def run_test(cfg, model, distributed):
if distributed:
model = model.module
torch.cuda.empty_cache()
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference",
dataset_name)
os.makedirs(output_folder, exist_ok=True)
output_folders[idx] = output_folder
# make test dataloader.
data_loaders_test = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
# test for each dataset.
for output_folder, dataset_name, data_loader_test in zip(
output_folders, dataset_names, data_loaders_test):
inference(
model,
data_loader_test,
dataset_name,
mem_active=has_memory(cfg.MODEL.IA_STRUCTURE),
output_folder=output_folder,
)
synchronize()
def __init__(
self,
cfg_file_path,
model_weight_url,
detect_rate,
common_cate,
device,
exclude_class=[],
):
# TODO: add exclude class
cfg = base_cfg.clone()
cfg.merge_from_file(cfg_file_path)
cfg.MODEL.WEIGHT = model_weight_url
cfg.MODEL.IA_STRUCTURE.MEMORY_RATE *= detect_rate
if common_cate:
cfg.MODEL.ROI_ACTION_HEAD.NUM_CLASSES = 15
cfg.MODEL.ROI_ACTION_HEAD.NUM_PERSON_MOVEMENT_CLASSES = 6
cfg.MODEL.ROI_ACTION_HEAD.NUM_OBJECT_MANIPULATION_CLASSES = 5
cfg.MODEL.ROI_ACTION_HEAD.NUM_PERSON_INTERACTION_CLASSES = 4
cfg.freeze()
self.cfg = cfg
self.model = build_detection_model(cfg)
self.model.eval()
self.model.to(device)
self.has_memory = has_memory(cfg.MODEL.IA_STRUCTURE)
self.mem_len = cfg.MODEL.IA_STRUCTURE.LENGTH
self.mem_rate = cfg.MODEL.IA_STRUCTURE.MEMORY_RATE
self.has_object = has_object(cfg.MODEL.IA_STRUCTURE)
checkpointer = ActionCheckpointer(cfg, self.model)
self.mem_pool = MemoryPool()
self.object_pool = MemoryPool()
self.mem_timestamps = []
self.obj_timestamps = []
self.pred_pos = 0
print("Loading action model weight from {}.".format(cfg.MODEL.WEIGHT))
_ = checkpointer.load(cfg.MODEL.WEIGHT)
print("Action model weight successfully loaded.")
self.transforms, self.person_transforms, self.object_transforms = self.build_transform(
)
self.device = device
self.cpu_device = torch.device("cpu")
self.exclude_class = exclude_class
def __init__(self, dim_person, dim_mem, dim_out, structure_cfg):
super(IAStructure, self).__init__()
self.dim_person = dim_person
self.dim_others = dim_mem
self.dim_inner = structure_cfg.DIM_INNER
self.dim_out = dim_out
self.max_person = structure_cfg.MAX_PERSON
self.max_object = structure_cfg.MAX_OBJECT
self.mem_len = structure_cfg.LENGTH[0] + structure_cfg.LENGTH[1] + 1
self.mem_feature_len = self.mem_len * structure_cfg.MAX_PER_SEC
self.I_block_list = structure_cfg.I_BLOCK_LIST
bias = not structure_cfg.NO_BIAS
conv_init_std = structure_cfg.CONV_INIT_STD
self.has_P = has_person(structure_cfg)
self.has_O = has_object(structure_cfg)
self.has_M = has_memory(structure_cfg)
self.person_dim_reduce = nn.Conv3d(dim_person, self.dim_inner, 1,
bias) # reduce person query
init_layer(self.person_dim_reduce, conv_init_std, bias)
self.reduce_dropout = nn.Dropout(structure_cfg.DROPOUT)
if self.has_M:
self.mem_dim_reduce = nn.Conv3d(dim_mem, self.dim_inner, 1, bias)
init_layer(self.mem_dim_reduce, conv_init_std, bias)
if self.has_P:
self.person_key_dim_reduce = nn.Conv3d(dim_person, self.dim_inner,
1,
bias) # reduce person key
init_layer(self.person_key_dim_reduce, conv_init_std, bias)
if self.has_O:
self.object_dim_reduce = nn.Conv3d(dim_person, self.dim_inner, 1,
bias)
init_layer(self.object_dim_reduce, conv_init_std, bias)
def main():
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Inference")
parser.add_argument(
"--config-file",
default="",
metavar="FILE",
help="path to config file",
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
# Merge config file.
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
# Print experimental infos.
save_dir = ""
logger = setup_logger("alphaction", save_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(cfg)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + get_pretty_env_info())
# Build the model.
model = build_detection_model(cfg)
model.to("cuda")
# load weight.
output_dir = cfg.OUTPUT_DIR
checkpointer = ActionCheckpointer(cfg, model, save_dir=output_dir)
checkpointer.load(cfg.MODEL.WEIGHT)
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
mem_active = has_memory(cfg.MODEL.IA_STRUCTURE)
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference",
dataset_name)
os.makedirs(output_folder, exist_ok=True)
output_folders[idx] = output_folder
# Do inference.
data_loaders_test = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
for output_folder, dataset_name, data_loader_test in zip(
output_folders, dataset_names, data_loaders_test):
inference(
model,
data_loader_test,
dataset_name,
mem_active=mem_active,
output_folder=output_folder,
)
synchronize()
def train(cfg,
local_rank,
distributed,
tblogger=None,
transfer_weight=False,
adjust_lr=False,
skip_val=False,
no_head=False):
# build the model.
model = build_detection_model(cfg)
device = torch.device("cuda")
model.to(device)
# make solver.
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[local_rank],
output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
arguments["person_pool"] = MemoryPool()
output_dir = cfg.OUTPUT_DIR
# load weight.
save_to_disk = get_rank() == 0
checkpointer = ActionCheckpointer(cfg, model, optimizer, scheduler,
output_dir, save_to_disk)
extra_checkpoint_data = checkpointer.load(
cfg.MODEL.WEIGHT,
model_weight_only=transfer_weight,
adjust_scheduler=adjust_lr,
no_head=no_head)
arguments.update(extra_checkpoint_data)
# make dataloader.
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments['iteration'],
)
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
val_period = cfg.SOLVER.EVAL_PERIOD
mem_active = has_memory(cfg.MODEL.IA_STRUCTURE)
# make validation dataloader if necessary
if not skip_val:
dataset_names_val = cfg.DATASETS.TEST
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
else:
dataset_names_val = []
data_loaders_val = []
# training
do_train(
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
tblogger,
val_period,
dataset_names_val,
data_loaders_val,
distributed,
mem_active,
)
return model