def make_ddqn_agent(obs_space_dim, action_space_dim):
gamma = 1
obs_low = np.array([-1] * obs_space_dim)
obs_high = np.array([1] * obs_space_dim)
ac_low = np.array([-1] * action_space_dim)
ac_high = np.array([1] * action_space_dim)
obsSpace = gym.spaces.Box(obs_low, obs_high)
actSpace = gym.spaces.Box(ac_low, ac_high)
qFunc = q_functions.FCQuadraticStateQFunction(obsSpace.low.size,
actSpace.low.size,
n_hidden_channels=50,
n_hidden_layers=2,
action_space=actSpace)
optimizer = chainer.optimizers.Adam(eps=1e-2)
optimizer.setup(qFunc)
# Use AdditiveOU for exploration
ou_sigma = (actSpace.high - actSpace.low) * 0.25
explorer = explorers.AdditiveOU(sigma=ou_sigma)
# DQN uses Experience Replay.
# Specify a replay buffer and its capacity.
replay_buffer = chainerrl.replay_buffer.ReplayBuffer(capacity=10**6)
phi = lambda x: x.astype(np.float32, copy=False)
agent = chainerrl.agents.DoubleDQN(qFunc,
optimizer,
replay_buffer,
gamma,
explorer,
replay_start_size=500,
update_interval=1,
target_update_interval=100,
phi=phi)
return agent
def build_agent() -> DoubleDQN:
# observation:
# friction on each somite (#somite)
# tension on each somite except for both ends (#somite - 2)
# cos(somite phases), sin(somites phases) (#oscillator x 2)
# cos(gripper phases), sin(gripper phases) (#gripper x 2)
obs_size = config.somites + (config.somites - 2) + config.oscillators*2 + config.grippers*2
# actions: feedbacks to somite oscillators, feedbacks to gripper oscillators
action_size = config.oscillators + config.grippers
action_space = spaces.Box(low=-F_OUTPUT_BOUND, high=F_OUTPUT_BOUND, shape=(action_size))
q_func = q_functions.FCQuadraticStateQFunction(
obs_size,
action_size,
n_hidden_channels=6,
n_hidden_layers=2,
action_space=action_space)
opt = optimizers.Adam()
opt.setup(q_func)
explorer = explorers.AdditiveOU(sigma=OU_SIGMA)
rbuf = replay_buffer.ReplayBuffer(capacity=5 * 10 ** 5)
phi = lambda x: x.astype(np.float32)
agent = DoubleDQN(q_func, opt, rbuf, gamma=GAMMA, explorer=explorer,
phi=lambda x: x.astype(np.float32), gpu=GPU, replay_start_size=10000)
return agent
def dqn_q_values_and_neuronal_net(self, args, action_space, obs_size,
obs_space):
"""
learning process
"""
if isinstance(action_space, spaces.Box):
action_size = action_space.low.size
# Use NAF to apply DQN to continuous action spaces
q_func = q_functions.FCQuadraticStateQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
action_space=action_space)
# Use the Ornstein-Uhlenbeck process for exploration
ou_sigma = (action_space.high - action_space.low) * 0.2
explorer = explorers.AdditiveOU(sigma=ou_sigma)
else:
n_actions = action_space.n
# print("n_actions: ", n_actions)
q_func = q_functions.FCStateQFunctionWithDiscreteAction(
obs_size,
n_actions,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers)
# print("q_func ", q_func)
# Use epsilon-greedy for exploration
explorer = explorers.LinearDecayEpsilonGreedy(
args.start_epsilon, args.end_epsilon,
args.final_exploration_steps, action_space.sample)
# print("explorer: ", explorer)
if args.noisy_net_sigma is not None:
links.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
# print("obs_space.low : ", obs_space.shape)
chainerrl.misc.draw_computational_graph(
[q_func(np.zeros_like(obs_space.low, dtype=np.float32)[None])],
os.path.join(args.outdir, 'model'))
opt = optimizers.Adam()
opt.setup(q_func)
rbuf_capacity = 5 * 10**5
if args.minibatch_size is None:
args.minibatch_size = 32
if args.prioritized_replay:
betasteps = (args.steps - args.replay_start_size) \
// args.update_interval
rbuf = replay_buffer.PrioritizedReplayBuffer(rbuf_capacity,
betasteps=betasteps)
else:
rbuf = replay_buffer.ReplayBuffer(rbuf_capacity)
return q_func, opt, rbuf, explorer
def build_agent() -> DDPG:
# observation:
# friction on each somite (#somite)
# tension on each somite except for both ends (#somite - 2)
# cos(somite phases), sin(somites phases) (#oscillator x 2)
# cos(gripper phases), sin(gripper phases) (#gripper x 2)
obs_size = config.somites + (config.somites - 2) + config.oscillators*2 + config.grippers*2
# actions: feedbacks to somite oscillators, feedbacks to gripper oscillators
action_size = config.oscillators + config.grippers
q_func = q_functions.FCBNLateActionSAQFunction(
obs_size,
action_size,
n_hidden_channels=6,
n_hidden_layers=2,
normalize_input=True)
pi = policy.FCBNDeterministicPolicy(
obs_size,
action_size=action_size,
n_hidden_channels=6,
n_hidden_layers=2,
min_action=-F_OUTPUT_BOUND,
max_action=F_OUTPUT_BOUND,
bound_action=True,
normalize_input=True)
model = DDPGModel(q_func=q_func, policy=pi)
opt_actor = optimizers.Adam()
opt_actor.setup(model['policy'])
opt_critic = optimizers.Adam()
opt_critic.setup(model['q_function'])
opt_actor.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_actor')
opt_critic.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_critic')
rep_buf = replay_buffer.ReplayBuffer(capacity=1 * 10 ** 5)
explorer = explorers.AdditiveOU(sigma=OU_SIGMA)
phi = lambda x: x.astype(np.float32)
agent = DDPG(model, opt_actor, opt_critic, rep_buf, gamma=GAMMA, explorer=explorer,
phi=lambda x: x.astype(np.float32), gpu=GPU, replay_start_size=10000)
return agent
bound_action=True)
# The Model
model = DDPGModel(q_func=q_func, policy=pi)
opt_actor = optimizers.Adam(alpha=actor_lr)
opt_critic = optimizers.Adam(alpha=critic_lr)
opt_actor.setup(model['policy'])
opt_critic.setup(model['q_function'])
opt_actor.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_a')
opt_critic.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_c')
rbuf = replay_buffer.ReplayBuffer(replay_buffer_size)
ou_sigma = (action_space.high - action_space.low) * 0.2
explorer = explorers.AdditiveOU(sigma=ou_sigma)
# The agent
agent = DDPG(model,
opt_actor,
opt_critic,
rbuf,
gamma=gamma,
explorer=explorer,
replay_start_size=replay_start_size,
target_update_method=target_update_method,
target_update_interval=target_update_interval,
update_interval=update_interval,
soft_update_tau=soft_update_tau,
n_times_update=number_of_update_times,
phi=phi,
def main():
import logging
logging.basicConfig(level=logging.DEBUG)
parser = argparse.ArgumentParser()
parser.add_argument('--outdir',
type=str,
default='results',
help='Directory path to save output files.'
