Unexpected numpy.unique behavior

Question

I am using numpy.unique to get values, indices and counts on a masked array that has been flattened with numpy.ravel and am getting unexpected results.

If I test it manually as in the following it works:

>>> import numpy as np
>>> a = np.array([[1,2,3],[1,0,0],[2,1,5]])
>>> a
array([[1, 2, 3],
       [1, 0, 0],
       [2, 1, 5]])
>>> src = np.ma.masked_equal(a, 0)
>>> src
masked_array(data =
 [[1 2 3]
 [1 -- --]
 [2 1 5]],
             mask =
 [[False False False]
 [False  True  True]
 [False False False]],
       fill_value = 0)

>>> src = src.ravel()
>>> src
masked_array(data = [1 2 3 1 -- -- 2 1 5],
             mask = [False False False False  True  True False False False],
       fill_value = 0)

>>> s_values, s_idx, s_counts = np.unique(src, return_inverse=True, return_counts=True)

>>> s_values
masked_array(data = [1 2 3 5 --],
             mask = [False False False False  True],
       fill_value = 0)
>>> s_counts
array([3, 2, 1, 1, 2])

However, when I apply this same logic to a uint8 array from an image file I am getting the following:

>>> src_ds = '/Users/histo/S2_10_T_DN_2016_7_27_0_4328_repro.tif'
>>> src_ds = gdal.Open(src_ds)
>>> src = src_ds.GetRasterBand(1).ReadAsArray()
>>> src = np.ma.masked_equal(src, 0)
>>> src = src.ravel()
>>> s_values, s_idx, s_counts = np.unique(src, return_index=True, return_inverse=True)
>>> s_values
masked_array(data = [3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54
 55 56 57 58 59 60 61 62 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63
 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 --
 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63
 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 --
 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63
 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 --
 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63
 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 --
 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63
 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 --
 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63
 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 --
 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63
 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 --
 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63
 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 --
 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63
 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 --
 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63
 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 --
 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63
 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 --
 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63
 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 -- 63 64 65 66
 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91
 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111
 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129
 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147
 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165
 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183
 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201
 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219
 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237
 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255],
             mask = [False False False False False False False False False False False False
 False False False False False False False False False False False False
 False False False False False False False False False False False False
 False False False False False False False False False False False False
 False False False False False False False False False False False False
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
 False  True False  True False  True False  True False  True False  True
...

For some reason 63 and mask are repeating over and over as unique values which, of course, throws out the counts and indices and makes the results unusable for any subsequent analysis.

I have tried opening the data with ndimage and the results are the same and I have tried with other images. Again, I get repeating value mask value... as unique values.

Very strange. Has anybody seen this?

Answer 1

TL;DR

Solve the issue by changing the default fill value for masked arrays:

import numpy as np

x = np.array([64, 0, 1, 2, 3, 63, 63, 0, 0, 0, 1, 2, 0, 63, 0], dtype='uint8')
y = np.ma.masked_equal(x, 0)
v, i, c = np.unique(y, return_index=True, return_counts=True)
print(v)
# [1 2 3 -- 63 -- 63 -- 64]

np.ma.core.default_filler['u'] = 0  # fill value for unsigned integers
np.ma.core.default_filler['i'] = 0  # fill value for unsigned integers

v, i, c = np.unique(y, return_index=True, return_counts=True)
print(v)
# [-- 1 2 3 63 64]

---

Details

This problem is reproducible and does not apply specifically to images (note it only happens when return_index=True ):

import numpy as np

x = np.array([64, 0, 1, 2, 3, 62, 62, 0, 0, 0, 1, 2, 0, 62, 0], dtype='uint8')
y = np.ma.masked_equal(x, 0)
v, i, c = np.unique(y, return_index=True, return_counts=True)
print(v)
# [1 2 3 62 -- 64]

x = np.array([64, 0, 1, 2, 3, 63, 63, 0, 0, 0, 1, 2, 0, 63, 0], dtype='uint8')
y = np.ma.masked_equal(x, 0)
v, i, c = np.unique(y, return_index=True, return_counts=True)
print(v)
# [1 2 3 -- 63 -- 63 -- 64]

x = np.array([64, 0, 1, 2, 3, 63, 63, 0, 0, 0, 1, 2, 0, 63, 0], dtype='uint8')
y = np.ma.masked_equal(x, 0)
v = np.unique(y)
print(v)
# [1 2 3 63 64 --]

Obviously this has to do with the number 63. What's special about 63? Binary it's all ones (with two leading zeros in case of 8 bit representation).

>>> bin(63)
'0b111111'

However, I do not know why this specific number (eg 127 works) causes weird behavior in np.unique when combined with masks.

This has to do with sorting the array. Without return_indices np.unique internally uses np.sort and otherwise it uses np.argsort :

x = np.array([64, 0, 1, 2, 3, 63, 63, 0, 0, 0, 1, 2, 0, 63, 0], dtype='uint8')
y = np.ma.masked_equal(x, 0)
print(np.sort(y))
# [1 1 2 2 3 63 63 63 64 -- -- -- -- -- --]
print(y[np.argsort(y)])
# [1 1 2 2 3 -- 63 63 -- -- -- -- 63 -- 64]

By default np.ma.argsort used the function np.ma.default_fill_value to replace masked values during sorting. The default fill value for int (which is used for uint8 too) is 999999. This is in binary 0b111101000010 00111111 - The lowest 8 bit are equal to 63! Thus, for the sorting algorithm masked values and 63 are equivalent and they are jumbled together.

It is possible to change the default fill value and thus solve the problem:

np.ma.core.default_filler['u'] = 0  # fill value for unsigned integers
np.ma.core.default_filler['i'] = 0  # fill value for unsigned integers

---

As a workaround you can convert your data to another data type such as 16 bit integers of floating point:

x = np.array([64, 0, 1, 2, 3, 63, 63, 0, 0, 0, 1, 2, 0, 63, 0], dtype='uint8')
y = np.ma.masked_equal(x.astype('int16'), 9)
v, i, c = np.unique(y, return_index=True, return_counts=True)
print(v)
# [1 2 3 63 64 --]

Another alternative for the particular use case would be not to use masked arrays at all. Since only 0 is replaced with a mask it would be easy to simply ignore 0 in the result.