def _test_training(self,
gpu,
steps=5000,
load_model=False,
require_success=True):
random_seed.set_random_seed(1)
logging.basicConfig(level=logging.DEBUG)
env = self.make_env_and_successful_return(test=False)[0]
test_env, successful_return = self.make_env_and_successful_return(
test=True)
agent = self.make_agent(env, gpu)
if load_model:
print("Load agent from", self.agent_dirname)
agent.load(self.agent_dirname)
agent.replay_buffer.load(self.rbuf_filename)
# Train
train_agent_with_evaluation(
agent=agent,
env=env,
steps=steps,
outdir=self.tmpdir,
eval_interval=200,
eval_n_steps=None,
eval_n_episodes=5,
successful_score=1,
eval_env=test_env,
)
# Test
n_test_runs = 5
eval_returns = run_evaluation_episodes(
test_env,
agent,
n_steps=None,
n_episodes=n_test_runs,
)
n_succeeded = np.sum(np.asarray(eval_returns) >= successful_return)
if require_success:
assert n_succeeded == n_test_runs
# Save
agent.save(self.agent_dirname)
agent.replay_buffer.save(self.rbuf_filename)
def _test_abc(self,
steps=100000,
require_success=True,
gpu=-1,
load_model=False):
env, _ = self.make_env_and_successful_return(test=False)
test_env, successful_return = self.make_env_and_successful_return(
test=True)
agent = self.make_agent(env, gpu)
if load_model:
print("Load agent from", self.agent_dirname)
agent.load(self.agent_dirname)
max_episode_len = None if self.episodic else 2
# Train
train_agent_with_evaluation(
agent=agent,
env=env,
eval_env=test_env,
steps=steps,
outdir=self.tmpdir,
eval_interval=200,
eval_n_steps=None,
eval_n_episodes=5,
successful_score=successful_return,
train_max_episode_len=max_episode_len,
)
# Test
n_test_runs = 5
eval_returns, _ = run_evaluation_episodes(
test_env,
agent,
n_steps=None,
n_episodes=n_test_runs,
max_episode_len=max_episode_len,
)
if require_success:
n_succeeded = np.sum(np.asarray(eval_returns) >= successful_return)
assert n_succeeded == n_test_runs
# Save
agent.save(self.agent_dirname)
replay_start_size=5 * 10**4,
# replay_start_size=5*10**3,
target_update_interval=10**4,
clip_delta=True,
update_interval=4,
batch_accumulator="sum",
phi=phi,
gpu=0,
episodic_update_len=4,
recurrent=True)
if ns.demo:
agent.load(ns.model)
test_env = pfrl.wrappers.Render(test_env)
test_env = sleep_wrapper.SleepWrapper(test_env, ns.sleep)
experiments.eval_performance(env=test_env,
agent=agent,
n_steps=100000,
n_episodes=None)
else:
experiments.train_agent_with_evaluation(agent=agent,
env=env,
steps=ns.steps,
eval_n_steps=None,
checkpoint_freq=None,
eval_n_episodes=10,
eval_interval=100000,
outdir=ns.model,
eval_env=test_env,
use_tensorboard=True)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--env", type=str, default="BreakoutNoFrameskip-v4")
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("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
parser.add_argument("--gpu", type=int, default=0)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load", type=str, default=None)
parser.add_argument("--final-exploration-frames", type=int, default=10**6)
parser.add_argument("--final-epsilon", type=float, default=0.1)
parser.add_argument("--eval-epsilon", type=float, default=0.05)
parser.add_argument("--steps", type=int, default=10**7)
parser.add_argument(
"--max-frames",
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help="Maximum number of frames for each episode.",
)
parser.add_argument("--replay-start-size", type=int, default=5 * 10**4)
parser.add_argument("--target-update-interval", type=int, default=10**4)
parser.add_argument("--eval-interval", type=int, default=10**5)
parser.add_argument("--update-interval", type=int, default=4)
parser.add_argument("--eval-n-runs", type=int, default=10)
parser.add_argument("--batch-size", type=int, default=32)
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test,
)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, args.eval_epsilon)
if args.monitor:
env = pfrl.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training")
if args.render:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
n_actions = env.action_space.n
n_atoms = 51
v_max = 10
v_min = -10
q_func = torch.nn.Sequential(
pfrl.nn.LargeAtariCNN(),
pfrl.q_functions.DistributionalFCStateQFunctionWithDiscreteAction(
512,
n_actions,
n_atoms,
v_min,
v_max,
n_hidden_channels=0,
n_hidden_layers=0,
),
)
# Use the same hyper parameters as https://arxiv.org/abs/1707.06887
opt = torch.optim.Adam(q_func.parameters(),
2.5e-4,
eps=1e-2 / args.batch_size)
