Add files via upload

src/loss_functions/losses.py

@@ -0,0 +1,252 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class Hill(nn.Module):
+    r""" Hill as described in the paper "Robust Loss Design for Multi-Label Learning with Missing Labels "
+
+    .. math::
+        Loss = y \times (1-p_{m})^\gamma\log(p_{m}) + (1-y) \times -(\lambda-p){p}^2 
+
+    where : math:`\lambda-p` is the weighting term to down-weight the loss for possibly false negatives,
+          : math:`m` is a margin parameter, 
+          : math:`\gamma` is a commonly used value same as Focal loss.
+
+    .. note::
+        Sigmoid will be done in loss. 
+
+    Args:
+        lambda (float): Specifies the down-weight term. Default: 1.5. (We did not change the value of lambda in our experiment.)
+        margin (float): Margin value. Default: 1 . (Margin value is recommended in [0.5,1.0], and different margins have little effect on the result.)
+        gamma (float): Commonly used value same as Focal loss. Default: 2
+
+    """
+
+    def __init__(self, lamb: float = 1.5, margin: float = 1.0, gamma: float = 2.0,  reduction: str = 'sum') -> None:
+        super(Hill, self).__init__()
+        self.lamb = lamb
+        self.margin = margin
+        self.gamma = gamma
+        self.reduction = reduction
+
+    def forward(self, logits, targets):
+        """
+        call function as forward
+
+        Args:
+            logits : The predicted logits before sigmoid with shape of :math:`(N, C)`
+            targets : Multi-label binarized vector with shape of :math:`(N, C)`
+
+        Returns:
+            torch.Tensor: loss
+        """
+
+        # Calculating predicted probability
+        logits_margin = logits - self.margin
+        pred_pos = torch.sigmoid(logits_margin)
+        pred_neg = torch.sigmoid(logits)
+
+        # Focal margin for postive loss
+        pt = (1 - pred_pos) * targets + (1 - targets)
+        focal_weight = pt ** self.gamma
+
+        # Hill loss calculation
+        los_pos = targets * torch.log(pred_pos)
+        los_neg = (1-targets) * -(self.lamb - pred_neg) * pred_neg ** 2
+
+        loss = -(los_pos + los_neg)
+        loss *= focal_weight
+
+        if self.reduction == 'mean':
+            return loss.mean()
+        elif self.reduction == 'sum':
+            return loss.sum()
+        else:
+            return loss
+
+
+class SPLC(nn.Module):
+    r""" SPLC loss as described in the paper "Simple Loss Design for Multi-Label Learning with Missing Labels "
+
+    .. math::
+        &L_{SPLC}^+ = loss^+(p)
+        &L_{SPLC}^- = \mathbb{I}(p\leq \tau)loss^-(p) + (1-\mathbb{I}(p\leq \tau))loss^+(p)
+
+    where :math:'\tau' is a threshold to identify missing label 
+          :math:`$\mathbb{I}(\cdot)\in\{0,1\}$` is the indicator function, 
+          :math: $loss^+(\cdot), loss^-(\cdot)$ refer to loss functions for positives and negatives, respectively.
+
+    .. note::
+        SPLC can be combinded with various multi-label loss functions. 
+        SPLC performs best combined with Focal margin loss in our paper. Code of SPLC with Focal margin loss is released here.
+        Since the first epoch can recall few missing labels with high precision, SPLC can be used ater the first epoch.
+        Sigmoid will be done in loss. 
