def test_01_lwg_trainer(self):
opt = EasyDict()
opt.cfg_path = self.cfg_path
opt.gpu_ids = "2"
opt.model_id = "debug"
opt.output_dir = "../tests/data"
opt.verbose = True
opt.temporal = False
opt.load_iter = -1
cfg = setup(opt)
device = torch.device("cuda:{}".format(cfg.local_rank))
lwg_trainer = LWGTrainer(cfg, device)
lwg_trainer.gpu_wrapper()
def process_data():
global args, iPER_train_txt, iPER_val_txt
# 1. dumps videos to frames
# 1. prepreocess
vid_names = get_video_dirs(iPER_train_txt) + get_video_dirs(iPER_val_txt)
src_paths = prepare_src_path(vid_names)
args.src_path = "|".join(src_paths)
# print(args.src_path)
# set this as empty when preprocessing the training dataset.
args.ref_path = ""
cfg = setup(args)
# 1. human estimation, including 2D pose, tracking, 3D pose, parsing, and front estimation.
human_estimate(opt=cfg)
# 2. digital deformation.
digital_deform(opt=cfg)
# 3. check
meta_src_proc = cfg.meta_data["meta_src"]
invalid_meta_process = []
print(f"Check all videos have been processed...")
for meta_proc in tqdm(meta_src_proc):
process_info = ProcessInfo(meta_proc)
process_info.deserialize()
# check it has been processed successfully
if not process_info.check_has_been_processed(process_info.vid_infos,
verbose=False):
invalid_meta_process.append(meta_proc)
num_invalid = len(invalid_meta_process)
if num_invalid > 0:
for meta_proc in invalid_meta_process:
print(f"invalid meta proc {meta_proc}")
else:
print(f"process successfully.")
def process_data():
# 1. preprocess
src_paths = prepare_src_path()
args.src_path = "|".join(src_paths)
print(args.src_path)
# set this as empty when preprocessing the training dataset.
args.ref_path = ""
cfg = setup(args)
# 1. human estimation, including 2D pose, tracking, 3D pose, parsing, and front estimation.
human_estimate(opt=cfg)
# 2. digital deformation.
digital_deform(opt=cfg)
# 3. check
meta_src_proc = cfg.meta_data["meta_src"]
invalid_meta_process = []
for meta_proc in meta_src_proc:
process_info = ProcessInfo(meta_proc)
process_info.deserialize()
# check it has been processed successfully
if not process_info.check_has_been_processed(process_info.vid_infos,
verbose=False):
invalid_meta_process.append(meta_proc)
num_invalid = len(invalid_meta_process)
if num_invalid > 0:
for meta_proc in invalid_meta_process:
print(f"invalid meta proc {meta_proc}")
else:
print(f"process successfully.")
parser.add_argument("--ref_path", type=str, default="", help=ref_path_format)
args = parser.parse_args()
# symlink from the actual assets directory to this current directory
work_asserts_dir = os.path.join("./assets")
if not os.path.exists(work_asserts_dir):
os.symlink(osp.abspath(args.assets_dir),
osp.abspath(work_asserts_dir),
target_is_directory=(platform.system() == "Windows"))
args.cfg_path = osp.join(work_asserts_dir, "configs", "deploy.toml")
if __name__ == "__main__":
# run imitator
cfg = setup(args)
run_imitator(cfg)
# # or use the system call wrapper
# cmd = [
# sys.executable, "-m", "iPERCore.services.run_imitator",
# "--gpu_ids", args.gpu_ids,
# "--image_size", str(args.image_size),
# "--num_source", str(args.num_source),
# "--output_dir", args.output_dir,
# "--model_id", args.model_id,
# "--src_path", args.src_path,
# "--ref_path", args.ref_path
# ]
#
# subprocess.call(cmd)
def main():
# set this as empty when preprocessing the training dataset.
args.ref_path = ""
cfg = setup(args)
cfg.num_source = 4
cfg.time_step = 2
dataset = DatasetFactory.get_by_name(cfg.dataset_mode,
cfg,
is_for_train=True)
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=2,
num_workers=4,
pin_memory=True)
device = torch.device("cuda:0")
flow_comp = FlowCompositionForTrainer(cfg)
flow_comp = flow_comp.to(device)
for inputs in tqdm(dataloader):
images = inputs["images"].to(device, non_blocking=True)
smpls = inputs["smpls"].to(device, non_blocking=True)
masks = inputs["masks"].to(device, non_blocking=True)
offsets = inputs["offsets"].to(
device, non_blocking=True) if "offsets" in inputs else 0
links_ids = inputs["links_ids"].to(
device, non_blocking=True) if "links_ids" in inputs else None
aug_bg = inputs["bg"].to(device,
non_blocking=True) if "bg" in inputs else None
ns = cfg.num_source
src_img = images[:, 0:ns].contiguous()
src_smpl = smpls[:, 0:ns].contiguous()
tsf_img = images[:, ns:].contiguous()
tsf_smpl = smpls[:, ns:].contiguous()
src_mask = masks[:, 0:ns].contiguous()
ref_mask = masks[:, ns:].contiguous()
##################################################################
# input_G_bg (bs, 1, 4, 512, 512): for background inpainting network,
# input_G_src (bs, ns, 6, 512, 512): for source identity network,
# input_G_tsf (bs, nt, 6, 512, 512): for transfer network,
# Tst (bs, ns, nt, 512, 512, 2): the transformation flows from source (s_i) to target (t_j);
# Ttt (): if temporal is True, transformation from last time target (t_{j-1)
# to current time target (t_j), otherwise, it is None.
#
# src_mask (bs, ns, 1, 512, 512): the source masks;
# tsf_mask (bs, nt, 1, 512, 512): the target masks;
# head_bbox (ns, 4): the head bounding boxes of all targets;
# body_bbox (ns, 4): the body bounding boxes of all targets;
# uv_img (bs, 3, 512, 512): the extracted uv images, for visualization.
################################################################
input_G_bg, input_G_src, input_G_tsf, Tst, Ttt, src_mask, tsf_mask, head_bbox, body_bbox, uv_img = \
flow_comp(src_img, tsf_img, src_smpl, tsf_smpl, src_mask=src_mask, ref_mask=ref_mask,
links_ids=links_ids, offsets=offsets, temporal=cfg.temporal)
visualizer.vis_named_img("input_G_bg", input_G_bg[0:1, 0, 0:3])
visualizer.vis_named_img("input_G_src", input_G_src[0, :, 0:3])
visualizer.vis_named_img("input_G_tsf", input_G_tsf[0, :, 0:3])
visualizer.vis_named_img("src_mask", src_mask[0])
visualizer.vis_named_img("tsf_mask", tsf_mask[0])
visualizer.vis_named_img("src_img", src_img[0])
visualizer.vis_named_img("tsf_img", tsf_img[0])
visualizer.vis_named_img("uv_img", uv_img[0:1])
if aug_bg is not None:
visualizer.vis_named_img("aug_bg", aug_bg)
print(f"aug_bg = {aug_bg.shape}")
print(f"input_G_bg = {input_G_bg.shape}")
print(f"input_G_src = {input_G_src.shape}")
print(f"input_G_tsf = {input_G_tsf.shape}")
print(f"Tst = {Tst.shape}")
print(f"src_mask = {src_mask.shape}")
print(f"tsf_mask = {tsf_mask.shape}")
print(f"head_bbox = {head_bbox.shape}")
print(f"body_bbox = {body_bbox.shape}")
print(f"uv_img = {uv_img.shape}")