rbuf = replay_buffers.ReplayBuffer(10**6)
explorer = explorers.LinearDecayEpsilonGreedy(
1.0,
args.final_epsilon,
args.final_exploration_frames,
lambda: np.random.randint(n_actions),
)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
agent = pfrl.agents.CategoricalDQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
batch_accumulator="mean",
phi=phi,
)
if args.load:
agent.load(args.load)
if args.demo:
eval_stats = experiments.eval_performance(env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs)
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,
save_best_so_far_agent=False,
eval_env=eval_env,
)
def main():
import logging
logging.basicConfig(level=logging.INFO)
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="CartPole-v1")
parser.add_argument("--seed", type=int, default=0)
parser.add_argument("--gpu", type=int, default=0)
parser.add_argument("--final-exploration-steps", type=int, default=1000)
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**8)
parser.add_argument("--prioritized-replay", action="store_true")
parser.add_argument("--replay-start-size", type=int, default=50)
parser.add_argument("--target-update-interval", type=int, default=100)
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=1000)
parser.add_argument("--n-hidden-channels", type=int, default=12)
parser.add_argument("--n-hidden-layers", type=int, default=3)
parser.add_argument("--gamma", type=float, default=0.95)
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=1.0)
args = parser.parse_args()
# Set a random seed used in PFRL
utils.set_random_seed(args.seed)
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
env = gym.make(args.env)
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 = pfrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = pfrl.wrappers.Monitor(env, args.outdir)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = pfrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if (args.render_eval and test) or (args.render_train and not test):
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.max_episode_steps
obs_size = env.observation_space.low.size
action_space = env.action_space
n_atoms = 51
v_max = 500
v_min = 0
n_actions = action_space.n
q_func = q_functions.DistributionalFCStateQFunctionWithDiscreteAction(
obs_size,
n_actions,
n_atoms,
v_min,
v_max,
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,
)
opt = torch.optim.Adam(q_func.parameters(), 1e-3)
rbuf_capacity = 50000 # 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_buffers.PrioritizedReplayBuffer(rbuf_capacity,
betasteps=betasteps)
else:
rbuf = replay_buffers.ReplayBuffer(rbuf_capacity)
agent = pfrl.agents.CategoricalDQN(
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 _objective_core(
# optuna parameters
trial,
# training parameters
env_id,
outdir,
seed,
monitor,
gpu,
steps,
train_max_episode_len,
eval_n_episodes,
eval_interval,
batch_size,
# hyperparameters
hyperparams,
):
# Set a random seed used in PFRL
utils.set_random_seed(seed)
# Set different random seeds for train and test envs.
train_seed = seed
test_seed = 2**31 - 1 - seed
def make_env(test=False):
env = gym.make(env_id)
if not isinstance(env.observation_space, gym.spaces.Box):
raise ValueError(
"Supported only Box observation environments, but given: {}".format(
env.observation_space
)
)
if len(env.observation_space.shape) != 1:
raise ValueError(
"Supported only observation spaces with ndim==1, but given: {}".format(
env.observation_space.shape
)
)
if not isinstance(env.action_space, gym.spaces.Discrete):
raise ValueError(
"Supported only discrete action environments, but given: {}".format(
env.action_space
)
)
env_seed = test_seed if test else train_seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = pfrl.wrappers.CastObservationToFloat32(env)
if monitor:
env = pfrl.wrappers.Monitor(env, outdir)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = pfrl.wrappers.ScaleReward(env, hyperparams["reward_scale_factor"])
return env
env = make_env(test=False)
obs_space = env.observation_space
obs_size = obs_space.low.size
action_space = env.action_space
n_actions = action_space.n
# create model & q_function
model = MLP(
in_size=obs_size, out_size=n_actions, hidden_sizes=hyperparams["hidden_sizes"]
)
q_func = q_functions.SingleModelStateQFunctionWithDiscreteAction(model=model)
# Use epsilon-greedy for exploration
start_epsilon = 1
explorer = explorers.LinearDecayEpsilonGreedy(
