loss.py

from __future__ import absolute_import
import torch
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd.function import Function
from torch.autograd import Variable


class OriTripletLoss(nn.Module):
    """Triplet loss with hard positive/negative mining.
    
    Reference:
    Hermans et al. In Defense of the Triplet Loss for Person Re-Identification. arXiv:1703.07737.
    Code imported from https://github.com/Cysu/open-reid/blob/master/reid/loss/triplet.py.
    
    Args:
    - margin (float): margin for triplet.
    """
    
    def __init__(self, batch_size, margin=0.3):
        super(OriTripletLoss, self).__init__()
        self.margin = margin
        self.ranking_loss = nn.MarginRankingLoss(margin=margin)

    def forward(self, inputs, targets):
        """
        Args:
        - inputs: feature matrix with shape (batch_size, feat_dim)
        - targets: ground truth labels with shape (num_classes)
        """
        n = inputs.size(0)
        
        # Compute pairwise distance, replace by the official when merged
        dist = torch.pow(inputs, 2).sum(dim=1, keepdim=True).expand(n, n)
        dist = dist + dist.t()
        dist.addmm_(1, -2, inputs, inputs.t())
        dist = dist.clamp(min=1e-12).sqrt()  # for numerical stability
        
        # For each anchor, find the hardest positive and negative
        mask = targets.expand(n, n).eq(targets.expand(n, n).t())
        dist_ap, dist_an = [], []
        for i in range(n):
            dist_ap.append(dist[i][mask[i]].max().unsqueeze(0))
            dist_an.append(dist[i][mask[i] == 0].min().unsqueeze(0))
        dist_ap = torch.cat(dist_ap)
        dist_an = torch.cat(dist_an)
        
        # Compute ranking hinge loss
        y = torch.ones_like(dist_an)
        loss = self.ranking_loss(dist_an, dist_ap, y)
        
        # compute accuracy
        correct = torch.ge(dist_an, dist_ap).sum().item()
        return loss, correct



        
        
# Adaptive weights
def softmax_weights(dist, mask):  # softmax_weights(dist_ap, is_pos)  # softmax_weights(-dist_an, is_neg)
    max_v = torch.max(dist * mask, dim=1, keepdim=True)[0]
    diff = dist - max_v
    Z = torch.sum(torch.exp(diff) * mask, dim=1, keepdim=True) + 1e-6 # avoid division by zero
    W = torch.exp(diff) * mask / Z  # 对应行相除
    return W

def normalize(x, axis=-1):
    """Normalizing to unit length along the specified dimension.
    Args:
      x: pytorch Variable
    Returns:
      x: pytorch Variable, same shape as input
    """
    x = 1. * x / (torch.norm(x, 2, axis, keepdim=True).expand_as(x) + 1e-12)
    return x

class TripletLoss_WRT(nn.Module):
    """Weighted Regularized Triplet'."""

    def __init__(self):
        super(TripletLoss_WRT, self).__init__()
        self.ranking_loss = nn.SoftMarginLoss()

    def forward(self, inputs, targets, normalize_feature=False):
        if normalize_feature:
            inputs = normalize(inputs, axis=-1)
        dist_mat = pdist_torch(inputs, inputs)  # 计算两个输入的欧氏距离，输入是两个包含32个样本的batch

        N = dist_mat.size(0)  # 尺寸为64*64 所以N为64
        # shape [N, N] # targets 为 torch.Size([64])
        # eq():逐元素的比较，若相同位置的两个元素相同，则返回True；若不同，返回False。
        # ne():逐元素的比较，若相同位置的两个元素相同，则返回False；若不同，返回True。
        # float():把true变为1.0  把false变为0.0
        is_pos = targets.expand(N, N).eq(targets.expand(N, N).t()).float()
        is_neg = targets.expand(N, N).ne(targets.expand(N, N).t()).float()

        # `dist_ap` means distance(anchor, positive)
        # both `dist_ap` and `relative_p_inds` with shape [N, 1]
        dist_ap = dist_mat * is_pos  # 64*64 anchor, positive 的距离
        dist_an = dist_mat * is_neg  # 64*64 anchor, negative 的距离

        weights_ap = softmax_weights(dist_ap, is_pos)
        weights_an = softmax_weights(-dist_an, is_neg)
        furthest_positive = torch.sum(dist_ap * weights_ap, dim=1)  # [N]
        closest_negative = torch.sum(dist_an * weights_an, dim=1)

        y = furthest_positive.new().resize_as_(furthest_positive).fill_(1)  # 创建一个与furthest_positive相等尺寸的全1张量
        loss = self.ranking_loss(closest_negative - furthest_positive, y)


