def invoke_agent_env(params, alg):
"""Returns the wrapped env and string name of agent, then Use `eval(agent)` to activate it from main script
"""
if params.mode == "Atari":
env = wrap_deepmind(make_atari("{}NoFrameskip-v4".format(
params.env_name, skip_frame_k=params.skip_frame_k)),
skip_frame_k=params.skip_frame_k)
if params.debug_flg:
agent = "{}_debug".format(alg)
else:
agent = "{}".format(alg)
else:
agent = "{}".format(alg)
if params.mode == "CartPole":
env = MyWrapper(gym.make("CartPole-v0"))
elif params.mode == "CartPole-p":
env = CartPole_Pixel(gym.make("CartPole-v0"))
return agent, env
try:
os.system("rm -rf {}".format(logdirs.log_DQfD))
except:
pass
parser = argparse.ArgumentParser()
parser.add_argument("--mode", default="CartPole", help="game env type")
parser.add_argument("--num_episodes",
default=100,
type=int,
help="game env type")
args = parser.parse_args()
if args.mode == "CartPole":
env = MyWrapper(gym.make("CartPole-v0"))
elif args.mode == "Atari":
env = wrap_deepmind(make_atari("PongNoFrameskip-v4"))
params = Parameters(algo="DQfD", mode=args.mode)
params.num_episodes = args.num_episodes
replay_buffer = PrioritizedReplayBuffer(
params.memory_size, alpha=params.prioritized_replay_alpha)
Beta = AnnealingSchedule(start=params.prioritized_replay_beta_start,
end=params.prioritized_replay_beta_end,
decay_steps=params.decay_steps)
agent = DQfD(args.mode, Model, Model, env.action_space.n, params,
logdirs.model_DQN)
if params.policy_fn == "Eps":
Epsilon = AnnealingSchedule(start=params.epsilon_start,
end=params.epsilon_end,
decay_type="linear")
optimizer = tf.train.RMSPropOptimizer(anneal_lr.get_value(), 0.99, 0.0, 1e-6)
if params.loss_fn == "huber":
loss_fn = tf.losses.huber_loss
elif params.loss_fn == "mse":
loss_fn = tf.losses.mean_squared_error
else:
assert False, "Choose the loss_fn from either huber or mse"
grad_clip_fn = gradient_clip_fn(flag=params.grad_clip_flg)
summary_writer = tf.contrib.summary.create_file_writer(params.log_dir)
now = datetime.now()
if params.mode == "CartPole":
env = MyWrapper(gym.make("CartPole-v0"))
params.log_dir = "../../logs/logs/" + now.strftime(
"%Y%m%d-%H%M%S") + "-DDDP/"
params.model_dir = "../../logs/models/" + now.strftime(
"%Y%m%d-%H%M%S") + "-DDDP/"
agent = Double_DQN_cartpole(Model, optimizer, loss_fn, grad_clip_fn,
env.action_space.n, params)
elif params.mode == "CartPole-p":
env = CartPole_Pixel(gym.make("CartPole-v0"))
params.log_dir = "../../logs/logs/" + now.strftime(
"%Y%m%d-%H%M%S") + "-DDDP-p/"
params.model_dir = "../../logs/models/" + now.strftime(
"%Y%m%d-%H%M%S") + "-DDDP-p/"
agent = Double_DQN(Model_p, optimizer, loss_fn, grad_clip_fn,
env.action_space.n, params)
default="../../logs/models/DQN/",
help="directory for trained model")
parser.add_argument("--debug_flg",
default=False,
type=bool,
help="debug mode or not")
parser.add_argument("--google_colab",
default=False,
type=bool,
help="if you are executing this on GoogleColab")
params = parser.parse_args()
params.goal = 195
params.test_episodes = 10
if params.mode == "CartPole":
env = MyWrapper(gym.make("CartPole-v0"))
elif params.mode == "CartPole-p":
env = CartPole_Pixel(gym.make("CartPole-v0"))
if params.google_colab:
# mount your drive on google colab
from google.colab import drive
drive.mount("/content/gdrive")
params.log_dir = "/content/gdrive/My Drive/logs/logs/DQN/{}".format(
params.env_name)
params.model_dir = "/content/gdrive/My Drive/logs/models/DQN/{}".format(
params.env_name)
os.makedirs(params.log_dir)
os.makedirs(params.model_dir)
assert os.path.isdir(