def train_agent(args, driving_style):
from agents.dagger.agent import ensure_mnet2_baseline_weights
if args.agent == c.DAGGER or args.agent == c.DAGGER_MNET2:
train_dagger(args)
elif args.agent == c.BOOTSTRAPPED_PPO2:
from agents.bootstrap_rl.train import train
net_path = args.net_path
if not net_path:
log.info('Bootstrapping from baseline agent')
net_path = ensure_mnet2_baseline_weights(args.net_path)
if not args.sync and not args.eval_only:
args.sync = True
log.warning('Detected training RL in async mode which '
'can cause unequal time deltas. '
'Switching to synchronous mode. '
'Use --sync to avoid this.')
sim_args = get_sim_args_from_command_args(args)
train.run(args.env_id, resume_dir=args.resume_train,
bootstrap_net_path=net_path, agent_name=args.agent,
render=args.render, camera_rigs=[c.DEFAULT_CAM],
is_sync=args.sync, driving_style=driving_style,
is_remote_client=args.remote, eval_only=args.eval_only,
sim_args=sim_args)
else:
raise Exception('Agent type not recognized')
def train_dagger(args):
"""
Run the first iteration of DAgger where our policy is random.
"""
from agents.dagger.train import train
train.run(resume_dir=args.resume_train, data_dir=args.recording_dir,
agent_name=args.agent,
overfit=args.overfit, eval_only=args.eval_only,
tf_debug=args.tf_debug,
freeze_pretrained=args.freeze_pretrained,
train_args_collection_name=args.train_args_collection)