Unexpected numpy.unique behavior with return_inverse=True on uint8 dtype #8664
Description
numpy.unique returns erroneous results when the number 63 is encountered in masked uint8 array with return_index=True:
>>> 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 -- `This is reproducible:
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 --]A workaround has been suggested on stackoverflow by changing the dtype to int16:
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'), 0)
v, i, c = np.unique(y, return_index=True, return_counts=True)
print(v)
# [1 2 3 63 64 --]This works but greatly increases the memory footprint which is undesirable.