Focal loss (#325)

* added the Focal loss

* focal loss inherits from _WeightedLoss and improved documentation

* [MONAI] python code formatting

Co-authored-by: monai-bot <[email protected]>

docs/source/losses.rst

@@ -22,6 +22,10 @@ Segmentation Losses
 .. autoclass:: GeneralizedDiceLoss
     :members:
 
+`FocalLoss`
+~~~~~~~~~~~
+.. autoclass:: monai.losses.focal_loss.FocalLoss
+    :members:
 
 .. automodule:: monai.losses.tversky
 .. currentmodule:: monai.losses.tversky

monai/losses/__init__.py

@@ -10,4 +10,5 @@
 # limitations under the License.
 
 from .dice import *
+from .focal_loss import *
 from .tversky import *

monai/losses/focal_loss.py

@@ -0,0 +1,98 @@
+# Copyright 2020 MONAI Consortium
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#     http://www.apache.org/licenses/LICENSE-2.0
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn.functional as F
+from torch.nn.modules.loss import _WeightedLoss
+
+
+class FocalLoss(_WeightedLoss):
+    """
+    PyTorch implementation of the Focal Loss.
+    [1] "Focal Loss for Dense Object Detection", T. Lin et al., ICCV 2017
+    """
+
+    def __init__(self, gamma=2.0, weight=None, reduction="mean"):
+        """
+        Args:
+            gamma: (float) value of the exponent gamma in the definition
+            of the Focal loss.
+            weight: (tensor) weights to apply to the
+            voxels of each class. If None no weights are applied.
+            This corresponds to the weights \alpha in [1].
+            reduction: (string) reduction operation to apply on the loss batch.
+            It can be 'mean', 'sum' or 'none' as in the standard PyTorch API
+            for loss functions.
+
+        Exemple:
+            pred = torch.tensor([[1, 0], [0, 1], [1, 0]], dtype=torch.float32)
+            grnd = torch.tensor([0, 1 ,0], dtype=torch.int64)
+            fl = FocalLoss()
+            fl(pred, grnd)
+        """
+        super(FocalLoss, self).__init__(weight=weight, reduction=reduction)
+        self.gamma = gamma
+
+    def forward(self, input, target):
+        """
+        Args:
+            input: (tensor): the shape should be BNH[WD].
+            target: (tensor): the shape should be BNH[WD].
+        """
+        i = input
+        t = target
+
+        if t.dim() < i.dim():
+            # Add a class dimension to the ground-truth segmentation.
+            t = t.unsqueeze(1)  # N,H,W => N,1,H,W
+
+        # Change the shape of input and target to
+        # num_batch x num_class x num_voxels.
+        if input.dim() > 2:
+            i = i.view(i.size(0), i.size(1), -1)  # N,C,H,W => N,C,H*W
+            t = t.view(t.size(0), t.size(1), -1)  # N,1,H,W => N,1,H*W
+        else:  # Compatibility with classification.
+            i = i.unsqueeze(2)  # N,C => N,C,1
+            t = t.unsqueeze(2)  # N,1 => N,1,1
+
+        # Compute the log proba (more stable numerically than softmax).
+        logpt = F.log_softmax(i, dim=1)  # N,C,H*W
+        # Keep only log proba values of the ground truth class for each voxel.
+        logpt = logpt.gather(1, t)  # N,C,H*W => N,1,H*W
+        logpt = torch.squeeze(logpt, dim=1)  # N,1,H*W => N,H*W
+
+        # Get the proba
+        pt = torch.exp(logpt)  # N,H*W
+
+        if self.weight is not None:
+            if self.weight.type() != i.data.type():
+                self.weight = self.weight.type_as(i.data)
+            # Convert the weight to a map in which each voxel
+            # has the weight associated with the ground-truth label
+            # associated with this voxel in target.
+            at = self.weight[None, :, None]  # C => 1,C,1
+            at = at.expand((t.size(0), -1, t.size(2)))  # 1,C,1 => N,C,H*W
+            at = at.gather(1, t.data)  # selection of the weights  => N,1,H*W
+            at = torch.squeeze(at, dim=1)  # N,1,H*W => N,H*W
+            # Multiply the log proba by their weights.
+            logpt = logpt * at
+
+        # Compute the loss mini-batch.
+        weight = torch.pow(-pt + 1.0, self.gamma)
+        loss = torch.mean(-weight * logpt, dim=1)  # N
+
+        if self.reduction == "sum":
+            return loss.sum()
+        elif self.reduction == "none":
+            return loss
+        # Default is mean reduction.
+        else:
+            return loss.mean()

