if args.checkpoint is None and args.checkpoints is None:
sys.exit(1)
# check that server is running
while not is_server_running(server):
print('Start the server: python3 env_server.py')
sleep(sleep_interval)
# extract waypoints
trajectory_file = osp.join(trajectory_dir, args.trajectory)
_, _, waypoints = prepare_trajectories(signal_dir, [trajectory_file],
use_inputs=use_inputs,
use_outputs=use_outputs,
action_scale=action_scale,
lookback=lookback)
# create environment
env_fns = [
make_env(MantsinenBasic, model_path, model_dir, signal_dir, server,
data, nsteps, lookback, use_inputs, use_outputs, action_scale,
tstep) for data in waypoints
]
env = MeveaVecEnv(env_fns)
# checkpoints
if args.checkpoints is not None:
# check that server is running
while not is_server_running(server):
print('Start the server: python3 env_server.py')
sleep(sleep_interval)
# prepare training data
trajectory_files = [
osp.join(trajectory_dir, fpath) for fpath in os.listdir(trajectory_dir)
if fpath.endswith('csv')
]
bc_train, bc_val, waypoints = prepare_trajectories(
signal_dir,
trajectory_files,
use_inputs=use_inputs,
use_outputs=use_outputs,
action_scale=action_scale,
lookback=lookback,
tstep=tstep)
print(bc_train.shape, bc_val.shape)
# create environments
env_fns = [
make_env(MantsinenBasic, model_path, model_dir, signal_dir, server,
waypoints, nsteps, lookback, use_inputs, use_outputs,
action_scale, tstep)
]
env = MeveaVecEnv(env_fns)
try:
use_signals = False
# process arguments
parser = argparse.ArgumentParser()
parser.add_argument('-m', '--model', help='File path to the Mevea model.', default='C:\\Users\\mevea\\MeveaModels\\Mantsinen\\Models\\Mantsinen300M\\300M_fixed.mvs')
parser.add_argument('-s', '--server', help='Server URL.', default='http://127.0.0.1:5000')
parser.add_argument('-t', '--trajectories', help='Trajectory files.', default='trajectory1.csv,trajectory2.csv,trajectory3.csv,trajectory4.csv')
parser.add_argument('-f', '--frequency', help='Action frequency.', default=0.1, type=float)
parser.add_argument('-o', '--output', help='Result output.', default='models/mevea/mantsinen/ppo')
args = parser.parse_args()
# prepare training data
trajectory_files = [osp.join(trajectory_dir, fpath) for fpath in args.trajectories.split(',')]
trajectory_data = prepare_trajectories(signal_dir, trajectory_files, args.frequency, use_signals=use_signals)
# create environments
env_fns = [make_env(MantsinenBasic, args.model, signal_dir, data, args.server, args.frequency, use_signals) for data in trajectory_data]
env = MeveaVecEnv(env_fns)
# create model
model = ppo(MlpPolicy, env, runner=MeveaRunner, verbose=1)
# pretrain model
model.pretrain(trajectory_data, n_epochs=20000)
model.save('{0}/model_checkpoints/rl_model_0_steps.zip'.format(args.output))