losses.py

#!/usr/bin/env python
# coding: utf-8

# In[2]:


import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable


# https://github.com/CoinCheung/pytorch-loss/blob/master/focal_loss.py
class FocalLoss(nn.Module):

    def __init__(self,
                 alpha=0.25,
                 gamma=2,
                 reduction='mean',):
        super(FocalLoss, self).__init__()
        self.alpha = alpha
        self.gamma = gamma
        self.reduction = reduction
        self.crit = nn.BCEWithLogitsLoss(reduction='none')

    def forward(self, logits, label):
        '''
        Usage is same as nn.BCEWithLogits:
            >>> criteria = FocalLossV1()
            >>> logits = torch.randn(8, 19, 384, 384)
            >>> lbs = torch.randint(0, 2, (8, 19, 384, 384)).float()
            >>> loss = criteria(logits, lbs)
        '''
        probs = torch.sigmoid(logits)
        coeff = torch.abs(label - probs).pow(self.gamma).neg()
        log_probs = torch.where(logits >= 0,
                F.softplus(logits, -1, 50),
                logits - F.softplus(logits, 1, 50))
        log_1_probs = torch.where(logits >= 0,
                -logits + F.softplus(logits, -1, 50),
                -F.softplus(logits, 1, 50))
        loss = label * self.alpha * log_probs + (1. - label) * (1. - self.alpha) * log_1_probs
        loss = loss * coeff

        if self.reduction == 'mean':
            loss = loss.mean()
        if self.reduction == 'sum':
            loss = loss.sum()
        return loss # 하나의 mini-batch 만큼의 loss 평균


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._C.set_grad_enabled(False)
                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._C.set_grad_enabled(True)
                torch.set_grad_enabled(True)
            loss *= one_sided_w
        return -loss.mean() # .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._C.set_grad_enabled(False)
                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._C.set_grad_enabled(True)
                torch.set_grad_enabled(True)
            self.loss *= self.asymmetric_w

        return -self.loss.mean() #.sum()