def test_1d_coordinates(self):
img_size = (21, 1, 1, 1, 1)
win_size = (15, 1, 1)
sampling_map = np.zeros(img_size)
coords = balanced_spatial_coordinates(10, img_size, win_size,
sampling_map)
self.assertAllEqual(coords.shape, (10, N_SPATIAL))
self.assertCoordinatesAreValid(coords, sampling_map)
sampling_map[9, 0, 0, 0, 0] = 1.0
coords = balanced_spatial_coordinates(500, img_size, win_size,
sampling_map)
# better test?
self.assertTrue(np.sum(np.all(coords == [9, 0, 0], axis=1)) >= 200)
def test_classes_balances(self):
# Set the random state to prevent false positive
np.random.seed(0)
# Setting these too high inflats the run time
number_of_repetitions = 1000
samples_per_repetition = 10
num_classes = 3
# Create a map with almost all background, one pixel of each
# other label
img_size = (50, 25, 10, 1, 1)
win_size = (8, 7, 2, 1, 1)
sampling_map = np.zeros(img_size)
sampling_map[6, 5, 2:, 0, 0] = 1
sampling_map[11, 13:, 3, 0, 0] = 2
# Accumulate the number of times each class is sampled
accum = np.zeros((num_classes))
for _ in range(number_of_repetitions):
coords = balanced_spatial_coordinates(samples_per_repetition,
img_size, win_size,
sampling_map)
# Be sure to sample the correct number
self.assertAllEqual(coords.shape,
(samples_per_repetition, N_SPATIAL))
# Convert to np.ndarry indexable
for coord in coords.astype(int):
x, y, z = coord
label = int(sampling_map[x, y, z])
accum[label] = accum[label] + 1
# Each class should be within 2 decimal places of 1.0/num_classes
accum = np.divide(accum, accum.sum())
self.assertAllClose(accum,
np.ones((num_classes)) * 1.0 / num_classes,
rtol=1e-2,
atol=1e-2)