+
+    Args:
+        tau (float): threshold value. Default: 0.6
+        change_epoch (int): which epoch to combine SPLC. Default: 1
+        margin (float): Margin value. Default: 1
+        gamma (float): Hard mining value. Default: 2
+        reduction (string, optional): Specifies the reduction to apply to the output:
+            ``'none'`` | ``'mean'`` | ``'sum'``. ``'none'``: no reduction will be applied,
+            ``'mean'``: the sum of the output will be divided by the number of
+            elements in the output, ``'sum'``: the output will be summed. Default: ``'sum'``
+
+        """
+
+    def __init__(self,
+                 tau: float = 0.7,
+                 change_epoch: int = 1,
+                 margin: float = 1.0,
+                 gamma: float = 2.0,
+                 reduction: str = 'sum') -> None:
+        super(SPLC, self).__init__()
+        self.tau = tau
+        self.change_epoch = change_epoch
+        self.margin = margin
+        self.gamma = gamma
+        self.reduction = reduction
+
+    def forward(self, logits: torch.Tensor, targets: torch.LongTensor,
+                epoch) -> torch.Tensor:
+        """
+        call function as forward
+
+        Args:
+            logits : The predicted logits before sigmoid with shape of :math:`(N, C)`
+            targets : Multi-label binarized vector with shape of :math:`(N, C)`
+            epoch : The epoch of current training.
+
+        Returns:
+            torch.Tensor: loss
+        """
+
+        # Subtract margin for positive logits
+        logits = torch.where(targets == 1, logits-self.margin, logits)
+
+        # SPLC missing label correction
+        if epoch >= self.change_epoch:
+            targets = torch.where(
+                torch.sigmoid(logits) > self.tau,
+                torch.tensor(1).cuda(), targets)
+
+        pred = torch.sigmoid(logits)
+
+        #Focal margin for postive loss
+        pt = (1 - pred) * targets + pred * (1 - targets)
+        focal_weight = pt**self.gamma
+
+        los_pos = targets * F.logsigmoid(logits)
+        los_neg = (1 - targets) * F.logsigmoid(-logits)
+
+        loss = -(los_pos + los_neg)
+        loss *= focal_weight
+
+        if self.reduction == 'mean':
+            return loss.mean()
+        elif self.reduction == 'sum':
+            return loss.sum()
+        else:
+            return loss
+
+
+class AsymmetricLoss(nn.Module):
+    def __init__(self, gamma_neg=4, gamma_pos=1, clip=0.05, eps=1e-8, disable_torch_grad_focal_loss=True):
+        super(AsymmetricLoss, self).__init__()
+
+        self.gamma_neg = gamma_neg
+        self.gamma_pos = gamma_pos
+        self.clip = clip
+        self.disable_torch_grad_focal_loss = disable_torch_grad_focal_loss
+        self.eps = eps
+
+    def forward(self, x, y):
+        """"
+        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:
+            if self.disable_torch_grad_focal_loss:
+                torch.set_grad_enabled(False)
+            pt0 = xs_pos * y
+            pt1 = xs_neg * (1 - y)  # pt = p if t > 0 else 1-p
+            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)
+            if self.disable_torch_grad_focal_loss:
+                torch.set_grad_enabled(True)
+            loss *= one_sided_w
+
+        return -loss.sum()
+
+
+class AsymmetricLossOptimized(nn.Module):
+    ''' Notice - optimized version, minimizes memory allocation and gpu uploading,
+    favors inplace operations'''
+
+    def __init__(self, gamma_neg=4, gamma_pos=1, clip=0.05, eps=1e-8, disable_torch_grad_focal_loss=False):
+        super(AsymmetricLossOptimized, self).__init__()
+
+        self.gamma_neg = gamma_neg
+        self.gamma_pos = gamma_pos
+        self.clip = clip
+        self.disable_torch_grad_focal_loss = disable_torch_grad_focal_loss
+        self.eps = eps
+
+        # prevent memory allocation and gpu uploading every iteration, and encourages inplace operations
+        self.targets = self.anti_targets = self.xs_pos = self.xs_neg = self.asymmetric_w = self.loss = None
+
+    def forward(self, x, y):
+        """"
+        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:
+            if self.disable_torch_grad_focal_loss:
+                torch.set_grad_enabled(False)
+            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)
+            if self.disable_torch_grad_focal_loss:
+                torch.set_grad_enabled(True)
+            self.loss *= self.asymmetric_w
+
+        return -self.loss.sum()