loss.py

from typing import Optional, Sequence

import torch
from torch import Tensor
from torch import nn
from torch.nn import functional as F



class HeatmapLoss(nn.Module):
    """
    loss for detection heatmap
    """
    def __init__(self):
        super(HeatmapLoss, self).__init__()

    def forward(self, pred, gt):
        loss = ((pred - gt)**2)
        loss = loss.mean(dim=3).mean(dim=2).mean(dim=1).mean(dim=0)
        return  loss

class FocalLoss(nn.Module):
  '''nn.Module warpper for focal loss'''
  def __init__(self):
    super(FocalLoss, self).__init__()

  def forward(self, pred, gt):
    ''' Modified focal loss. Exactly the same as CornerNet.
      Runs faster and costs a little bit more memory
      Arguments:
       pred (batch x c x h x w)
       gt (batch x c x h x w)
     '''
    # https://arxiv.org/pdf/1808.01244.pdf (1)
    alpha = 2
    beta = 4

    pos_inds = gt.eq(1).float()
    pos_loss = torch.log(pred) * torch.pow(1 - pred, alpha) * pos_inds
    pos_loss = pos_loss.sum()

    neg_inds = gt.lt(1).float()
    neg_weights = torch.pow(1 - gt, beta)
    neg_loss = torch.log(1 - pred) * torch.pow(pred, alpha) * neg_weights * neg_inds

    loss = 0

    num_pos  = pos_inds.float().sum()
    neg_loss = neg_loss.sum()

    if num_pos == 0:
        loss = loss - neg_loss
    else:
        loss = loss - (pos_loss + neg_loss) / num_pos
    return loss


class MultipleFocalLoss(nn.Module):
    """ Focal Loss, as described in https://arxiv.org/abs/1708.02002.

    It is essentially an enhancement to cross entropy loss and is
    useful for classification tasks when there is a large class imbalance.
    x is expected to contain raw, unnormalized scores for each class.
    y is expected to contain class labels.

    Shape:
        - x: (batch_size, C) or (batch_size, C, d1, d2, ..., dK), K > 0.
        - y: (batch_size,) or (batch_size, d1, d2, ..., dK), K > 0.
    """

    def __init__(self,
                 alpha: Optional[Tensor] = None,
                 gamma: float = 0.,
                 reduction: str = 'mean',
                 ignore_index: int = -100):
        """Constructor.

        Args:
            alpha (Tensor, optional): Weights for each class. Defaults to None.
            gamma (float, optional): A constant, as described in the paper.
                Defaults to 0.
            reduction (str, optional): 'mean', 'sum' or 'none'.
                Defaults to 'mean'.
            ignore_index (int, optional): class label to ignore.
                Defaults to -100.
        """
        if reduction not in ('mean', 'sum', 'none'):
            raise ValueError(
                'Reduction must be one of: "mean", "sum", "none".')

        super().__init__()
        self.alpha = alpha
        self.gamma = gamma
        self.ignore_index = ignore_index
        self.reduction = reduction

        self.nll_loss = nn.NLLLoss(
            weight=alpha, reduction='none', ignore_index=ignore_index)

    def __repr__(self):
        arg_keys = ['alpha', 'gamma', 'ignore_index', 'reduction']
        arg_vals = [self.__dict__[k] for k in arg_keys]
        arg_strs = [f'{k}={v}' for k, v in zip(arg_keys, arg_vals)]
        arg_str = ', '.join(arg_strs)
        return f'{type(self).__name__}({arg_str})'

    def forward(self, x: Tensor, y: Tensor) -> Tensor:
        if x.ndim > 2:
            # (N, C, d1, d2, ..., dK) --> (N * d1 * ... * dK, C)
            c = x.shape[1]
            x = x.permute(0, *range(2, x.ndim), 1).reshape(-1, c)
            # (N, d1, d2, ..., dK) --> (N * d1 * ... * dK,)
            y = y.view(-1)

        unignored_mask = y != self.ignore_index
        y = y[unignored_mask]
        if len(y) == 0:
            return 0.
        x = x[unignored_mask]

        # compute weighted cross entropy term: -alpha * log(pt)
        # (alpha is already part of self.nll_loss)
        log_p = F.log_softmax(x, dim=-1)
        ce = self.nll_loss(log_p, y)

        # get true class column from each row
        all_rows = torch.arange(len(x))
        log_pt = log_p[all_rows, y]

        # compute focal term: (1 - pt)^gamma
        pt = log_pt.exp()
        focal_term = (1 - pt)**self.gamma

        # the full loss: -alpha * ((1 - pt)^gamma) * log(pt)
        loss = focal_term * ce

        if self.reduction == 'mean':
            loss = loss.mean()
        elif self.reduction == 'sum':
            loss = loss.sum()

        return loss


def focal_loss(alpha: Optional[Sequence] = None,
               gamma: float = 0.,
               reduction: str = 'mean',
               ignore_index: int = -100,
               device='cpu',
               dtype=torch.float32) -> FocalLoss:
    """Factory function for FocalLoss.

    Args:
        alpha (Sequence, optional): Weights for each class. Will be converted
            to a Tensor if not None. Defaults to None.
        gamma (float, optional): A constant, as described in the paper.
            Defaults to 0.
        reduction (str, optional): 'mean', 'sum' or 'none'.
            Defaults to 'mean'.
        ignore_index (int, optional): class label to ignore.
            Defaults to -100.
        device (str, optional): Device to move alpha to. Defaults to 'cpu'.
        dtype (torch.dtype, optional): dtype to cast alpha to.
            Defaults to torch.float32.

    Returns:
        A FocalLoss object
    """
    if alpha is not None:
        if not isinstance(alpha, Tensor):
            alpha = torch.tensor(alpha)
        alpha = alpha.to(device=device, dtype=dtype)

    fl = MultipleFocalLoss(
        alpha=alpha,
        gamma=gamma,
        reduction=reduction,
        ignore_index=ignore_index)
    return fl