def load_config(args):
"""
Given the arguemnts, load and initialize the configs.
Args:
args (argument): arguments includes `shard_id`, `num_shards`,
`init_method`, `cfg_file`, and `opts`.
"""
# Setup cfg.
cfg = get_cfg()
# Load config from cfg.
if args.cfg_file is not None:
cfg.merge_from_file(args.cfg_file)
# Load config from command line, overwrite config from opts.
if args.opts is not None:
cfg.merge_from_list(args.opts)
# Inherit parameters from args.
if hasattr(args, "num_shards") and hasattr(args, "shard_id"):
cfg.NUM_SHARDS = args.num_shards
cfg.SHARD_ID = args.shard_id
if hasattr(args, "rng_seed"):
cfg.RNG_SEED = args.rng_seed
if hasattr(args, "output_dir"):
cfg.OUTPUT_DIR = args.output_dir
# Create the checkpoint dir.
cu.make_checkpoint_dir(cfg.OUTPUT_DIR)
return cfg
def test_dataset():
from config.defaults import get_cfg
cfg = get_cfg()
dataset = ExampleDataset(cfg, train=True)
x, y = dataset[4]
print(x.shape)
print(y)
def main(args):
cfg = get_cfg()
if args.cfg_file:
cfg.merge_from_file(args.cfg_file)
if args.opts is not None:
cfg.merge_from_list(args.opts)
cfg.freeze()
solver = Solver(cfg)
if cfg.MODE in ["train", "training"]:
solver.train(cfg.TRAIN.NUM_EPOCHS)
elif cfg.MODE in ['validate', 'validation']:
solver.evaluate(split=cfg.VAL.SPLIT)
elif cfg.MODE in ['test', 'testing']:
solver.inference(split=cfg.TEST.SPLIT, batch_size=cfg.TEST.BATCH_SIZE)
def load_config(args):
"""
Given the arguemnts, load and initialize the configs.
Args:
args (argument): arguments includes `shard_id`, `num_shards`,
`init_method`, `cfg_file`, and `opts`.
"""
# Setup cfg.
cfg = get_cfg()
# Load config from cfg.
if args.cfg_file is not None:
cfg.merge_from_file(args.cfg_file)
# Load config from command line, overwrite config from opts.
if args.opts is not None:
cfg.merge_from_list(args.opts)
if args.test:
cfg.TRAIN.ENABLE = False
cfg.TEST.ENABLE = True
# Inherit parameters from args.
if hasattr(args, "num_shards") and hasattr(args, "shard_id"):
cfg.NUM_SHARDS = args.num_shards
cfg.SHARD_ID = args.shard_id
if hasattr(args, "rng_seed"):
cfg.RNG_SEED = args.rng_seed
if hasattr(args, "output_dir"):
cfg.OUTPUT_DIR = args.output_dir
cfg_file_name = args.cfg_file.split('/')[-1].split('.yaml')[0]
cfg.LOG_NAME = cfg_file_name + '.log'
cfg.OUTPUT_DIR = os.path.join(cfg.OUTPUT_DIR, cfg_file_name)
cfg.TEST.OUTPUT_DIR = os.path.join(cfg.TEST.OUTPUT_DIR, cfg_file_name)
# Create the checkpoint dir.
cu.make_checkpoint_dir(cfg.OUTPUT_DIR)
return cfg
fuser_input = env_agent_cat_features[:,
smpl_bgn:smpl_end].contiguous(
)
fuser_input = fuser_input.view(-1, 2, ft_sz).permute(1, 0, 2)
fuser_output = self.agents_environment_fuser(fuser_input)
fuser_output = torch.mean(fuser_output, dim=0)
context_features[:, smpl_bgn:smpl_end] = fuser_output.view(
bsz, -1, ft_sz)
return self.event_detector(context_features.permute(0, 2, 1))
if __name__ == '__main__':
cfg = get_cfg()
model = EventDetection(cfg)
batch_size = 1
temporal_dim = 100
box_dim = 4
feature_dim = 2304
env_input = torch.randn(batch_size, temporal_dim, feature_dim)
agent_input = torch.randn(batch_size, temporal_dim, box_dim, feature_dim)
agent_padding_mask = torch.tensor(
np.random.randint(0, 1, (batch_size, temporal_dim, box_dim))).bool()
a, b, c = model(env_input, agent_input, agent_padding_mask)
print(a.shape, b.shape, c.shape)