        # compute accuracy
        correct = torch.ge(closest_negative, furthest_positive).sum().item()  # closest_negative > furthest_positive 输出1 选择正确
        return loss, correct
        
def pdist_torch(emb1, emb2):
    '''
    compute the euclidean（欧几里得） distance matrix between embeddings1 and embeddings2
    using gpu
    '''
    m, n = emb1.shape[0], emb2.shape[0]
    # torch.pow(emb1, 2) 是对emb1中每一个元素取2次方（平方）
    emb1_pow = torch.pow(emb1, 2).sum(dim = 1, keepdim = True).expand(m, n)  # sum后默认会降维，而keepdim=True会保证不降维
    emb2_pow = torch.pow(emb2, 2).sum(dim = 1, keepdim = True).expand(n, m).t()  # t()是求转置
    dist_mtx = emb1_pow + emb2_pow
    dist_mtx = dist_mtx.addmm_(1, -2, emb1, emb2.t())  # beta=1, alpha=-2，大于1.7.0的pytorch，可能需要把1 -2移到后面
    # dist_mtx = dist_mtx.clamp(min = 1e-12)
    dist_mtx = dist_mtx.clamp(min = 1e-12).sqrt()  # clamp(min = 1e-12)代表对dist_mtx的每个元素限制最小值不能小于1e-12，比如如果为0，则变成1e-12
    return dist_mtx    


def pdist_np(emb1, emb2):
    '''
    compute the euclidean distance matrix between embeddings1 and embeddings2
    using cpu
    '''
    m, n = emb1.shape[0], emb2.shape[0]
    emb1_pow = np.square(emb1).sum(axis = 1)[..., np.newaxis]
    emb2_pow = np.square(emb2).sum(axis = 1)[np.newaxis, ...]
    dist_mtx = -2 * np.matmul(emb1, emb2.T) + emb1_pow + emb2_pow
    # dist_mtx = np.sqrt(dist_mtx.clip(min = 1e-12))
    return dist_mtx


class CrossEntropyLabelSmooth(nn.Module):
    """Cross entropy loss with label smoothing regularizer.

    Reference:
    Szegedy et al. Rethinking the Inception Architecture for Computer Vision. CVPR 2016.
    Equation: y = (1 - epsilon) * y + epsilon / K.

    Args:
        num_classes (int): number of classes.
        epsilon (float): weight.
    """
    def __init__(self, num_classes, epsilon=0.1, use_gpu=True):
        super(CrossEntropyLabelSmooth, self).__init__()
        self.num_classes = num_classes
        self.epsilon = epsilon
        self.use_gpu = use_gpu
        self.logsoftmax = nn.LogSoftmax(dim=1)

    def forward(self, inputs, targets):
        """
        Args:
            inputs: prediction matrix (before softmax) with shape (batch_size, num_classes)
            targets: ground truth labels with shape (num_classes)
        """
        log_probs = self.logsoftmax(inputs)
        targets = torch.zeros(log_probs.size()).scatter_(1, targets.unsqueeze(1).data.cpu(), 1)
        if self.use_gpu: targets = targets.cuda()
        targets = (1 - self.epsilon) * targets + self.epsilon / self.num_classes
        loss = (- targets * log_probs).mean(0).sum()
        return loss

class CenterLoss(nn.Module):
    """Center loss.

    Reference:
    Wen et al. A Discriminative Feature Learning Approach for Deep Face Recognition. ECCV 2016.

    Args:
        num_classes (int): number of classes.
        feat_dim (int): feature dimension.
    """

    def __init__(self, num_classes=187, feat_dim=2048, use_gpu=True):
        super(CenterLoss, self).__init__()
        self.num_classes = num_classes
        self.feat_dim = feat_dim
        self.use_gpu = use_gpu

        if self.use_gpu:
            self.centers = nn.Parameter(torch.randn(self.num_classes, self.feat_dim).cuda())
        else:
            self.centers = nn.Parameter(torch.randn(self.num_classes, self.feat_dim))

    def forward(self, x, labels):
        """
        Args:
            x: feature matrix with shape (batch_size, feat_dim).
            labels: ground truth labels with shape (num_classes).
        """
        assert x.size(0) == labels.size(0), "features.size(0) is not equal to labels.size(0)"

        batch_size = x.size(0)
        distmat = torch.pow(x, 2).sum(dim=1, keepdim=True).expand(batch_size, self.num_classes) + \
                  torch.pow(self.centers, 2).sum(dim=1, keepdim=True).expand(self.num_classes, batch_size).t()
        distmat.addmm_(1, -2, x, self.centers.t())

        classes = torch.arange(self.num_classes).long()
        if self.use_gpu: classes = classes.cuda()
        labels = labels.unsqueeze(1).expand(batch_size, self.num_classes)
        mask = labels.eq(classes.expand(batch_size, self.num_classes))

        dist = distmat * mask.float()
        loss = dist.clamp(min=1e-12, max=1e+12).sum() / batch_size
        #dist = []
        #for i in range(batch_size):
        #    value = distmat[i][mask[i]]
        #    value = value.clamp(min=1e-12, max=1e+12)  # for numerical stability
        #    dist.append(value)
        #dist = torch.cat(dist)
        #loss = dist.mean()
        return loss