tests/test_focal_loss.py

@@ -0,0 +1,218 @@
+# Copyright 2020 MONAI Consortium
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#     http://www.apache.org/licenses/LICENSE-2.0
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch.nn.functional as F
+import torch.nn as nn
+import torch.optim as optim
+import torch
+import unittest
+from monai.losses import FocalLoss
+
+
+class TestFocalLoss(unittest.TestCase):
+    def test_consistency_with_cross_entropy_2d(self):
+        # For gamma=0 the focal loss reduces to the cross entropy loss
+        focal_loss = FocalLoss(gamma=0.0, reduction="mean")
+        ce = nn.CrossEntropyLoss(reduction="mean")
+        max_error = 0
+        class_num = 10
+        batch_size = 128
+        for _ in range(100):
+            # Create a random tensor of shape (batch_size, class_num, 8, 4)
+            x = torch.rand(batch_size, class_num, 8, 4, requires_grad=True)
+            # Create a random batch of classes
+            l = torch.randint(low=0, high=class_num, size=(batch_size, 8, 4))
+            l = l.long()
+            if torch.cuda.is_available():
+                x = x.cuda()
+                l = l.cuda()
+            output0 = focal_loss.forward(x, l)
+            output1 = ce.forward(x, l)
+            a = float(output0.cpu().detach())
+            b = float(output1.cpu().detach())
+            if abs(a - b) > max_error:
+                max_error = abs(a - b)
+        self.assertAlmostEqual(max_error, 0.0, places=3)
+
+    def test_consistency_with_cross_entropy_classification(self):
+        # for gamma=0 the focal loss reduces to the cross entropy loss
+        focal_loss = FocalLoss(gamma=0.0, reduction="mean")
+        ce = nn.CrossEntropyLoss(reduction="mean")
+        max_error = 0
+        class_num = 10
+        batch_size = 128
+        for _ in range(100):
+            # Create a random scores tensor of shape (batch_size, class_num)
+            x = torch.rand(batch_size, class_num, requires_grad=True)
+            # Create a random batch of classes
+            l = torch.randint(low=0, high=class_num, size=(batch_size,))
+            l = l.long()
+            if torch.cuda.is_available():
+                x = x.cuda()
+                l = l.cuda()
+            output0 = focal_loss.forward(x, l)
+            output1 = ce.forward(x, l)
+            a = float(output0.cpu().detach())
+            b = float(output1.cpu().detach())
+            if abs(a - b) > max_error:
+                max_error = abs(a - b)
+        self.assertAlmostEqual(max_error, 0.0, places=3)
+
+    def test_bin_seg_2d(self):
+        # 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()
+
+        # initialize the mean dice loss
+        loss = FocalLoss()
+
+        # focal loss for pred_very_good should be close to 0
+        focal_loss_good = float(loss.forward(pred_very_good, target).cpu())
+        self.assertAlmostEqual(focal_loss_good, 0.0, places=3)
+
+        # Same test, but for target with a class dimension
+        target = target.unsqueeze(1)  # shape (1, 1, H, W)
+        focal_loss_good = float(loss.forward(pred_very_good, target).cpu())
+        self.assertAlmostEqual(focal_loss_good, 0.0, places=3)
+
+    def test_empty_class_2d(self):
+        num_classes = 2
+        # define 2d examples
+        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)  # shape (1, H, W)
+        pred_very_good = 1000 * F.one_hot(target, num_classes=num_classes).permute(0, 3, 1, 2).float()
+
+        # initialize the mean dice loss
+        loss = FocalLoss()
+
+        # focal loss for pred_very_good should be close to 0
+        focal_loss_good = float(loss.forward(pred_very_good, target).cpu())
+        self.assertAlmostEqual(focal_loss_good, 0.0, places=3)
+
+    def test_multi_class_seg_2d(self):
+        num_classes = 6  # labels 0 to 5
+        # define 2d examples
+        target = torch.tensor([[0, 0, 0, 0], [0, 1, 2, 0], [0, 3, 4, 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=num_classes).permute(0, 3, 1, 2).float()
+
+        # initialize the mean dice loss
+        loss = FocalLoss()
+
+        # focal loss for pred_very_good should be close to 0
+        focal_loss_good = float(loss.forward(pred_very_good, target).cpu())
+        self.assertAlmostEqual(focal_loss_good, 0.0, places=3)
+
+    def test_bin_seg_3d(self):
+        # define 2d 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()
+
+        # initialize the mean dice loss
+        loss = FocalLoss()
+
+        # focal loss for pred_very_good should be close to 0
+        focal_loss_good = float(loss.forward(pred_very_good, target).cpu())
+        self.assertAlmostEqual(focal_loss_good, 0.0, 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.
+        """
+        learning_rate = 0.001
+        max_iter = 20
+
+        # 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]],
+            ]
+        )
+        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()
+
+        # initialize the loss
+        loss = FocalLoss()
+
+        # initialize an SGD
+        optimizer = optim.SGD(net.parameters(), lr=learning_rate, momentum=0.9)
+
+        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())
+
+        # 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)
+
+
+if __name__ == "__main__":
+    unittest.main()