Asymmetric Loss

chebai/loss/asymmetric_loss.py

@@ -0,0 +1,205 @@
+import torch
+import torch.nn as nn
+
+
+class AsymmetricLoss(nn.Module):
+    """
+    Asymmetric Loss for multi-label and single-label classification tasks.
+
+    Implementation from: https://github.com/Alibaba-MIIL/ASL
+
+    Asymmetric Loss from: "Asymmetric Loss For Multi-Label Classification"
+    https://openaccess.thecvf.com/content/ICCV2021/papers/Ben-Baruch_Asymmetric_Loss_For_Multi-Label_Classification_ICCV_2021_paper.pdf
+
+    Args:
+        gamma_neg (float): Negative focusing parameter. Default is 4.
+        gamma_pos (float): Positive focusing parameter. Default is 1.
+        clip (float, optional): Asymmetric clipping value for negative probabilities. Default is 0.05.
+        eps (float, optional): Small epsilon value for numerical stability. Default is 1e-8.
+        reduction (str, optional): Specifies the reduction method: 'none' | 'mean' | 'sum'. Default is 'mean'.
+        task_type (str, optional): Type of task: 'multi-label' or 'single-label'. Default is 'multi-label'.
+        disable_torch_grad_focal_loss (bool, optional): Whether to disable gradient computation during focal loss calculation. Default is True.
+        optimized (bool, optional): Whether to use optimized version with inplace operations (only for multi-label). Default is False.
+    """
+
+    def __init__(
+        self,
+        gamma_neg=4,
+        gamma_pos=1,
+        clip=0.05,
+        eps=1e-8,
+        reduction="mean",
+        task_type="multi-label",
+        disable_torch_grad_focal_loss=True,
+        optimized=False,
+    ):
+        super(AsymmetricLoss, self).__init__()
+        self.gamma_neg = gamma_neg
+        self.gamma_pos = gamma_pos
+        self.clip = clip
+        self.eps = eps
+        self.reduction = reduction
+        self.task_type = task_type
+        self.disable_torch_grad_focal_loss = disable_torch_grad_focal_loss
+        self.optimized = optimized
+        self.logsoftmax = nn.LogSoftmax(dim=-1)
+        self.targets_classes = []
+
+        # For optimized version: pre-allocate tensors
+        if self.optimized and self.task_type == "multi-label":
+            self.targets = None
+            self.anti_targets = None
+            self.xs_pos = None
+            self.xs_neg = None
+            self.asymmetric_w = None
+            self.loss = None
+
+    def forward(self, inputs, targets, **kwargs):
+        """
+        Forward pass to compute the Asymmetric Loss.
+
+        Args:
+            inputs: Predictions (logits) from the model.
+            targets: Ground truth labels.
+            **kwargs: Additional keyword arguments (for compatibility with training framework).
+
+        Returns:
+            Loss tensor with the reduction option applied.
+        """
+        if self.task_type == "multi-label":
+            if self.optimized:
+                return self._multi_label_asymmetric_loss_optimized(inputs, targets)
+            else:
+                return self._multi_label_asymmetric_loss(inputs, targets)
+        elif self.task_type == "single-label":
+            return self._single_label_asymmetric_loss(inputs, targets)
+        else:
+            raise ValueError(
+                f"Unsupported task_type '{self.task_type}'. Use 'multi-label' or 'single-label'."
+            )
+
+    def _multi_label_asymmetric_loss(self, x, y):
+        """
+        Standard asymmetric loss for multi-label classification.
+
+        Parameters
+        ----------
+        x: input logits
+        y: targets (multi-label binarized vector)
+        """
+        # Calculating Probabilities
+        x_sigmoid = torch.sigmoid(x)
+        xs_pos = x_sigmoid
+        xs_neg = 1 - x_sigmoid
+
+        # Asymmetric Clipping
+        if self.clip is not None and self.clip > 0:
+            xs_neg = (xs_neg + self.clip).clamp(max=1)
+
+        # Basic CE calculation
+        los_pos = y * torch.log(xs_pos.clamp(min=self.eps))
