def test_average_precision_real_data(self, ovule_label):
label = ovule_label[64:128, 64:128, 64:128]
ltb = LabelToAffinities((1, 2, 4, 6), aggregate_affinities=True)
pred = ltb(label)
label = torch.tensor(label.reshape((1, 1) + label.shape).astype('int64'))
pred = torch.tensor(np.expand_dims(pred, 0))
ap = BoundaryAveragePrecision()
assert ap(pred, label) > 0.5
def test_label_to_boundary_with_ignore(self):
size = 20
label = _diagonal_label_volume(size, init=-1)
transform = LabelToAffinities(offsets=(2, 4), ignore_index=-1, aggregate_affinities=True)
result = transform(label)
assert result.shape == (2,) + label.shape
assert np.array_equal(np.unique(result), [-1, 0, 1])
def test_label_to_boundary_no_aggregate(self):
size = 20
label = _diagonal_label_volume(size)
# this transform will produce 6 channels
transform = LabelToAffinities(offsets=(2, 4), aggregate_affinities=False)
result = transform(label)
assert result.shape == (6,) + label.shape
assert np.array_equal(np.unique(result), [0, 1])
def test_average_precision_real_data(self):
l_file = 'resources/sample_patch.h5'
with h5py.File(l_file, 'r') as f:
label = f['big_label'][...]
ltb = LabelToAffinities((1, 2, 4, 6), aggregate_affinities=True)
pred = ltb(label)
# don't compare instances smaller than 25K voxels
ap = BoundaryAveragePrecision(min_instance_size=25000)
assert ap(pred, label) > 0.5