start_epsilon=start_epsilon,
end_epsilon=hyperparams["end_epsilon"],
decay_steps=hyperparams["decay_steps"],
random_action_func=action_space.sample,
)
opt = optim.Adam(
q_func.parameters(), lr=hyperparams["lr"], eps=hyperparams["adam_eps"]
)
rbuf_capacity = steps
rbuf = replay_buffers.ReplayBuffer(rbuf_capacity)
agent = DQN(
q_func,
opt,
rbuf,
gpu=gpu,
gamma=hyperparams["gamma"],
explorer=explorer,
replay_start_size=hyperparams["replay_start_size"],
target_update_interval=hyperparams["target_update_interval"],
update_interval=hyperparams["update_interval"],
minibatch_size=batch_size,
)
eval_env = make_env(test=True)
evaluation_hooks = [OptunaPrunerHook(trial=trial)]
_, eval_stats_history = experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=steps,
eval_n_steps=None,
eval_n_episodes=eval_n_episodes,
eval_interval=eval_interval,
outdir=outdir,
eval_env=eval_env,
train_max_episode_len=train_max_episode_len,
evaluation_hooks=evaluation_hooks,
)
score = _get_score_from_eval_stats_history(eval_stats_history)
return score
def main():
import logging
logging.basicConfig(level=logging.INFO)
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)
parser.add_argument(
"--actor-learner",
action="store_true",
help="Enable asynchronous sampling with asynchronous actor(s)",
) # NOQA
parser.add_argument(
"--num-envs",
type=int,
default=1,
help=("The number of environments for sampling (only effective with"
" --actor-learner enabled)"),
) # NOQA
args = parser.parse_args()
# Set a random seed used in PFRL
utils.set_random_seed(args.seed)
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print("Output files are saved in {}".format(args.outdir))
# Set different random seeds for different subprocesses.
# If seed=0 and processes=4, subprocess seeds are [0, 1, 2, 3].
# If seed=1 and processes=4, subprocess seeds are [4, 5, 6, 7].
process_seeds = np.arange(args.num_envs) + args.seed * args.num_envs
assert process_seeds.max() < 2**32
def clip_action_filter(a):
return np.clip(a, action_space.low, action_space.high)
def make_env(idx=0, test=False):
env = gym.make(args.env)
# Use different random seeds for train and test envs
process_seed = int(process_seeds[idx])
env_seed = 2**32 - 1 - process_seed if test else process_seed
utils.set_random_seed(env_seed)
# Cast observations to float32 because our model uses float32
env = pfrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = pfrl.wrappers.Monitor(env, args.outdir)
if isinstance(env.action_space, spaces.Box):
utils.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 = pfrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if (args.render_eval and test) or (args.render_train and not test):
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.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:
pnn.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
opt = optim.Adam(q_func.parameters())
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_buffers.PrioritizedReplayBuffer(rbuf_capacity,
betasteps=betasteps)
else:
rbuf = replay_buffers.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"],
))
elif not args.actor_learner:
print(
"WARNING: Since https://github.com/pfnet/pfrl/pull/112 we have started"
" setting `eval_during_episode=True` in this script, which affects the"
" timings of evaluation phases.")
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,
eval_during_episode=True,
)
else:
# using impala mode when given num of envs
# When we use multiple envs, it is critical to ensure each env
# can occupy a CPU core to get the best performance.
# Therefore, we need to prevent potential CPU over-provision caused by
# multi-threading in Openmp and Numpy.
# Disable the multi-threading on Openmp and Numpy.
os.environ["OMP_NUM_THREADS"] = "1" # NOQA
(
make_actor,
learner,
poller,
exception_event,
) = agent.setup_actor_learner_training(args.num_envs)
poller.start()
learner.start()
experiments.train_agent_async(
processes=args.num_envs,
make_agent=make_actor,
make_env=make_env,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
stop_event=learner.stop_event,
exception_event=exception_event,
)
poller.stop()
learner.stop()
poller.join()
learner.join()
def main():
import logging
parser = argparse.ArgumentParser()
parser.add_argument("--gpu",
type=int,
default=0,
help="GPU to use, set to -1 if no GPU.")
parser.add_argument("--num-envs",
type=int,
default=1,
help="Number of envs run in parallel.")