' If it does not exist, it will be created.')
parser.add_argument('--env', type=str, default='Humanoid-v2')
parser.add_argument('--seed',
type=int,
default=0,
help='Random seed [0, 2 ** 32)')
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--final-exploration-steps', type=int, default=10**6)
parser.add_argument('--actor-lr', type=float, default=1e-4)
parser.add_argument('--critic-lr', type=float, default=1e-3)
parser.add_argument('--load', type=str, default='')
parser.add_argument('--steps', type=int, default=10**7)
parser.add_argument('--n-hidden-channels', type=int, default=300)
parser.add_argument('--n-hidden-layers', type=int, default=3)
parser.add_argument('--replay-start-size', type=int, default=5000)
parser.add_argument('--n-update-times', type=int, default=1)
parser.add_argument('--target-update-interval', type=int, default=1)
parser.add_argument('--target-update-method',
type=str,
default='soft',
choices=['hard', 'soft'])
parser.add_argument('--soft-update-tau', type=float, default=1e-2)
parser.add_argument('--update-interval', type=int, default=4)
parser.add_argument('--eval-n-runs', type=int, default=100)
parser.add_argument('--eval-interval', type=int, default=10**5)
parser.add_argument('--gamma', type=float, default=0.995)
parser.add_argument('--minibatch-size', type=int, default=200)
parser.add_argument('--render', action='store_true')
parser.add_argument('--demo', action='store_true')
parser.add_argument('--use-bn', action='store_true', default=False)
parser.add_argument('--monitor', action='store_true')
parser.add_argument('--reward-scale-factor', type=float, default=1e-2)
args = parser.parse_args()
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print('Output files are saved in {}'.format(args.outdir))
# Set a random seed used in ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
def clip_action_filter(a):
return np.clip(a, action_space.low, action_space.high)
def reward_filter(r):
return r * args.reward_scale_factor
def make_env(test):
env = gym.make(args.env)
# Use different random seeds for train and test envs
env_seed = 2**32 - 1 - args.seed if test else args.seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = chainerrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = chainerrl.wrappers.Monitor(env, args.outdir)
if isinstance(env.action_space, spaces.Box):
misc.env_modifiers.make_action_filtered(env, clip_action_filter)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = chainerrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if args.render and not test:
env = chainerrl.wrappers.Render(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_size = np.asarray(env.observation_space.shape).prod()
action_space = env.action_space
action_size = np.asarray(action_space.shape).prod()
if args.use_bn:
q_func = q_functions.FCBNLateActionSAQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
normalize_input=True)
pi = policy.FCBNDeterministicPolicy(
obs_size,
action_size=action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
min_action=action_space.low,
max_action=action_space.high,
bound_action=True,
normalize_input=True)
else:
q_func = q_functions.FCSAQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers)
pi = policy.FCDeterministicPolicy(
obs_size,
action_size=action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
min_action=action_space.low,
max_action=action_space.high,
bound_action=True)
model = DDPGModel(q_func=q_func, policy=pi)
opt_a = optimizers.Adam(alpha=args.actor_lr)
opt_c = optimizers.Adam(alpha=args.critic_lr)
opt_a.setup(model['policy'])
opt_c.setup(model['q_function'])
opt_a.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_a')
opt_c.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_c')
rbuf = replay_buffer.ReplayBuffer(5 * 10**5)
def random_action():
a = action_space.sample()
if isinstance(a, np.ndarray):
a = a.astype(np.float32)
return a
ou_sigma = (action_space.high - action_space.low) * 0.2
explorer = explorers.AdditiveOU(sigma=ou_sigma)
agent = DDPG(model,
opt_a,
opt_c,
rbuf,
gamma=args.gamma,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_method=args.target_update_method,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
soft_update_tau=args.soft_update_tau,
n_times_update=args.n_update_times,
gpu=args.gpu,
minibatch_size=args.minibatch_size)
if len(args.load) > 0:
agent.load(args.load)
eval_env = make_env(test=True)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
max_episode_len=timestep_limit)
print('n_runs: {} mean: {} median: {} stdev {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_env=eval_env,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
train_max_episode_len=timestep_limit)
def main():
import logging
logging.basicConfig(level=logging.DEBUG)
parser = argparse.ArgumentParser()
parser.add_argument('--outdir',
type=str,
default='results',
help='Directory path to save output files.'
' If it does not exist, it will be created.')
