- "p": np.array([[0, 0.90822864, 0.008192], [0, 1, 0], [0, 0, 1]]),
- "t": np.array([[1, 0, 0], [0, 1, 0], [-0.15773032, 1.19465634, 0]]),
+ "d": np.array([[1, 0, 0], [1.10104433, 0, -0.00901975], [0, 0, 1]], dtype=np.float16),
+ "p": np.array([[0, 0.90822864, 0.008192], [0, 1, 0], [0, 0, 1]], dtype=np.float16),
+ "t": np.array([[1, 0, 0], [0, 1, 0], [-0.15773032, 1.19465634, 0]], dtype=np.float16),
  }
  rgb2lms = np.array([[0.3904725 , 0.54990437, 0.00890159],
  [0.07092586, 0.96310739, 0.00135809],
- [0.02314268, 0.12801221, 0.93605194]])
+ [0.02314268, 0.12801221, 0.93605194]], dtype=np.float16)
  # Precomputed inverse
  lms2rgb = np.array([[ 2.85831110e+00, -1.62870796e+00, -2.48186967e-02],
  [-2.10434776e-01, 1.15841493e+00, 3.20463334e-04],
- [-4.18895045e-02, -1.18154333e-01, 1.06888657e+00]])
+ [-4.18895045e-02, -1.18154333e-01, 1.06888657e+00]], dtype=np.float16)
  # first go from RBG to LMS space
  lms = transform_colorspace(rgb, rgb2lms)
  # Calculate image as seen by the color blind
@@ -101,7 +101,6 @@ def daltonize(rgb, color_deficit='d'):
  # rgb - sim_rgb contains the color information that dichromats
  # cannot see. err2mod rotates this to a part of the spectrum that
  # they can see.
- rgb = rgb.convert('RGB')
  err = transform_colorspace(rgb - sim_rgb, err2mod)
  dtpn = err + rgb
  return dtpn
@@ -471,9 +470,6 @@ def inverse_gamma_correction(linear_rgb, gamma=2.4):
  rgb[idx, i] = 255 * 12.92 * linear_rgb[idx, i]
  idx = np.logical_not(idx)
  rgb[idx, i] = 255 * (1.055 * linear_rgb[idx, i]**(1/gamma) - 0.055)
- max_value = np.max(rgb)
- if max_value > 255:
- rgb *= 255 / max_value
  return np.round(rgb)
 
 
@@ -505,7 +501,7 @@ def main():
  if args.type is None:
  args.type = "d"
 
- orig_img = np.asarray(Image.open(args.input_image).convert("RGB"), dtype=float)
+ orig_img = np.asarray(Image.open(args.input_image).convert("RGB"), dtype=np.float16)
  orig_img = gamma_correction(orig_img, args.gamma)
 
  if args.simulate: