def test_empty_class_2d(self):
num_classes = 2
target = torch.tensor([[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0],
[0, 0, 0, 0]])
# add another dimension corresponding to the batch (batch size = 1 here)
target = target.unsqueeze(0)
pred_very_good = 1000 * F.one_hot(
target, num_classes=num_classes).permute(0, 3, 1, 2).float()
pred_very_poor = 1000 * F.one_hot(
1 - target, num_classes=num_classes).permute(0, 3, 1, 2).float()
for w_mode in ["default", "GDL"]:
# initialize the loss
loss = GeneralizedWassersteinDiceLoss(dist_matrix=np.array(
[[0.0, 1.0], [1.0, 0.0]]),
weighting_mode=w_mode)
# loss for pred_very_good should be close to 0
loss_good = float(loss.forward(pred_very_good, target))
self.assertAlmostEqual(loss_good, 0.0, places=3)
# loss for pred_very_poor should be close to 1
loss_poor = float(loss.forward(pred_very_poor, target))
self.assertAlmostEqual(loss_poor, 1.0, places=3)
def test_bin_seg_3d(self):
# define 3d examples
target = torch.tensor([
# raw 0
[[0, 0, 0, 0], [0, 1, 1, 0], [0, 1, 1, 0], [0, 0, 0, 0]],
# raw 1
[[0, 0, 0, 0], [0, 1, 1, 0], [0, 1, 1, 0], [0, 0, 0, 0]],
# raw 2
[[0, 0, 0, 0], [0, 1, 1, 0], [0, 1, 1, 0], [0, 0, 0, 0]],
])
# add another dimension corresponding to the batch (batch size = 1 here)
target = target.unsqueeze(0) # shape (1, H, W, D)
pred_very_good = 1000 * F.one_hot(target, num_classes=2).permute(
0, 4, 1, 2, 3).float()
pred_very_poor = 1000 * F.one_hot(1 - target, num_classes=2).permute(
0, 4, 1, 2, 3).float()
for w_mode in ["default", "GDL"]:
# initialize the loss
loss = GeneralizedWassersteinDiceLoss(dist_matrix=np.array(
[[0.0, 1.0], [1.0, 0.0]]),
weighting_mode=w_mode)
# mean dice loss for pred_very_good should be close to 0
loss_good = float(loss.forward(pred_very_good, target))
self.assertAlmostEqual(loss_good, 0.0, places=3)
# mean dice loss for pred_very_poor should be close to 1
loss_poor = float(loss.forward(pred_very_poor, target))
self.assertAlmostEqual(loss_poor, 1.0, places=3)
def test_bin_seg_2d(self):
target = torch.tensor([[0, 0, 0, 0], [0, 1, 1, 0], [0, 1, 1, 0],
[0, 0, 0, 0]])
# add another dimension corresponding to the batch (batch size = 1 here)
target = target.unsqueeze(0)
pred_very_good = 1000 * F.one_hot(target, num_classes=2).permute(
0, 3, 1, 2).float()
pred_very_poor = 1000 * F.one_hot(1 - target, num_classes=2).permute(
0, 3, 1, 2).float()
# initialize the loss
loss = GeneralizedWassersteinDiceLoss(
dist_matrix=np.array([[0.0, 1.0], [1.0, 0.0]]))
# the loss for pred_very_good should be close to 0
loss_good = float(loss.forward(pred_very_good, target))
self.assertAlmostEqual(loss_good, 0.0, places=3)
# same test, but with target with a class dimension
target_4dim = target.unsqueeze(1)
loss_good = float(loss.forward(pred_very_good, target_4dim))
self.assertAlmostEqual(loss_good, 0.0, places=3)
# the loss for pred_very_poor should be close to 1
loss_poor = float(loss.forward(pred_very_poor, target))
self.assertAlmostEqual(loss_poor, 1.0, places=3)
def test_different_target_data_type(self):
"""
Test if the loss is compatible with all the integer types
for the target segmentation.
"""
# define 2d examples
target = torch.tensor([[0, 0, 0, 0], [0, 1, 1, 0], [0, 1, 1, 0],
[0, 0, 0, 0]])
# add another dimension corresponding to the batch (batch size = 1 here)
target = target.unsqueeze(0) # shape (1, H, W)
pred_very_good = 1000 * F.one_hot(target, num_classes=2).permute(
0, 3, 1, 2).float()
target_uint8 = target.to(torch.uint8)
target_int8 = target.to(torch.int8)
target_short = target.short()
target_int = target.int()
target_long = target.long()
target_list = [
target_uint8, target_int8, target_short, target_int, target_long
]
for w_mode in ["default", "GDL"]:
# initialize the loss
loss = GeneralizedWassersteinDiceLoss(dist_matrix=np.array(
[[0.0, 1.0], [1.0, 0.0]]),
weighting_mode=w_mode)
# The test should pass irrespectively of the integer type used
for t in target_list:
# the loss for pred_very_good should be close to 0
loss_good = float(loss.forward(pred_very_good, t))
self.assertAlmostEqual(loss_good, 0.0, places=3)