parser.add_argument('--env', type=str, default='Pendulum-v0')
parser.add_argument('--seed',
type=int,
default=0,
help='Random seed [0, 2 ** 32)')
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--final-exploration-steps', type=int, default=10**4)
parser.add_argument('--start-epsilon', type=float, default=1.0)
parser.add_argument('--end-epsilon', type=float, default=0.1)
parser.add_argument('--noisy-net-sigma', type=float, default=None)
parser.add_argument('--demo', action='store_true', default=False)
parser.add_argument('--load', type=str, default=None)
parser.add_argument('--steps', type=int, default=10**5)
parser.add_argument('--prioritized-replay', action='store_true')
parser.add_argument('--replay-start-size', type=int, default=1000)
parser.add_argument('--target-update-interval', type=int, default=10**2)
parser.add_argument('--target-update-method', type=str, default='hard')
parser.add_argument('--soft-update-tau', type=float, default=1e-2)
parser.add_argument('--update-interval', type=int, default=1)
parser.add_argument('--eval-n-runs', type=int, default=100)
parser.add_argument('--eval-interval', type=int, default=10**4)
parser.add_argument('--n-hidden-channels', type=int, default=100)
parser.add_argument('--n-hidden-layers', type=int, default=2)
parser.add_argument('--gamma', type=float, default=0.99)
parser.add_argument('--minibatch-size', type=int, default=None)
parser.add_argument('--render-train', action='store_true')
parser.add_argument('--render-eval', action='store_true')
parser.add_argument('--monitor', action='store_true')
parser.add_argument('--reward-scale-factor', type=float, default=1e-3)
args = parser.parse_args()
# Set a random seed used in ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print('Output files are saved in {}'.format(args.outdir))
def clip_action_filter(a):
return np.clip(a, action_space.low, action_space.high)
def make_env(test):
env = gym.make(args.env)
# Use different random seeds for train and test envs
env_seed = 2**32 - 1 - args.seed if test else args.seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = chainerrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = chainerrl.wrappers.Monitor(env, args.outdir)
if isinstance(env.action_space, spaces.Box):
misc.env_modifiers.make_action_filtered(env, clip_action_filter)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = chainerrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if ((args.render_eval and test) or (args.render_train and not test)):
env = chainerrl.wrappers.Render(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_space = env.observation_space
obs_size = obs_space.low.size
action_space = env.action_space
if isinstance(action_space, spaces.Box):
action_size = action_space.low.size
# Use NAF to apply DQN to continuous action spaces
q_func = q_functions.FCQuadraticStateQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
action_space=action_space)
# Use the Ornstein-Uhlenbeck process for exploration
ou_sigma = (action_space.high - action_space.low) * 0.2
explorer = explorers.AdditiveOU(sigma=ou_sigma)
else:
n_actions = action_space.n
q_func = q_functions.FCStateQFunctionWithDiscreteAction(
obs_size,
n_actions,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers)
# Use epsilon-greedy for exploration
explorer = explorers.LinearDecayEpsilonGreedy(
args.start_epsilon, args.end_epsilon, args.final_exploration_steps,
action_space.sample)
if args.noisy_net_sigma is not None:
links.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
# Draw the computational graph and save it in the output directory.
chainerrl.misc.draw_computational_graph(
[q_func(np.zeros_like(obs_space.low, dtype=np.float32)[None])],
os.path.join(args.outdir, 'model'))
opt = optimizers.Adam()
opt.setup(q_func)
rbuf_capacity = 5 * 10**5
if args.minibatch_size is None:
args.minibatch_size = 32
if args.prioritized_replay:
betasteps = (args.steps - args.replay_start_size) \
// args.update_interval
rbuf = replay_buffer.PrioritizedReplayBuffer(rbuf_capacity,
betasteps=betasteps)
else:
rbuf = replay_buffer.ReplayBuffer(rbuf_capacity)
agent = DQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=args.gamma,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
minibatch_size=args.minibatch_size,
target_update_method=args.target_update_method,
soft_update_tau=args.soft_update_tau,
)
if args.load:
agent.load(args.load)
eval_env = make_env(test=True)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
max_episode_len=timestep_limit)
print('n_runs: {} mean: {} median: {} stdev {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
eval_env=eval_env,
train_max_episode_len=timestep_limit)
def main():
import logging
logging.basicConfig(level=logging.DEBUG)
parser = argparse.ArgumentParser()
parser.add_argument('--outdir', type=str, default='dqn_out')
parser.add_argument('--env', type=str, default='Pendulum-v0')
parser.add_argument('--seed', type=int, default=None)
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--final-exploration-steps', type=int, default=10**4)
parser.add_argument('--start-epsilon', type=float, default=1.0)
parser.add_argument('--end-epsilon', type=float, default=0.1)
parser.add_argument('--demo', action='store_true', default=False)
parser.add_argument('--load', type=str, default=None)
parser.add_argument('--steps', type=int, default=10**5)
parser.add_argument('--prioritized-replay', action='store_true')
parser.add_argument('--episodic-replay', action='store_true')
parser.add_argument('--replay-start-size', type=int, default=1000)
parser.add_argument('--target-update-interval', type=int, default=10**2)
parser.add_argument('--target-update-method', type=str, default='hard')
parser.add_argument('--soft-update-tau', type=float, default=1e-2)
parser.add_argument('--update-interval', type=int, default=1)
parser.add_argument('--eval-n-runs', type=int, default=100)
parser.add_argument('--eval-interval', type=int, default=10**4)
parser.add_argument('--n-hidden-channels', type=int, default=100)
parser.add_argument('--n-hidden-layers', type=int, default=2)
parser.add_argument('--gamma', type=float, default=0.99)
parser.add_argument('--minibatch-size', type=int, default=None)
parser.add_argument('--render-train', action='store_true')
parser.add_argument('--render-eval', action='store_true')
parser.add_argument('--monitor', action='store_true')
parser.add_argument('--reward-scale-factor', type=float, default=1e-3)
args = parser.parse_args()
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print('Output files are saved in {}'.format(args.outdir))
if args.seed is not None:
misc.set_random_seed(args.seed)
def clip_action_filter(a):
return np.clip(a, action_space.low, action_space.high)
def make_env(for_eval):
env = gym.make(args.env)
if args.monitor:
env = gym.wrappers.Monitor(env, args.outdir)
if isinstance(env.action_space, spaces.Box):
misc.env_modifiers.make_action_filtered(env, clip_action_filter)
if not for_eval:
misc.env_modifiers.make_reward_filtered(
env, lambda x: x * args.reward_scale_factor)
if ((args.render_eval and for_eval)
or (args.render_train and not for_eval)):
misc.env_modifiers.make_rendered(env)
return env
env = make_env(for_eval=False)
timestep_limit = env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_space = env.observation_space
obs_size = obs_space.low.size
action_space = env.action_space
if isinstance(action_space, spaces.Box):
action_size = action_space.low.size
# Use NAF to apply DQN to continuous action spaces
q_func = q_functions.FCQuadraticStateQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
action_space=action_space)
# Use the Ornstein-Uhlenbeck process for exploration
ou_sigma = (action_space.high - action_space.low) * 0.2
explorer = explorers.AdditiveOU(sigma=ou_sigma)
else:
n_actions = action_space.n
q_func = q_functions.FCStateQFunctionWithDiscreteAction(
obs_size,
n_actions,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers)
# Use epsilon-greedy for exploration
explorer = explorers.LinearDecayEpsilonGreedy(
args.start_epsilon, args.end_epsilon, args.final_exploration_steps,
action_space.sample)
# Draw the computational graph and save it in the output directory.
chainerrl.misc.draw_computational_graph(
[q_func(np.zeros_like(obs_space.low, dtype=np.float32)[None])],
os.path.join(args.outdir, 'model'))
opt = optimizers.Adam()
opt.setup(q_func)
rbuf_capacity = 5 * 10**5
if args.episodic_replay:
if args.minibatch_size is None:
args.minibatch_size = 4
if args.prioritized_replay:
betasteps = (args.steps - args.replay_start_size) \
// args.update_interval
rbuf = replay_buffer.PrioritizedEpisodicReplayBuffer(
rbuf_capacity, betasteps=betasteps)
else:
rbuf = replay_buffer.EpisodicReplayBuffer(rbuf_capacity)
else:
if args.minibatch_size is None:
args.minibatch_size = 32
if args.prioritized_replay:
betasteps = (args.steps - args.replay_start_size) \
// args.update_interval
rbuf = replay_buffer.PrioritizedReplayBuffer(rbuf_capacity,
betasteps=betasteps)
else:
rbuf = replay_buffer.ReplayBuffer(rbuf_capacity)
def phi(obs):
return obs.astype(np.float32)
agent = DQN(q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=args.gamma,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
phi=phi,
minibatch_size=args.minibatch_size,
target_update_method=args.target_update_method,
soft_update_tau=args.soft_update_tau,
episodic_update=args.episodic_replay,
episodic_update_len=16)
if args.load:
agent.load(args.load)
eval_env = make_env(for_eval=True)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_runs=args.eval_n_runs,
max_episode_len=timestep_limit)
print('n_runs: {} mean: {} median: {} stdev {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_runs=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
eval_env=eval_env,
max_episode_len=timestep_limit)
def main():
import logging
logging.basicConfig(level=logging.DEBUG)
parser = argparse.ArgumentParser()
parser.add_argument('--outdir', type=str, default='out')
parser.add_argument('--env', type=str, default='Humanoid-v1')
parser.add_argument('--seed', type=int, default=None)
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--final-exploration-steps', type=int, default=10**6)
parser.add_argument('--actor-lr', type=float, default=1e-4)
parser.add_argument('--critic-lr', type=float, default=1e-3)
parser.add_argument('--load', type=str, default='')
parser.add_argument('--steps', type=int, default=10**7)
parser.add_argument('--n-hidden-channels', type=int, default=300)
parser.add_argument('--n-hidden-layers', type=int, default=3)
parser.add_argument('--replay-start-size', type=int, default=5000)
parser.add_argument('--n-update-times', type=int, default=1)
parser.add_argument('--target-update-frequency', type=int, default=1)
parser.add_argument('--target-update-method',
type=str,
default='soft',
choices=['hard', 'soft'])
parser.add_argument('--soft-update-tau', type=float, default=1e-2)
parser.add_argument('--update-frequency', type=int, default=4)
parser.add_argument('--eval-n-runs', type=int, default=100)
parser.add_argument('--eval-frequency', type=int, default=10**5)
parser.add_argument('--gamma', type=float, default=0.995)
parser.add_argument('--minibatch-size', type=int, default=200)
parser.add_argument('--render', action='store_true')
parser.add_argument('--demo', action='store_true')
parser.add_argument('--use-bn', action='store_true', default=False)
parser.add_argument('--monitor', action='store_true')
parser.add_argument('--reward-scale-factor', type=float, default=1e-2)
args = parser.parse_args()
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print('Output files are saved in {}'.format(args.outdir))
if args.seed is not None:
misc.set_random_seed(args.seed)
def clip_action_filter(a):
return np.clip(a, action_space.low, action_space.high)
def reward_filter(r):
return r * args.reward_scale_factor
def make_env():
env = gym.make(args.env)
if args.monitor:
env = gym.wrappers.Monitor(env, args.outdir)
if isinstance(env.action_space, spaces.Box):
misc.env_modifiers.make_action_filtered(env, clip_action_filter)
misc.env_modifiers.make_reward_filtered(env, reward_filter)
if args.render:
misc.env_modifiers.make_rendered(env)
def __exit__(self, *args):
pass
env.__exit__ = __exit__
return env
env = make_env()
timestep_limit = env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_size = np.asarray(env.observation_space.shape).prod()
action_space = env.action_space
action_size = np.asarray(action_space.shape).prod()
if args.use_bn:
q_func = q_functions.FCBNLateActionSAQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
normalize_input=True)
pi = policy.FCBNDeterministicPolicy(
obs_size,
action_size=action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
min_action=action_space.low,
max_action=action_space.high,
bound_action=True,
normalize_input=True)
else:
q_func = q_functions.FCSAQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers)
pi = policy.FCDeterministicPolicy(
obs_size,
action_size=action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
min_action=action_space.low,
max_action=action_space.high,
bound_action=True)
model = DDPGModel(q_func=q_func, policy=pi)
opt_a = optimizers.Adam(alpha=args.actor_lr)
opt_c = optimizers.Adam(alpha=args.critic_lr)
opt_a.setup(model['policy'])
opt_c.setup(model['q_function'])
opt_a.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_a')
opt_c.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_c')
rbuf = replay_buffer.ReplayBuffer(5 * 10**5)
def phi(obs):
return obs.astype(np.float32)
def random_action():
a = action_space.sample()
if isinstance(a, np.ndarray):
a = a.astype(np.float32)
return a
ou_sigma = (action_space.high - action_space.low) * 0.2
explorer = explorers.AdditiveOU(sigma=ou_sigma)
agent = DDPG(model,
opt_a,
opt_c,
rbuf,
gamma=args.gamma,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_method=args.target_update_method,
target_update_frequency=args.target_update_frequency,
update_frequency=args.update_frequency,
soft_update_tau=args.soft_update_tau,
n_times_update=args.n_update_times,
phi=phi,
gpu=args.gpu,
minibatch_size=args.minibatch_size)
agent.logger.setLevel(logging.DEBUG)
if len(args.load) > 0:
agent.load(args.load)
if args.demo:
mean, median, stdev = experiments.eval_performance(
env=env,
agent=agent,
n_runs=args.eval_n_runs,
max_episode_len=timestep_limit)
print('n_runs: {} mean: {} median: {} stdev'.format(
args.eval_n_runs, mean, median, stdev))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_runs=args.eval_n_runs,
eval_frequency=args.eval_frequency,
outdir=args.outdir,
max_episode_len=timestep_limit)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--outdir',
type=str,
default='/tmp/chainerRL_results',
help='Directory path to save output files.'
