def test_empty_class_2d(self):
num_classes = 2
M = np.array([[0., 1.], [1., 0.]])
# define 2d examples
target = torch.tensor([[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0],
[0, 0, 0, 0]])
if torch.cuda.is_available():
target = target.cuda()
# 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=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()
# initialize the loss
loss = GeneralizedWassersteinDiceLoss(dist_matrix=M)
# loss for pred_very_good should be close to 0
loss_good = float(loss.forward(pred_very_good, target).cpu())
self.assertAlmostEqual(loss_good, 0., places=3)
# loss for pred_very_poor should be close to 1
loss_poor = float(loss.forward(pred_very_poor, target).cpu())
self.assertAlmostEqual(loss_poor, 1., places=3)
def test_different_target_data_type(self):
"""
Test if the loss is compatible with all the integer types
for the target segmentation.
"""
M = np.array([[0., 1.], [1., 0.]])
# define 2d examples
target = torch.tensor([[0, 0, 0, 0], [0, 1, 1, 0], [0, 1, 1, 0],
[0, 0, 0, 0]])
if torch.cuda.is_available():
target = target.cuda()
# 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
]
# initialize the loss
loss = GeneralizedWassersteinDiceLoss(dist_matrix=M)
# The test should work whatever integer type is 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).cpu())
self.assertAlmostEqual(loss_good, 0., places=3)
def test_bin_seg_3d(self):
M = np.array([[0., 1.], [1., 0.]])
# 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]]
])
if torch.cuda.is_available():
target = target.cuda()
# 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 SUPPORTED_WEIGHTING:
# initialize the loss
loss = GeneralizedWassersteinDiceLoss(dist_matrix=M,
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).cpu())
self.assertAlmostEqual(loss_good, 0., places=3)
# mean dice loss for pred_very_poor should be close to 1
loss_poor = float(loss.forward(pred_very_poor, target).cpu())
self.assertAlmostEqual(loss_poor, 1., places=3)
def test_convergence(self):
"""
The goal of this test is to assess if the gradient of the loss function
is correct by testing if we can train a one layer neural network
to segment one image.
We verify that the loss is decreasing in almost all SGD steps.
"""
M = np.array([[0., 1.], [1., 0.]])
learning_rate = 0.01
max_iter = 50
# define a simple 3d example
target_seg = 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]]
])
if torch.cuda.is_available():
target_seg = target_seg.cuda()
target_seg = torch.unsqueeze(target_seg, dim=0)
image = 12 * target_seg + 27
image = image.float()
num_classes = 2
num_voxels = 3 * 4 * 4
# define a one layer model
class OnelayerNet(nn.Module):
def __init__(self):
super(OnelayerNet, self).__init__()
self.layer = nn.Linear(num_voxels, num_voxels * num_classes)
def forward(self, x):
x = x.view(-1, num_voxels)
x = self.layer(x)
x = x.view(-1, num_classes, 3, 4, 4)
return x
# initialise the network
net = OnelayerNet()
if torch.cuda.is_available():
net = net.cuda()
# initialize the loss
loss = GeneralizedWassersteinDiceLoss(dist_matrix=M)
# initialize an SGD
optimizer = optim.SGD(net.parameters(), lr=learning_rate, momentum=0.9)
# initial difference between pred and target
pred_start = torch.argmax(net(image), dim=1)
diff_start = torch.norm(pred_start.float() - target_seg.float())
loss_history = []
# train the network
for _ in range(max_iter):
# set the gradient to zero
optimizer.zero_grad()
# forward pass
output = net(image)
loss_val = loss(output, target_seg)
# backward pass
loss_val.backward()
optimizer.step()
# stats
loss_history.append(loss_val.item())
# difference between pred and target after training
pred_end = torch.argmax(net(image), dim=1)
diff_end = torch.norm(pred_end.float() - target_seg.float())
# count the number of SGD steps in which the loss decreases
num_decreasing_steps = 0
for i in range(len(loss_history) - 1):
if loss_history[i] > loss_history[i + 1]:
num_decreasing_steps += 1
decreasing_steps_ratio = float(num_decreasing_steps) / (
len(loss_history) - 1)
# verify that the loss is decreasing for sufficiently many SGD steps
self.assertTrue(decreasing_steps_ratio > 0.9)
# check that the predicted segmentation has improved
self.assertGreater(diff_start, diff_end)