# read mean and stdev for un-normalizing predictions
pose_stats_file = osp.join(opt.data_dir, opt.dataset, opt.scene, 'stats.txt')
pose_m, pose_s = np.loadtxt(pose_stats_file) # mean and stdev
# Load the dataset
kwargs = dict(scene=opt.scene,
data_path=opt.data_dir,
train=False,
transform=data_transform,
target_transform=target_transform,
seed=opt.seed)
if opt.model == 'AtLoc':
if opt.dataset == '7Scenes':
data_set = SevenScenes(**kwargs)
elif opt.dataset == 'RobotCar':
data_set = RobotCar(**kwargs)
elif opt.dataset == 'Topo':
data_set = Topo(**kwargs)
elif opt.dataset == 'comballaz':
data_set = Topo2(**kwargs)
elif opt.dataset == 'EPFL':
data_set = Topo3(**kwargs)
else:
raise NotImplementedError
elif opt.model == 'AtLocPlus':
kwargs = dict(kwargs,
dataset=opt.dataset,
skip=opt.skip,
steps=opt.steps,
variable_skip=opt.variable_skip)
data_set = MF(real=opt.real, **kwargs)
opt = Options().parse()
if opt.val:
print('processing VAL data using {:d} cores'.format(opt.nThreads))
else:
print('processing TRAIN data using {:d} cores'.format(opt.nThreads))
# create data loader
transform = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize(opt.cropsize),
transforms.Lambda(lambda x: np.asarray(x))
])
dset = RobotCar(scene=opt.scene,
data_path=opt.data_dir,
train=not opt.val,
transform=transform,
undistort=True)
loader = DataLoader(dset, batch_size=opt.batchsize, num_workers=opt.nThreads)
# gather information about output filenames
base_dir = osp.join(opt.data_dir, opt.dataset, opt.scene)
if opt.val:
split_filename = osp.join(base_dir, 'test_split.txt')
else:
split_filename = osp.join(base_dir, 'train_split.txt')
with open(split_filename, 'r') as f:
seqs = [l.rstrip() for l in f if not l.startswith('#')]
im_filenames = []
for seq in seqs:
'pose_stats.txt')
pose_m, pose_s = np.loadtxt(pose_stats_file) # 获取pose统计文件中的均值和标准差
# 加载测试数据集
kwargs = dict(scene=opt.scene,
data_path=opt.data_dir,
train=False,
transform=data_transform,
target_transform=target_transform,
seed=opt.seed)
if opt.model == 'AtLoc':
if opt.dataset == '7Scenes':
data_set = SevenScenes(**kwargs) # 获取处理后的7Scenes数据
elif opt.dataset == 'RobotCar':
data_set = RobotCar(**kwargs) # 获取处理后的RobotCar数据
else:
raise NotImplementedError
L = len(data_set)
# num_workers：使用的子进程数，0为不使用多进程
# 是否将tensor数据复制到CUDA pinned memory中，pin memory中的数据转到GPU中会快一些
kwargs = {'num_workers': opt.nThreads, 'pin_memory': True} if cuda else {}
# Dataset负责表示数据集，它可以每次使用__getitem__返回一个样本
# 而torch.utils.data.Dataloader提供了对batch的处理，如shuffle等
# Dataset被封装在了Dataloader中
loader = DataLoader(data_set, batch_size=1, shuffle=False, **kwargs)
pred_poses = np.zeros((L, 7)) # save all predicted poses
from torch.utils.data import DataLoader
from tools.options import Options
opt = Options().parse()
data_transform = transforms.Compose([
transforms.Resize(opt.cropsize),
transforms.RandomCrop(opt.cropsize),
transforms.ToTensor()])
# dataset loader
kwargs = dict(scene=opt.scene, data_path=opt.data_dir, train=True, real=False, transform=data_transform)
if opt.dataset == '7Scenes':
dset = SevenScenes(**kwargs)
elif opt.dataset == 'RobotCar':
dset = RobotCar(**kwargs)
else:
raise NotImplementedError
# accumulate
loader = DataLoader(dset, batch_size=opt.batch_size, num_workers=opt.nThreads)
acc = np.zeros((3, opt.cropsize, opt.cropsize))
sq_acc = np.zeros((3, opt.cropsize, opt.cropsize))
for batch_idx, (imgs, _) in enumerate(loader):
imgs = imgs.numpy()
acc += np.sum(imgs, axis=0)
sq_acc += np.sum(imgs ** 2, axis=0)
if batch_idx % 50 == 0:
print('Accumulated {:d} / {:d}'.format(batch_idx * opt.batch_size, len(dset)))
tforms.append(transforms.Normalize(mean=stats[0], std=np.sqrt(stats[1])))
data_transform = transforms.Compose(tforms)
target_transform = transforms.Lambda(lambda x: torch.from_numpy(x).float())
# Load the dataset
kwargs = dict(scene=opt.scene,
data_path=opt.data_dir,
transform=data_transform,
target_transform=target_transform,
seed=opt.seed)
if opt.model == 'AtLoc':
if opt.dataset == '7Scenes':
train_set = SevenScenes(train=True, **kwargs)
val_set = SevenScenes(train=False, **kwargs)
elif opt.dataset == 'RobotCar':
train_set = RobotCar(train=True, **kwargs)
val_set = RobotCar(train=False, **kwargs)
elif opt.dataset == 'Topo':
train_set = Topo(train=True, **kwargs)
val_set = Topo(train=False, **kwargs)
elif opt.dataset == 'comballaz':
train_set = Topo2(train=True, **kwargs)
val_set = Topo2(train=False, **kwargs)
elif opt.dataset == 'EPFL':
train_set = Topo3(train=True, **kwargs)
val_set = Topo3(train=False, **kwargs)
else:
raise NotImplementedError
elif opt.model == 'AtLocPlus':
kwargs = dict(kwargs,