triplet loss

---

---

为了打好基础，还是得深入理解代码啊啊啊，虽然看到代码都头疼，能咋整，还是一点一点来吧T-T，仅自留，毕竟我还是研0的小白。

---

先写loss部分吧，后面再慢慢写model、train、test部分~（师兄说主要看这四个部分就可啦~）

---

一、loss部分

首先就是一个难样本挖掘的三元组损失。

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)      # 获得一个简单的距离triplet函数

    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)                                                         #n即batch_size
        
        # Compute pairwise distance, replace by the official when merged
        dist = torch.pow(inputs, 2).sum(dim=1, keepdim=True).expand(n, n)     # 每个数平方后， sum（保持行数n不变），再扩展成nxn维
        dist = dist + dist.t()                                                 #dis[i][j]代表的是第i个特征与第j个特征的平方的和
        dist.addmm_(1, -2, inputs, inputs.t())                                # 然后减去2倍的 第i个特征*第j个特征 从而通过完全平方式得到 (a-b)^2
        dist = dist.clamp(min=1e-12).sqrt()  # for numerical stability         #开方，clamp做简单数值处理(为了数值的稳定性)：小于min参数的dist元素值由min值取代。
                                               #根号下不能为0，0开根号没有问题的，但是梯度反向传播就会导致无穷大。
        # For each anchor, find the hardest positive and negative
        mask = targets.expand(n, n).eq(targets.expand(n, n).t())   #targets：样本对应的标签（ground truth labels with shape——num_classes）
        # targets有n个类别，所以将它扩展成n*n的矩阵，判断该矩阵和转置矩阵对应元素之间是否相等
        # 是否属于同一类别

        dist_ap, dist_an = [], []

        #分别提取出正样本和负样本
        #先过滤掉和它不同类别的样本对应的距离，剩下的就是和它同一类别的positive，然后再在剩下的positive中找到距离值最大的，就是我们需要的hard positive
        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))
        # cat使用用于将所有的hard样本距离拼接起来
        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

具体小细节：

---

然后就是叶茫这个baseline用的WRT损失喽
对着公式看：

我感觉吧，这部分和上面那个三元组本组上其实并没有上面太大区别，就加了个权重，加了个softmax

# Adaptive weights
def softmax_weights(dist, mask):
    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
    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)

        N = dist_mat.size(0)
        # shape [N, N]
        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
        dist_an = dist_mat * is_neg

        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)
        closest_negative = torch.sum(dist_an * weights_an, dim=1)

        y = furthest_positive.new().resize_as_(furthest_positive).fill_(1)
        loss = self.ranking_loss(closest_negative - furthest_positive, y)


        # compute accuracy
        correct = torch.ge(closest_negative, furthest_positive).sum().item()       #ge（）逐元素比较
        return loss, correct
        
def pdist_torch(emb1, emb2):
    '''
    compute the eucilidean distance matrix between embeddings1 and embeddings2
    using gpu
    '''
    m, n = emb1.shape[0], emb2.shape[0]
    emb1_pow = torch.pow(emb1, 2).sum(dim = 1, keepdim = True).expand(m, n)
    emb2_pow = torch.pow(emb2, 2).sum(dim = 1, keepdim = True).expand(n, m).t()
    dist_mtx = emb1_pow + emb2_pow
    dist_mtx = dist_mtx.addmm_(1, -2, emb1, emb2.t())
    # dist_mtx = dist_mtx.clamp(min = 1e-12)
    dist_mtx = dist_mtx.clamp(min = 1e-12).sqrt()
    return dist_mtx    


def pdist_np(emb1, emb2):
    '''
    compute the eucilidean 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