Triplet Loss 和 Center Loss以及在reid中的应用过程

1、
Triplet Loss 和 Center Loss详解和pytorch实现

2、
Bag of Tricks and A Strong Baseline for Deep Person Re-identification（论文阅读笔记）

triploss ， p k 实现是数据处理的时候通过 sampler 实现的 p 个人， k 张图片读取数据

if 'triplet' in cfg.DATALOADER.SAMPLER:  #这里实现 多少个人， 每个人多少张图片的batchsize
    train_loader = DataLoader(
        train_set, batch_size=cfg.SOLVER.IMS_PER_BATCH,
        sampler=RandomIdentitySampler(dataset.train, cfg.SOLVER.IMS_PER_BATCH, cfg.DATALOADER.NUM_INSTANCE),
        num_workers=num_workers, collate_fn=train_collate_fn
    )
elif cfg.DATALOADER.SAMPLER == 'softmax':
    print('using softmax sampler')
    train_loader = DataLoader(
        train_set, batch_size=cfg.SOLVER.IMS_PER_BATCH, shuffle=True, num_workers=num_workers,
        collate_fn=train_collate_fn
    )


class RandomIdentitySampler(Sampler):
    """
    Randomly sample N identities, then for each identity,
    randomly sample K instances, therefore batch size is N*K.
    Args:
    - data_source (list): list of (img_path, pid, camid).
    - num_instances (int): number of instances per identity in a batch.
    - batch_size (int): number of examples in a batch.
    """

    def __init__(self, data_source, batch_size, num_instances):  #默认参数是128  8， k的影响很大，16个人，每个人8张图片，
        self.data_source = data_source
        self.batch_size = batch_size
        self.num_instances = num_instances
        self.num_pids_per_batch = self.batch_size // self.num_instances
        self.index_dic = defaultdict(list) #dict with list value
        #{783: [0, 5, 116, 876, 1554, 2041],...,}
        for index, (_, pid, _) in enumerate(self.data_source):
            self.index_dic[pid].append(index)
        self.pids = list(self.index_dic.keys())

        # estimate number of examples in an epoch
        self.length = 0
        for pid in self.pids:
            idxs = self.index_dic[pid]
            num = len(idxs)
            if num < self.num_instances:
                num = self.num_instances
            self.length += num - num % self.num_instances

    def __iter__(self):
        batch_idxs_dict = defaultdict(list)

        for pid in self.pids:
            idxs = copy.deepcopy(self.index_dic[pid])
            if len(idxs) < self.num_instances:
                idxs = np.random.choice(idxs, size=self.num_instances, replace=True)
            random.shuffle(idxs)
            batch_idxs = []
            for idx in idxs:
                batch_idxs.append(idx)
                if len(batch_idxs) == self.num_instances:
                    batch_idxs_dict[pid].append(batch_idxs)
                    batch_idxs = []

        avai_pids = copy.deepcopy(self.pids)
        final_idxs = []

        while len(avai_pids) >= self.num_pids_per_batch:
            selected_pids = random.sample(avai_pids, self.num_pids_per_batch)
            for pid in selected_pids:
                batch_idxs = batch_idxs_dict[pid].pop(0)
                final_idxs.extend(batch_idxs)
                if len(batch_idxs_dict[pid]) == 0:
                    avai_pids.remove(pid)

        return iter(final_idxs)

    def __len__(self):
        return self.length