+        los_neg = (1 - y) * torch.log(xs_neg.clamp(min=self.eps))
+        loss = los_pos + los_neg
+
+        # Asymmetric Focusing
+        if self.gamma_neg > 0 or self.gamma_pos > 0:
+            grad_ctx = (
+                torch.no_grad() if self.disable_torch_grad_focal_loss else nullcontext()
+            )
+            with grad_ctx:
+                pt0 = xs_pos * y
+                pt1 = xs_neg * (1 - y)
+                pt = pt0 + pt1
+                one_sided_gamma = self.gamma_pos * y + self.gamma_neg * (1 - y)
+                one_sided_w = torch.pow(1 - pt, one_sided_gamma)
+            loss *= one_sided_w
+
+        loss = -loss
+
+        if self.reduction == "mean":
+            return loss.mean()
+        elif self.reduction == "sum":
+            return loss.sum()
+        return loss
+
+    def _multi_label_asymmetric_loss_optimized(self, x, y):
+        """
+        Optimized version - minimizes memory allocation and gpu uploading,
+        favors inplace operations.
+
+        Parameters
+        ----------
+        x: input logits
+        y: targets (multi-label binarized vector)
+        """
+        self.targets = y
+        self.anti_targets = 1 - y
+
+        # Calculating Probabilities
+        self.xs_pos = torch.sigmoid(x)
+        self.xs_neg = 1.0 - self.xs_pos
+
+        # Asymmetric Clipping
+        if self.clip is not None and self.clip > 0:
+            self.xs_neg.add_(self.clip).clamp_(max=1)
+
+        # Basic CE calculation
+        self.loss = self.targets * torch.log(self.xs_pos.clamp(min=self.eps))
+        self.loss.add_(self.anti_targets * torch.log(self.xs_neg.clamp(min=self.eps)))
+
+        # Asymmetric Focusing
+        if self.gamma_neg > 0 or self.gamma_pos > 0:
+            grad_ctx = (
+                torch.no_grad() if self.disable_torch_grad_focal_loss else nullcontext()
+            )
+            with grad_ctx:
+                self.xs_pos = self.xs_pos * self.targets
+                self.xs_neg = self.xs_neg * self.anti_targets
+                self.asymmetric_w = torch.pow(1 - self.xs_pos - self.xs_neg,
+                                              self.gamma_pos * self.targets + self.gamma_neg * self.anti_targets)
+            self.loss *= self.asymmetric_w
+
+        loss = -self.loss
+
+        if self.reduction == "mean":
+            return loss.mean()
+        elif self.reduction == "sum":
+            return loss.sum()
+        return loss
+
+    def _single_label_asymmetric_loss(self, inputs, target):
+        """
+        Asymmetric loss for single-label classification problems.
+
+        "input" dimensions: - (batch_size, number_classes)
+        "target" dimensions: - (batch_size)
+        """
+        num_classes = inputs.size()[-1]
+        log_preds = self.logsoftmax(inputs)
+        self.targets_classes = torch.zeros_like(inputs).scatter_(1, target.long().unsqueeze(1), 1)
+
+        # ASL weights
+        targets = self.targets_classes
+        anti_targets = 1 - targets
+        xs_pos = torch.exp(log_preds)
+        xs_neg = 1 - xs_pos
+        grad_ctx = (
+                torch.no_grad() if self.disable_torch_grad_focal_loss else nullcontext()
+            )
+        with grad_ctx:
+            xs_pos = xs_pos * targets
+            xs_neg = xs_neg * anti_targets
+            asymmetric_w = torch.pow(1 - xs_pos - xs_neg,
+                                 self.gamma_pos * targets + self.gamma_neg * anti_targets)
+            log_preds = log_preds * asymmetric_w
+
+            # loss calculation
+            loss = -self.targets_classes.mul(log_preds)
+
+        loss = loss.sum(dim=-1)
+        if self.reduction == "mean":
+            loss = loss.mean()
+        elif self.reduction == "sum":
+            loss = loss.sum()
+
+        return loss

configs/loss/asymmetric_loss.yml

@@ -0,0 +1,3 @@
+class_path: chebai.loss.asymmetric_loss.AsymmetricLoss
+init_args:
+  task_type: multi-label                   # 'multi-label' or 'single-label'