def init_data(args):
# split the train set into train + val
# for CIFAR, split 5k for val
# for ImageNet, split 3k for val
val_size = 5000 if 'cifar' in args.dataset else 3000
train_loader, val_loader, _ = get_split_dataset(
args.dataset, args.batch_size,
args.n_worker, val_size,
data_root=args.data_root,
shuffle=False
) # same sampling
return train_loader, val_loader
type=float,
default=0.0,
help=
"Distance of camera from origin, default is average of z_far, z_near of dataset (only for non-DTU)",
)
parser.add_argument("--fps", type=int, default=30, help="FPS of video")
return parser
args, conf = util.args.parse_args(extra_args)
args.resume = True
device = util.get_cuda(args.gpu_id[0])
dset = get_split_dataset(args.dataset_format,
args.datadir,
want_split=args.split,
training=False)
data = dset[args.subset]
data_path = data["path"]
print("Data instance loaded:", data_path)
images = data["images"] # (NV, 3, H, W)
poses = data["poses"] # (NV, 4, 4)
focal = data["focal"]
if isinstance(focal, float):
# Dataset implementations are not consistent about
# returning float or scalar tensor in case of fx=fy
focal = torch.tensor(focal, dtype=torch.float32)
focal = focal[None]
)
parser.add_argument(
"--fixed_test",
action="store_true",
default=None,
help="Freeze encoder weights and only train MLP",
)
return parser
args, conf = util.args.parse_args(extra_args,
training=True,
default_ray_batch_size=128)
device = util.get_cuda(args.gpu_id[0])
dset, val_dset, _ = get_split_dataset(args.dataset_format, args.datadir)
print("dset z_near {}, z_far {}, lindisp {}".format(dset.z_near, dset.z_far,
dset.lindisp))
net = make_model(conf["model"]).to(device=device)
net.stop_encoder_grad = args.freeze_enc
if args.freeze_enc:
print("Encoder frozen")
net.encoder.eval()
renderer = NeRFRenderer.from_conf(
conf["renderer"],
lindisp=dset.lindisp,
).to(device=device)
# Parallize
return parser
args, conf = util.args.parse_args(
extra_args,
default_conf="conf/resnet_fine_mv.conf",
default_expname="shapenet",
)
args.resume = True
device = util.get_cuda(args.gpu_id)
extra_gpus = list(map(int, args.extra_gpus.split()))
dset = get_split_dataset(args.dataset_format,
args.datadir,
want_split=args.split)
data_loader = torch.utils.data.DataLoader(dset,
batch_size=1,
shuffle=False,
num_workers=8,
pin_memory=False)
output_dir = args.output.strip()
has_output = len(output_dir) > 0
total_psnr = 0.0
total_ssim = 0.0
cnt = 0
if has_output:
return parser
args, conf = util.args.parse_args(
extra_args, default_conf="conf/resnet_fine_mv.conf", default_expname="shapenet",
)
args.resume = True
device = util.get_cuda(args.gpu_id[0])
only_load_these_ids = None
if args.viewlist:
with open(args.viewlist) as f:
only_load_these_ids = [l.split()[1] for l in f.readlines()]
dset = get_split_dataset(
args.dataset_format, args.datadir, want_split=args.split, training=False, only_load_these_ids=only_load_these_ids
)
data_loader = torch.utils.data.DataLoader(
dset, batch_size=1, shuffle=False, num_workers=8, pin_memory=False
)
output_dir = args.output.strip()
has_output = len(output_dir) > 0
total_psnr = 0.0
total_ssim = 0.0
cnt = 0
if has_output:
finish_path = os.path.join(output_dir, "finish.txt")
os.makedirs(output_dir, exist_ok=True)
