Ony have anchor and pos_sample

[Meimeiainaonao]

Thanks for your contribution！

In this code, I only found utilize the anchor and pos_sample as follows:

def __getitem__(self, index):
        path, target = self.samples[index]
        cam = self.cams[index]
        # pos_path, neg_path
        pos_path = self._get_pos_sample(target, index)

        sample = self.loader(path)
        pos0 = self.loader(pos_path[0])
        pos1 = self.loader(pos_path[1])
        pos2 = self.loader(pos_path[2])
        pos3 = self.loader(pos_path[3])

and in the forward as follows:

            outputs, f = model(inputs)
            _, pf = model(pos)
            #pf = Variable( pf, requires_grad=True)
            neg_labels = pos_labels
            # hard-neg
            # ----------------------------------
            nf_data = pf # 128*512
            # 128 is too much, we use pool size = 64
            rand = np.random.permutation(4*opt.batchsize)[0:opt.poolsize]
            nf_data = nf_data[rand,:]
            neg_labels = neg_labels[rand]
            nf_t = nf_data.transpose(0,1) # 512*128
            score = torch.mm(f.data, nf_t) # cosine 32*128 
            score, rank = score.sort(dim=1, descending = True) # score high == hard
            labels_cpu = labels.cpu()
            nf_hard = torch.zeros(f.shape).cuda()
            for k in range(now_batch_size):
                hard = rank[k,:]
                for kk in hard:
                    now_label = neg_labels[kk] 
                    anchor_label = labels_cpu[k]
                    if now_label != anchor_label:
                        nf_hard[k,:] = nf_data[kk,:]
                        break

I cannot understand neg_labels = pos_labels, In the paper, neg_lable should decribe as neg_lable !=pos_label.

Please help, Thx

[CynicalHeart]

我记得当初也被pos_labels名称迷惑到了，因为它的定义在上面：

inputs, labels, pos, pos_labels = data  # anchor、batch中类别索引列表、正例、正例类别索引列表
...
pos = pos.view(4 * opt.batchsize, c, h, w)  # 正例样本[5维->4维]
# copy pos 4times 复制四次,并使四个索引连续排布.eg:[213,213,213,213,132,132,132,132...]
pos_labels = pos_labels.repeat(4).reshape(4, opt.batchsize)  # [4, b]
pos_labels = pos_labels.transpose(0, 1).reshape(4 * opt.batchsize)  # [4*b]

它的含义是一个批次所有图片的类别标签×4

[Meimeiainaonao]

非常感谢同学帮我解答，但是我还是不明白，code中对tripletfolder.py中anchor、pos、neg进行了定义:

def _get_pos_sample(self, target, index):
    pos_index = np.argwhere(self.targets == target) # 正样本 与anchor同身份但是不相似的行人
    pos_index = pos_index.flatten()
    pos_index = np.setdiff1d(pos_index, index)
    rand = np.random.permutation(len(pos_index))
    result_path = []
    for i in range(4):
       t = i%len(rand)
       tmp_index = pos_index[rand[t]]
       result_path.append(self.samples[tmp_index][0])
    return result_path

def _get_neg_sample(self, target):
    neg_index = np.argwhere(self.targets != target)  # 负样本，与anchor不同身份但是相似的行人
    neg_index = neg_index.flatten()
    rand = random.randint(0,len(neg_index)-1)
    return self.samples[neg_index[rand]]

但在__getitem__中仅使用了pos_sample，没有neg_sample呐

        def __getitem__(self, index):
             path, target = self.samples[index]
             cam = self.cams[index]
             # pos_path, neg_path
             pos_path = self._get_pos_sample(target, index) # 仅使用了pos_sample，没有neg_sample呐     
             ...
             c,h,w = pos0.shape
             pos = torch.cat((pos0.view(1,c,h,w), pos1.view(1,c,h,w), pos2.view(1,c,h,w), pos3.view(1,c,h,w)), 0)
             pos_target = target
             return sample, target, pos, pos_targe

此外，在forward内，就仅有
outputs, f = model(inputs) _, pf = model(pos)# 仅计算了pos_f,也没有计算neg_f

想请教同学这是为啥呐，code中到底哪个位置使用了neg_sample, 或者给出了neg_sample的索引呐？
不胜感激~~

[CynicalHeart]

我记得没有任何位置用上neg_sample，正负例都是用的第一个函数。

[Meimeiainaonao]

正负样本不应该分别对应anchor同身份和不同身份的行人嘛？计算loss时，这个code内正负样本选择的都是同身份中的行人呐？

        neg_labels = pos_labels

        # hard-neg
        # ----------------------------------
        nf_data = pf # 负样本的特征直接取自于正样本的特征呐？后面计算的就都是正样本中，与anchor距离最远的正样本了？
        # 128 is too much, we use pool size = 64
        rand = np.random.permutation(4*opt.batchsize)[0:opt.poolsize]
        nf_data = nf_data[rand,:]
        neg_labels = neg_labels[rand]
        nf_t = nf_data.transpose(0,1) # 512*128
        score = torch.mm(f.data, nf_t) # cosine 32*128 
        score, rank = score.sort(dim=1, descending = True) # score high == hard
        labels_cpu = labels.cpu()

[layumi]

你好 ，@Meimeiainaonao
在后面 for 循环里
这一行 https://github.com/layumi/Person-reID-triplet-loss/blob/master/train_new.py#L238