' If it does not exist, it will be created.')
parser.add_argument('--seed',
type=int,
default=0,
help='Random seed [0, 2 ** 32)')
parser.add_argument('--final-exploration-steps', type=int, default=10**4)
parser.add_argument('--start-epsilon', type=float, default=1.0)
parser.add_argument('--end-epsilon', type=float, default=0.1)
parser.add_argument('--noisy-net-sigma', type=float, default=None)
parser.add_argument('--evaluate',
action='store_true',
default=False,
help="Run evaluation mode")
parser.add_argument('--load',
type=str,
default=None,
help="Load saved_model")
parser.add_argument('--steps', type=int, default=4 * 10**6)
parser.add_argument('--prioritized-replay', action='store_true')
parser.add_argument('--replay-start-size', type=int, default=1000)
parser.add_argument('--target-update-interval',
type=int,
default=5 * 10**2)
parser.add_argument('--target-update-method', type=str, default='hard')
parser.add_argument('--soft-update-tau', type=float, default=1e-2)
parser.add_argument('--update-interval', type=int, default=1)
parser.add_argument('--eval-n-runs', type=int, default=1)
parser.add_argument('--eval-interval',
type=int,
default=1e4,
help="After how many steps to evaluate the agent."
"(-1 -> always)")
parser.add_argument('--n-hidden-channels', type=int, default=20)
parser.add_argument('--n-hidden-layers', type=int, default=20)
parser.add_argument('--gamma', type=float, default=0.99)
parser.add_argument('--minibatch-size', type=int, default=None)
parser.add_argument('--render-train', action='store_true')
parser.add_argument('--render-eval', action='store_true')
parser.add_argument('--reward-scale-factor', type=float, default=1)
parser.add_argument('--time-step-limit', type=int, default=1e5)
parser.add_argument('--outdir-time-suffix',
choices=['empty', 'none', 'time'],
default='empty',
type=str.lower)
parser.add_argument('--checkpoint_frequency',
type=int,
default=1e3,
help="Nuber of steps to checkpoint after")
parser.add_argument('--verbose',
'-v',
action='store_true',
help='Use debug log-level')
args = parser.parse_args()
import logging
logging.basicConfig(
level=logging.INFO if not args.verbose else logging.DEBUG)
# Set a random seed used in ChainerRL ALSO SETS NUMPY SEED!
misc.set_random_seed(args.seed)
if args.outdir and not args.load:
outdir_suffix_dict = {
'none': '',
'empty': '',
'time': '%Y%m%dT%H%M%S.%f'
}
args.outdir = experiments.prepare_output_dir(
args,
args.outdir,
argv=sys.argv,
time_format=outdir_suffix_dict[args.outdir_time_suffix])
elif args.load:
if args.load.endswith(os.path.sep):
args.load = args.load[:-1]
args.outdir = os.path.dirname(args.load)
count = 0
fn = os.path.join(args.outdir.format(count), 'scores_{:>03d}')
while os.path.exists(fn.format(count)):
count += 1
os.rename(os.path.join(args.outdir, 'scores.txt'), fn.format(count))
if os.path.exists(os.path.join(args.outdir, 'best')):
os.rename(os.path.join(args.outdir, 'best'),
os.path.join(args.outdir, 'best_{:>03d}'.format(count)))
logging.info('Output files are saved in {}'.format(args.outdir))
def clip_action_filter(a):
return np.clip(a, action_space.low, action_space.high)
def make_env(test):
HOST = '' # The server's hostname or IP address
PORT = 54321 # The port used by the server
if test: # Just such that eval and train env don't have the same port
PORT += 1
# TODO don't hardcode env params
# TODO if we use this solution (i.e. write port to file and read it with FD) we would have to make sure that
# outdir doesn't append time strings. Otherwise it will get hard to use on the cluster
env = FDEnvSelHeur(host=HOST,
port=PORT,
num_heuristics=2,
config_dir=args.outdir)
# Use different random seeds for train and test envs
env_seed = 2**32 - 1 - args.seed if test else args.seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = chainerrl.wrappers.CastObservationToFloat32(env)
if isinstance(env.action_space, spaces.Box):
misc.env_modifiers.make_action_filtered(env, clip_action_filter)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = chainerrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if ((args.render_eval and test) or (args.render_train and not test)):
env = chainerrl.wrappers.Render(env)
return env
env = make_env(test=False)
# state = env.reset()
# while True:
# for x in [1,1,1,1,0,0,0,0]:
# state, reward, done, _ = env.step(x)
# print(x)
# if done:
# break
timestep_limit = args.time_step_limit
obs_space = env.observation_space
obs_size = obs_space.low.size
action_space = env.action_space
if isinstance(action_space, spaces.Box): # Usefull if we want to control
action_size = action_space.low.size # other continous parameters
# Use NAF to apply DQN to continuous action spaces
q_func = q_functions.FCQuadraticStateQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
action_space=action_space)
# Use the Ornstein-Uhlenbeck process for exploration
ou_sigma = (action_space.high - action_space.low) * 0.2
explorer = explorers.AdditiveOU(sigma=ou_sigma)
else:
n_actions = action_space.n
q_func = q_functions.FCStateQFunctionWithDiscreteAction(
obs_size,
n_actions,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers)
# q_func = FCDuelingDQN(
# obs_size, n_actions)
# Use epsilon-greedy for exploration
explorer = explorers.LinearDecayEpsilonGreedy(
args.start_epsilon, args.end_epsilon, args.final_exploration_steps,
action_space.sample)
if args.noisy_net_sigma is not None:
