src/common/loss.py

#!/usr/bin/env python
# encoding: utf-8
'''
*Copyright (c) 2023, Alibaba Group;
*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.

@author: Hey
@email: sanyuan.**@**.com
@tel: 137****6540
@datetime: 2022/11/26 21:05
@project: DeepProtFunc
@file: loss
@desc: some loss functions
'''

import torch
import torch.nn as nn
import torch.nn.functional as F


class FocalLoss(nn.Module):
    '''
    Focal loss
    '''
    def __init__(self, alpha=1, gamma=2, normalization=False, reduce=False):
        super(FocalLoss, self).__init__()
        self.alpha = alpha
        self.gamma = gamma
        self.normalization = normalization
        self.reduce = reduce

    def forward(self, inputs, targets):
        if self.normalization:
            '''
            reduction: the operation on the output loss, which can be set to 'none', 'mean', and 'sum'; 'none'will not perform any processing on the loss, 'mean' will calculate the mean of the loss, 'sum' will sum the loss, and the default is 'mean'
            '''
            bce = F.binary_cross_entropy_with_logits(inputs, targets, reduction='none')
            probs = torch.sigmoid(inputs)
        else:
            bce = F.binary_cross_entropy(inputs, targets, reduction='none')
            probs = inputs
        pt = targets * probs + (1 - targets) * (1 - probs)
        modulate = 1 if self.gamma is None else (1 - pt) ** self.gamma

        focal_loss = modulate * bce

        if self.alpha is not None:
            assert 0 <= self.alpha <= 1
            alpha_weights = targets * self.alpha + (1 - targets) * (1 - self.alpha)
            focal_loss *= alpha_weights
        if self.reduce:
            # global mean
            return torch.mean(focal_loss)
        else:
            # sum of all samples and calc the mean value
            return torch.mean(torch.sum(focal_loss, dim=1))


class MultiLabel_CCE(nn.Module):
    '''
    multi label cce
    '''
    def __init__(self, normalization=False):
        super(MultiLabel_CCE, self).__init__()
        self.normalization = normalization

    def forward(self, inputs, targets):
        """
        Cross entropy of multi-label classification
        Note：The shapes of y_true and y_pred are consistent, and the elements of y_true are either 0 or 1. 1 indicates that the corresponding class is a target class, and 0 indicates that the corresponding class is a non-target class.
        """
        if self.normalization:
            y_pred = nn.Sigmoid()(inputs)
        else:
            y_pred = inputs
        y_true = targets
        y_pred = (1 - 2 * y_true) * y_pred
        y_pred_neg = y_pred - y_true * 1e12
        y_pred_pos = y_pred - (1 - y_true) * 1e12
        zeros = torch.zeros_like(y_pred[..., :1])
        y_pred_neg = torch.cat((y_pred_neg, zeros), axis=-1)
        y_pred_pos = torch.cat((y_pred_pos, zeros), axis=-1)
        neg_loss = torch.logsumexp(y_pred_neg,  axis=-1)
        pos_loss = torch.logsumexp(y_pred_pos,  axis=-1)
        loss = torch.mean(neg_loss + pos_loss)
        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()


class ASLSingleLabel(nn.Module):
    '''
    This loss is intended for single-label classification problems（multi-class）
    '''
    def __init__(self, gamma_pos=0, gamma_neg=4, eps: float = 0.1, reduction='mean'):
        super(ASLSingleLabel, self).__init__()

        self.eps = eps
        self.logsoftmax = nn.LogSoftmax(dim=-1)
        self.targets_classes = []
        self.gamma_pos = gamma_pos
        self.gamma_neg = gamma_neg
        self.reduction = reduction

    def forward(self, inputs, target):
        '''
        "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
        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

        if self.eps > 0:  # label smoothing
            self.targets_classes = self.targets_classes.mul(1 - self.eps).add(self.eps / num_classes)

        # loss calculation
        loss = - self.targets_classes.mul(log_preds)

        loss = loss.sum(dim=-1)
        if self.reduction == 'mean':
            loss = loss.mean()

        return loss