def run_dqn_thread(baseDir, runNo, thread_id, args):
init_nn_library(True, '0')
runner, agent = run_dqn(**args)
runner.run()
with open(
baseDir + '/' + str(args['learning_rate']) + '-' + str(runNo) +
'.pkl', 'w') as f:
pickle.dump(agent.stats, f)
help='gym-env monitor directory')
parser.add_argument('--gpu',
type=str,
default=None,
help='Gpu to use (none for cpu, any for any gpu)')
args = parser.parse_args()
from algo.dqn import run_dqn, run_dqn_test
arguments = vars(args)
import numpy as np
if args.mode == "train":
runner, _ = run_dqn(**arguments)
if args.save_trajectory is not None:
from utils.trajectory_utils import TrajectorySaver
ts = TrajectorySaver(args.save_trajectory)
runner.listen(ts, None)
runner.run()
else:
stats = run_dqn_test(**arguments)
#from PIL import Image
#
#from envs.gym_env import gym_env
#from envs.env_transform import *
#from utils.preprocess import *
#from utils.network_utils import NetworkSaver
#from runner.runner import Runner
import tensorflow as tf from utils.network_utils import NetworkSaver from runner.runner import TrajRunner from utils.trajectory_utils import TrajectoryReplay from nets.net import init_nn_library arguments = vars(args) env = get_env(args.game, args.atari, args.env_transforms) if args.load_trajectory is None: dqn_args = arguments.copy() dqn_args['mode'] = 'test' dqn_args['replay_buffer_size'] = 0 runner = run_dqn(**dqn_args) replay_buffer = ReplayBuffer(args.replay_buffer_size, 1, args.update_frequency, args.replay_start_size, args.batch_size) else: init_nn_library(True, "1") runner = TrajRunner(args.max_step) replay_buffer = TrajectoryReplay(args.load_trajectory, args.batch_size) envOps = EnvOps(env.observation_space.shape, env.action_space.n, args.learning_rate) summary_writer = tf.summary.FileWriter(args.logdir, K.get_session().graph) if not args.logdir is None else None #model = EnvModelCartpole(envOps) model = globals()[args.env_model](envOps) env = EnvLearner(replay_buffer, model, summary_writer, args.reward_scale) runner.listen(env, None)
def run_dqn_thread(baseDir, runNo, args, max_episode, test_only):
init_nn_library(True, args['gpu'])
if not test_only:
runner, _ = run_dqn(**args)
runner.run()
run_test('dqn', args, baseDir, runNo, max_episode)