links.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
# Draw the computational graph and save it in the output directory.
if not args.load:
chainerrl.misc.draw_computational_graph(
[q_func(np.zeros_like(obs_space.low, dtype=np.float32)[None])],
os.path.join(args.outdir, 'model'))
opt = optimizers.Adam(eps=1e-2)
logging.info('Optimizer: %s', str(opt))
opt.setup(q_func)
opt.add_hook(GradientClipping(5))
rbuf_capacity = 5 * 10**5
if args.minibatch_size is None:
args.minibatch_size = 32
# args.minibatch_size = 16
if args.prioritized_replay:
betasteps = (args.steps - args.replay_start_size) \
// args.update_interval
rbuf = replay_buffer.PrioritizedReplayBuffer(rbuf_capacity,
betasteps=betasteps)
else:
rbuf = replay_buffer.ReplayBuffer(rbuf_capacity)
agent = DDQN(
q_func,
opt,
rbuf,
gamma=args.gamma,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
minibatch_size=args.minibatch_size,
target_update_method=args.target_update_method,
soft_update_tau=args.soft_update_tau,
)
t_offset = 0
if args.load: # Continue training model or load for evaluation
agent.load(args.load)
rbuf.load(os.path.join(args.load, 'replay_buffer.pkl'))
try:
t_offset = int(os.path.basename(args.load).split('_')[0])
except TypeError:
with open(os.path.join(args.load, 't.txt'), 'r') as fh:
data = fh.readlines()
t_offset = int(data[0])
except ValueError:
t_offset = 0
eval_env = make_env(test=False)
if args.evaluate:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
max_episode_len=timestep_limit)
print('n_runs: {} mean: {} median: {} stdev {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
criterion = 'steps' # can be made an argument if we support any other form of checkpointing
l = logging.getLogger('Checkpoint_Hook')
def checkpoint(env, agent, step):
if criterion == 'steps':
if step % args.checkpoint_frequency == 0:
save_agent_and_replay_buffer(
agent,
step,
args.outdir,
suffix='_chkpt',
logger=l,
chckptfrq=args.checkpoint_frequency)
else:
# TODO seems to checkpoint given wall_time we would have to modify the environment such that it tracks
# time or number of episodes
raise NotImplementedError
hooks = [checkpoint]
experiments.train_agent(agent=agent,
env=env,
steps=args.steps,
outdir=args.outdir,
step_hooks=hooks,
step_offset=t_offset)
def main(args):
import logging
logging.basicConfig(level=logging.INFO, filename='log')
if(type(args) is list):
args=make_args(args)
# Set a random seed used in ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu,))
if not os.path.exists(args.outdir):
os.makedirs(args.outdir)
print('Output files are saved in {}'.format(args.outdir))
def clip_action_filter(a):
return np.clip(a, action_space.low, action_space.high)
def make_env(test):
env = gym.make(args.env)
# Use different random seeds for train and test envs
env_seed = 2 ** 32 - 1 - args.seed if test else args.seed
env.seed(env_seed)
env = chainerrl.wrappers.CastObservationToFloat32(env)
if isinstance(env.action_space, spaces.Box):
misc.env_modifiers.make_action_filtered(env, clip_action_filter)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = chainerrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if ((args.render_eval and test) or
(args.render_train and not test)):
env = chainerrl.wrappers.Render(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_space = env.observation_space
obs_size = obs_space.low.size
action_space = env.action_space
if isinstance(action_space, spaces.Box):
print("Use NAF to apply DQN to continuous action spaces")
action_size = action_space.low.size
# Use NAF to apply DQN to continuous action spaces
q_func = q_functions.FCQuadraticStateQFunction(
obs_size, action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
action_space=action_space)
# Use the Ornstein-Uhlenbeck process for exploration
ou_sigma = (action_space.high - action_space.low) * 0.2
explorer = explorers.AdditiveOU(sigma=ou_sigma)
else:
print("not continuous action spaces")
n_actions = action_space.n
q_func = q_functions.FCStateQFunctionWithDiscreteAction(
obs_size, n_actions,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers)
# Use epsilon-greedy for exploration
explorer = explorers.LinearDecayEpsilonGreedy(
args.start_epsilon, args.end_epsilon, args.final_exploration_steps,
action_space.sample)
if args.noisy_net_sigma is not None:
links.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
# Draw the computational graph and save it in the output directory.
chainerrl.misc.draw_computational_graph(
[q_func(np.zeros_like(obs_space.low, dtype=np.float32)[None])],
os.path.join(args.outdir, 'model'))
opt = optimizers.Adam()
opt.setup(q_func)
rbuf_capacity = 5 * 10 ** 5
if args.minibatch_size is None:
args.minibatch_size = 32
if args.prioritized_replay:
betasteps = (args.steps - args.replay_start_size) \
// args.update_interval
rbuf = replay_buffer.PrioritizedReplayBuffer(
rbuf_capacity, betasteps=betasteps)
else:
rbuf = replay_buffer.ReplayBuffer(rbuf_capacity)
agent = DoubleDQN(q_func, opt, rbuf, gpu=args.gpu, gamma=args.gamma,
explorer=explorer, replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
minibatch_size=args.minibatch_size,
target_update_method=args.target_update_method,
soft_update_tau=args.soft_update_tau,
)
if args.load_agent:
agent.load(args.load_agent)
eval_env = make_env(test=True)
if (args.mode=='train'):
experiments.train_agent_with_evaluation(
agent=agent, env=env, steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs, eval_interval=args.eval_interval,
outdir=args.outdir, eval_env=eval_env,
step_offset=args.step_offset,
checkpoint_freq=args.checkpoint_freq,
train_max_episode_len=timestep_limit,
log_type=args.log_type
)
elif (args.mode=='check'):
from matplotlib import animation
import matplotlib.pyplot as plt
frames = []
for i in range(3):
obs = env.reset()
done = False
R = 0
t = 0
while not done and t < 200:
frames.append(env.render(mode = 'rgb_array'))
action = agent.act(obs)
obs, r, done, _ = env.step(action)
R += r
t += 1
print('test episode:', i, 'R:', R)
agent.stop_episode()
env.close()
from IPython.display import HTML
plt.figure(figsize=(frames[0].shape[1]/72.0, frames[0].shape[0]/72.0),dpi=72)
patch = plt.imshow(frames[0])
plt.axis('off')
def animate(i):
patch.set_data(frames[i])
anim = animation.FuncAnimation(plt.gcf(), animate, frames=len(frames),interval=50)
anim.save(args.save_mp4)
return anim
def chokoDQN(env, args=None):
args = args or []
if (type(args) is list):
args = make_args(args)
obs_space = env.observation_space
obs_size = obs_space.low.size * args.stack_k
action_space = env.action_space
if isinstance(action_space, spaces.Box):
action_size = action_space.low.size
q_func = q_functions.FCQuadraticStateQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
action_space=action_space)
# Use the Ornstein-Uhlenbeck process for exploration
ou_sigma = (action_space.high - action_space.low) * 0.2
explorer = explorers.AdditiveOU(sigma=ou_sigma)
else:
n_actions = action_space.n
q_func = q_functions.FCStateQFunctionWithDiscreteAction(
obs_size,
n_actions,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers)
# Use epsilon-greedy for exploration
explorer = explorers.LinearDecayEpsilonGreedy(
args.start_epsilon, args.end_epsilon, args.final_exploration_steps,
action_space.sample)
if args.noisy_net_sigma is not None:
links.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
opt = optimizers.Adam()
opt.setup(q_func)
rbuf_capacity = 5 * 10**5
if args.minibatch_size is None:
args.minibatch_size = 32
if args.prioritized_replay:
betasteps = (args.steps - args.replay_start_size) \
// args.update_interval
rbuf = replay_buffer.PrioritizedReplayBuffer(rbuf_capacity,
betasteps=betasteps)
else:
rbuf = replay_buffer.ReplayBuffer(rbuf_capacity)
agent = DQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=args.gamma,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
minibatch_size=args.minibatch_size,
target_update_method=args.target_update_method,
soft_update_tau=args.soft_update_tau,
)
return agent
def main():
import logging
logging.basicConfig(level=logging.DEBUG)
parser = argparse.ArgumentParser()
parser.add_argument('--outdir', type=str, default='out')
parser.add_argument('--env', type=str, default='griewank')
parser.add_argument('--seed',
type=int,
default=0,
help='Random seed [0, 2 ** 32)')
parser.add_argument('--obs-dim', type=int, default=1, help='int')
parser.add_argument('--action-dim', type=int, default=1, help='int')
parser.add_argument('--action-low',
type=int,
default=-100,
help='int. action state low limit.')
