def build_data_loader():
logger.info("build train dataset")
# train_dataset
train_dataset = TrkDataset()
logger.info("build dataset done")
train_sampler = None
train_loader = DataLoader(train_dataset,
batch_size=cfg.TRAIN.BATCH_SIZE,
num_workers=cfg.TRAIN.NUM_WORKERS,
pin_memory=True,
sampler=train_sampler)
return train_loader
def get_train_dataflow():
'''
training dataflow with data augmentation.
'''
ds = TrkDataset()
train_preproc = TrainingDataPreprocessor(cfg)
if cfg.TRAIN.NUM_WORKERS == 1:
ds = MapData(ds, train_preproc)
else:
ds = MultiProcessMapDataZMQ(ds, cfg.TRAIN.NUM_WORKERS, train_preproc)
ds = BatchData(ds, cfg.TRAIN.BATCH_SIZE)
return TPIterableDataset(ds)
def build_data_loader():
logger.info("build train dataset")
reload(pysot.core.config)
reload(pysot.datasets.dataset)
from pysot.datasets.dataset import TrkDataset
train_dataset = TrkDataset()
logger.info("build dataset done")
train_sampler = None
if get_world_size() > 1:
train_sampler = DistributedSampler(train_dataset)
train_loader = DataLoader(train_dataset,
batch_size=cfg.TRAIN.BATCH_SIZE,
num_workers=cfg.TRAIN.NUM_WORKERS,
pin_memory=True,
sampler=train_sampler)
return train_loader
def build_data_loader():
'''
:return: 建立train_loader,参数在config中指定
'''
logger.info("build train dataset")
# train_dataset
train_dataset = TrkDataset(
) ##为feature map生成anchor的位置信息，通过json文件加载训练数据集（数据集合一视频为单位，保证每个视频至少一个跟踪目标，每个目标跟踪标注信息至少有一帧），设置数据增强的参数
logger.info("build dataset done")
train_sampler = None
if get_world_size() > 1:
train_sampler = DistributedSampler(train_dataset)
train_loader = DataLoader(train_dataset,
batch_size=cfg.TRAIN.BATCH_SIZE,
num_workers=cfg.TRAIN.NUM_WORKERS,
pin_memory=True,
sampler=train_sampler)
return train_loader
def build_data_loader(mode='train'):
if mode == 'train':
logger.info("build train dataset")
# train_dataset
train_dataset = TrkDataset()
logger.info("build dataset done")
else:
logger.info("build val dataset")
# train_dataset
train_dataset = ValDataset()
logger.info("build dataset done")
train_sampler = None
if get_world_size() > 1:
train_sampler = DistributedSampler(train_dataset)
train_loader = DataLoader(train_dataset,
batch_size=cfg.TRAIN.BATCH_SIZE,
num_workers=cfg.TRAIN.NUM_WORKERS,
pin_memory=True,
sampler=train_sampler)
return train_loader
def build_data_loader():
logger.info("build train dataset")
# train_dataset
train_dataset = TrkDataset()
logger.info("build dataset done")
train_sampler = None
if get_world_size() > 1:
train_sampler = DistributedSampler(train_dataset)
# dataset:PyTorch已有的数据读取接口或者自定义的数据接口的输出
# batchsize：batch块的大小
# collate_fn：用来处理不同情况下的输入dataset的封装
# num_workers: 数据导入时需要的进程数量-0表示数据导入从主进程中进行
# pin_memory:如果是True，dataloader会在返回之前将tensors复制到cuda的固定内存（pinned memory）中
# sampler：采样器
# timeout：用来设置数据读取的超时时间的，超过时间没有读取到数据就会报错
train_loader = DataLoader(train_dataset,
batch_size=cfg.TRAIN.BATCH_SIZE,
num_workers=cfg.TRAIN.NUM_WORKERS,
pin_memory=True,
sampler=train_sampler)
return train_loader