def test_embed_mws_rand(self):
from torch_em.metric import EmbeddingMWSRandMetric
emebd = torch.from_numpy(np.random.rand(self.batch_size, 6, 128, 128))
gt = make_gt((128, 128), n_batches=self.batch_size, with_channels=True)
metric = EmbeddingMWSRandMetric(delta=2.0,
offsets=self.offsets,
min_seg_size=self.min_size)
self._test_metric(emebd, gt, metric, upper_bound=1.0)
def test_mws_sbd(self):
from torch_em.metric import MWSSBDMetric
affs = torch.from_numpy(np.random.rand(self.batch_size, 5, 128, 128))
gt = make_gt((128, 128),
n_batches=self.batch_size,
with_channels=True,
with_background=True)
metric = MWSSBDMetric(self.offsets, self.min_size)
self._test_metric(affs, gt, metric, upper_bound=1.0)
def test_hdbscan_sbd(self):
from torch_em.metric import HDBScanSBDMetric
embed = torch.from_numpy(np.random.rand(self.batch_size, 6, 128, 128))
gt = make_gt((128, 128),
n_batches=self.batch_size,
with_channels=True,
with_background=True)
metric = HDBScanSBDMetric(min_size=50, eps=1.0e-4)
self._test_metric(embed, gt, metric, upper_bound=1.0)
def _test_spoco(self, aux_loss):
from torch_em.loss import SPOCOLoss
loss = SPOCOLoss(delta_var=0.75, delta_dist=2.0, aux_loss=aux_loss)
input1 = torch.from_numpy(np.random.rand(2, 8, 64, 64))
input1.requires_grad = True
input1.retain_grad = True
input2 = torch.from_numpy(np.random.rand(2, 8, 64, 64))
target = make_gt(
(64, 64), n_batches=2, with_channels=True, dtype="int64") - 1
assert target.min() == 0
lval = loss((input1, input2), target)
self.assertNotEqual(lval.item(), 0.0)
lval.backward()
grads = input1.grad
self.assertEqual(grads.shape, input1.shape)
self.assertFalse(np.allclose(grads.numpy(), 0))
def test_multicut_voi(self):
from torch_em.metric import MulticutVOIMetric
bd = torch.from_numpy(np.random.rand(self.batch_size, 1, 128, 128))
gt = make_gt((128, 128), n_batches=self.batch_size, with_channels=True)
metric = MulticutVOIMetric(self.min_size)
self._test_metric(bd, gt, metric)
def test_mws_voi(self):
from torch_em.metric import MWSVOIMetric
affs = torch.from_numpy(np.random.rand(self.batch_size, 4, 128, 128))
gt = make_gt((128, 128), n_batches=self.batch_size, with_channels=True)
metric = MWSVOIMetric(self.offsets, min_seg_size=self.min_size)
self._test_metric(affs, gt, metric)