parser.add_argument('--action-high',
type=int,
default=100,
help='int. action state high limit.')
parser.add_argument('--gpu', type=int, default=-1)
# parser.add_argument('--final-exploration-steps',
# type=int, default=10 ** 4)
parser.add_argument('--actor-lr', type=float, default=1e-4)
parser.add_argument('--critic-lr', type=float, default=1e-3)
parser.add_argument('--load', type=str, default='')
parser.add_argument('--steps', type=int, default=10**4)
parser.add_argument('--n-hidden-channels', type=int, default=300)
parser.add_argument('--n-hidden-layers', type=int, default=3)
parser.add_argument('--replay-start-size', type=int, default=5000)
parser.add_argument('--n-update-times', type=int, default=1)
parser.add_argument('--target-update-interval', type=int, default=1)
parser.add_argument('--target-update-method',
type=str,
default='soft',
choices=['hard', 'soft'])
parser.add_argument('--soft-update-tau', type=float, default=1e-2)
parser.add_argument('--update-interval', type=int, default=4)
parser.add_argument('--eval-n-runs', type=int, default=100)
parser.add_argument('--eval-interval', type=int, default=10**3)
parser.add_argument('--gamma', type=float, default=0.995)
parser.add_argument('--minibatch-size', type=int, default=200)
parser.add_argument('--render', action='store_true')
parser.add_argument('--demo', action='store_true')
parser.add_argument('--use-bn', action='store_true', default=False)
parser.add_argument('--monitor', action='store_true')
parser.add_argument('--reward-scale-factor', type=float, default=1e-2)
parser.add_argument('--sigma-scale-factor', type=float, default=5e-2)
args = parser.parse_args()
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print('Output files are saved in {}'.format(args.outdir))
# Set a random seed used in ChainerRL
# misc.set_random_seed(args.seed, gpus=(args.gpu,))
misc.set_random_seed(args.seed)
def clip_action_filter(a):
print("clip_", a)
return np.clip(a, args.action_low, args.action_high)
def reward_filter(r):
return r * args.reward_scale_factor
def make_env(test,
env_name='griewank',
obs_dim=1,
action_dim=1,
action_low=-100,
action_high=100):
env = None
if env_name == 'griewank':
env = Griewank(test=test,
obs_dim=obs_dim,
action_dim=action_dim,
action_low=args.action_low,
action_high=args.action_high)
elif env_name == 'easy2d':
env = Easy2D(test=test,
obs_dim=obs_dim,
action_dim=action_dim,
action_low=args.action_low,
action_high=args.action_high)
# Use different random seeds for train and test envs
env_seed = 2**32 - 1 - args.seed if test else args.seed
env.seed(env_seed)
# if args.monitor:
# env = gym.wrappers.Monitor(env, args.outdir)
# if isinstance(env.action_space, spaces.Box):
# misc.env_modifiers.make_action_filtered(env, clip_action_filter)
if not test:
misc.env_modifiers.make_action_filtered(env, clip_action_filter)
misc.env_modifiers.make_reward_filtered(env, reward_filter)
# if args.render and not test:
# misc.env_modifiers.make_rendered(env)
return env
env = make_env(test=False,
env_name=args.env,
obs_dim=args.obs_dim,
action_dim=args.action_dim,
action_low=args.action_low,
action_high=args.action_high)
# timestep_limit = env.spec.tags.get(
# 'wrapper_config.TimeLimit.max_episode_steps')
timestep_limit = env.timestep_limit
obs_size = np.asarray(env.observation_space.shape).prod()
action_space = env.action_space
action_size = np.asarray(action_space.shape).prod()
print("obs_size, action_size", obs_size, action_size)
if args.use_bn:
q_func = q_functions.FCBNLateActionSAQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
normalize_input=True)
pi = policy.FCBNDeterministicPolicy(
obs_size,
action_size=action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
min_action=action_space.low,
max_action=action_space.high,
bound_action=True,
normalize_input=True)
else:
q_func = q_functions.FCSAQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers)
pi = policy.FCDeterministicPolicy(
obs_size,
action_size=action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
min_action=action_space.low,
max_action=action_space.high,
bound_action=True)
model = DDPGModel(q_func=q_func, policy=pi)
opt_a = optimizers.Adam(alpha=args.actor_lr)
opt_c = optimizers.Adam(alpha=args.critic_lr)
opt_a.setup(model['policy'])
opt_c.setup(model['q_function'])
opt_a.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_a')
opt_c.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_c')
rbuf = replay_buffer.ReplayBuffer(5 * 10**5)
def phi(obs):
return obs.astype(np.float32)
def random_action():
a = action_space.sample()
if isinstance(a, np.ndarray):
a = a.astype(np.float32)
return a
ou_sigma = (action_space.high - action_space.low) * args.sigma_scale_factor
explorer = explorers.AdditiveOU(sigma=ou_sigma)
agent = DDPG(model,
opt_a,
opt_c,
rbuf,
gamma=args.gamma,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_method=args.target_update_method,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
soft_update_tau=args.soft_update_tau,
n_times_update=args.n_update_times,
phi=phi,
gpu=args.gpu,
minibatch_size=args.minibatch_size)
if len(args.load) > 0:
agent.load(args.load)
eval_env = make_env(test=True,
env_name=args.env,
obs_dim=args.obs_dim,
action_dim=args.action_dim,
action_low=args.action_low,
action_high=args.action_high)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_runs=args.eval_n_runs,
max_episode_len=timestep_limit)
print('n_runs: {} mean: {} median: {} stdev {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_env=eval_env,
eval_n_runs=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
max_episode_len=timestep_limit)
def make_agent_ddpg(args, env):
obs_size = np.asarray(env.observation_space.shape).prod()
action_space = env.action_space
action_size = np.asarray(action_space.shape).prod()
q_func = FCSAQFunction(
obs_size, action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers)
pi = FCDeterministicPolicy(
obs_size, action_size=action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
min_action=action_space.low, max_action=action_space.high,
bound_action=True)
if args.gpu > -1:
q_func.to_gpu(args.gpu)
pi.to_gpu(args.gpu)
else:
q_func.to_cpu()
pi.to_cpu()
model = DDPGModel(q_func=q_func, policy=pi)
opt_a = optimizers.Adam(alpha=args.actor_lr)
opt_c = optimizers.Adam(alpha=args.critic_lr)
opt_a.setup(model['policy'])
opt_c.setup(model['q_function'])
opt_a.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_a')
opt_c.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_c')
rbuf = replay_buffer.ReplayBuffer(5 * 10 ** 5)
def phi(obs):
return obs.astype(np.float32)
# def random_action():
# a = action_space.sample()
# if isinstance(a, np.ndarray):
# a = a.astype(np.float32)
# return a
ou_sigma = (action_space.high - action_space.low) * 0.2
explorer = explorers.AdditiveOU(sigma=ou_sigma)
if args.skip_step == 0:
agent = DDPG(model, opt_a, opt_c, rbuf, gamma=args.gamma,
explorer=explorer, replay_start_size=args.replay_start_size,
target_update_method=args.target_update_method,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
soft_update_tau=args.soft_update_tau,
n_times_update=args.n_update_times,
phi=phi, gpu=args.gpu, minibatch_size=args.minibatch_size)
else:
agent = DDPGStep(model, opt_a, opt_c, rbuf, gamma=args.gamma,
explorer=explorer, replay_start_size=args.replay_start_size,
target_update_method=args.target_update_method,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
soft_update_tau=args.soft_update_tau,
n_times_update=args.n_update_times,
phi=phi, gpu=args.gpu, minibatch_size=args.minibatch_size, skip_step=args.skip_step)
if args.model_dir is not None:
agent.save(args.model_dir)
return agent
def main(args):
import logging
logging.basicConfig(level=logging.INFO, filename='log')
if(type(args) is list):
args=make_args(args)
if not os.path.exists(args.outdir):
os.makedirs(args.outdir)
# Set a random seed used in ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu,))
print('Output files are saved in {}'.format(args.outdir))
def clip_action_filter(a):
return np.clip(a, action_space.low, action_space.high)
def make_env(test):
env = gym.make(args.env)
# Use different random seeds for train and test envs
env_seed = 2 ** 32 - 1 - args.seed if test else args.seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = chainerrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = chainerrl.wrappers.Monitor(env, args.outdir)
if isinstance(env.action_space, spaces.Box):
misc.env_modifiers.make_action_filtered(env, clip_action_filter)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = chainerrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if ((args.render_eval and test) or
(args.render_train and not test)):
env = chainerrl.wrappers.Render(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_space = env.observation_space
obs_size = obs_space.low.size
action_space = env.action_space
if isinstance(action_space, spaces.Box):
print("Use NAF to apply DQN to continuous action spaces")
action_size = action_space.low.size
# Use NAF to apply DQN to continuous action spaces
q_func = q_functions.FCQuadraticStateQFunction(
obs_size, action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
action_space=action_space)
# Use the Ornstein-Uhlenbeck process for exploration
ou_sigma = (action_space.high - action_space.low) * 0.2
explorer = explorers.AdditiveOU(sigma=ou_sigma)
else:
print("not continuous action spaces")
n_actions = action_space.n
q_func = q_functions.FCStateQFunctionWithDiscreteAction(
obs_size, n_actions,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers)
# Use epsilon-greedy for exploration
explorer = explorers.LinearDecayEpsilonGreedy(
args.start_epsilon, args.end_epsilon, args.final_exploration_steps,
action_space.sample)
if args.noisy_net_sigma is not None:
links.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
# Draw the computational graph and save it in the output directory.
chainerrl.misc.draw_computational_graph(
[q_func(np.zeros_like(obs_space.low, dtype=np.float32)[None])],
os.path.join(args.outdir, 'model'))
opt = optimizers.Adam()
opt.setup(q_func)
rbuf_capacity = 5 * 10 ** 5
if args.minibatch_size is None:
args.minibatch_size = 32
if args.prioritized_replay:
betasteps = (args.steps - args.replay_start_size) \
// args.update_interval
rbuf = replay_buffer.PrioritizedReplayBuffer(
rbuf_capacity, betasteps=betasteps)
else:
rbuf = replay_buffer.ReplayBuffer(rbuf_capacity)
agent = DQN(q_func, opt, rbuf, gpu=args.gpu, gamma=args.gamma,
explorer=explorer, replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
minibatch_size=args.minibatch_size,
target_update_method=args.target_update_method,
soft_update_tau=args.soft_update_tau,
)
if args.load_agent:
agent.load(args.load_agent)
eval_env = make_env(test=True)
if (args.mode=='train'):
experiments.train_agent_with_evaluation(
agent=agent, env=env, steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs, eval_interval=args.eval_interval,
outdir=args.outdir, eval_env=eval_env,
step_offset=args.step_offset,
checkpoint_freq=args.checkpoint_freq,
train_max_episode_len=args.max_episode_len,
log_type=args.log_type
)
elif (args.mode=='check'):
return tools.make_video.check(env=env,agent=agent,save_mp4=args.save_mp4)
elif (args.mode=='growth'):
return tools.make_video.growth(env=env,agent=agent,outdir=args.outdir,max_num=args.max_episode_len,save_mp4=args.save_mp4)