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 32)")
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(
"--steps",
type=int,
default=2 * 10**6,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--eval-interval",
type=int,
default=100000,
help="Interval in timesteps between evaluations.",
)
parser.add_argument(
"--eval-n-runs",
type=int,
default=100,
help="Number of episodes run for each evaluation.",
)
parser.add_argument("--render",
action="store_true",
help="Render env states in a GUI window.")
parser.add_argument("--demo",
action="store_true",
help="Just run evaluation, not training.")
parser.add_argument("--load-pretrained",
action="store_true",
default=False)
parser.add_argument("--load",
type=str,
default="",
help="Directory to load agent from.")
parser.add_argument("--log-level",
type=int,
default=logging.INFO,
help="Level of the root logger.")
parser.add_argument("--monitor",
action="store_true",
help="Wrap env with gym.wrappers.Monitor.")
parser.add_argument(
"--log-interval",
type=int,
default=1000,
help=
"Interval in timesteps between outputting log messages during training",
)
parser.add_argument(
"--update-interval",
type=int,
default=2048,
help="Interval in timesteps between model updates.",
)
parser.add_argument(
"--epochs",
type=int,
default=10,
help="Number of epochs to update model for per PPO iteration.",
)
parser.add_argument("--batch-size",
type=int,
default=64,
help="Minibatch size")
current_dir = os.path.dirname(os.path.realpath(__file__))
parser.add_argument('--cfg',
type=str,
default=os.path.abspath(
os.path.join(current_dir, "raisimlib",
"raisimGymTorch", "raisimGymTorch",
"env", "envs", "rsg_anymal",
"cfg.yaml")),
help='configuration file')
cfg_abs_path = parser.parse_args().cfg
cfg = YAML().load(open(cfg_abs_path, 'r'))
args = parser.parse_args()
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL
utils.set_random_seed(args.seed)
# Set different random seeds for different subprocesses.
# If seed=0 and processes=4, subprocess seeds are [0, 1, 2, 3].
# If seed=1 and processes=4, subprocess seeds are [4, 5, 6, 7].
process_seeds = np.arange(args.num_envs) + args.seed * args.num_envs
assert process_seeds.max() < 2**31
args.outdir = experiments.prepare_output_dir(args, args.outdir)
def make_env(process_idx, test):
visualize = True if process_idx == 0 and args.render else False
env = myEnv(os.path.join(current_dir, "raisimlib", "rsc"), cfg,
visualize)
# Use different random seeds for train and test envs
process_seed = int(process_seeds[process_idx])
env_seed = 2**31 - 1 - process_seed if test else process_seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = pfrl.wrappers.CastObservationToFloat32(env)
env = NormalizeObsSpace(env)
env = pfrl.wrappers.NormalizeActionSpace(env)
if args.monitor:
env = pfrl.wrappers.Monitor(env, args.outdir)
return env
def make_batch_env(test):
return pfrl.envs.MultiprocessVectorEnv([
functools.partial(make_env, idx, test)
for idx, env in enumerate(range(args.num_envs))
])
# Only for getting timesteps, and obs-action spaces
sample_env = myEnv(os.path.join(current_dir, "raisimlib", "rsc"), cfg,
False)
timestep_limit = 100000 #sample_env.spec.max_episode_steps
obs_space = sample_env.observation_space
action_space = sample_env.action_space
print("Observation space:", obs_space)
print("Action space:", action_space)
assert isinstance(action_space, gym.spaces.Box)
# Normalize observations based on their empirical mean and variance
obs_normalizer = pfrl.nn.EmpiricalNormalization(obs_space.low.size,
clip_threshold=5)
obs_size = obs_space.low.size
action_size = action_space.low.size
policy = torch.nn.Sequential(
nn.Linear(obs_size, 64),
nn.Tanh(),
nn.Linear(64, 64),
nn.Tanh(),
nn.Linear(64, action_size),
pfrl.policies.GaussianHeadWithStateIndependentCovariance(
action_size=action_size,
var_type="diagonal",
var_func=lambda x: torch.exp(2 * x), # Parameterize log std
var_param_init=0, # log std = 0 => std = 1
),
)
vf = torch.nn.Sequential(
nn.Linear(obs_size, 64),
nn.Tanh(),
nn.Linear(64, 64),
nn.Tanh(),
nn.Linear(64, 1),
)
# While the original paper initialized weights by normal distribution,
# we use orthogonal initialization as the latest openai/baselines does.
def ortho_init(layer, gain):
nn.init.orthogonal_(layer.weight, gain=gain)
nn.init.zeros_(layer.bias)
ortho_init(policy[0], gain=1)
ortho_init(policy[2], gain=1)
ortho_init(policy[4], gain=1e-2)
ortho_init(vf[0], gain=1)
ortho_init(vf[2], gain=1)
ortho_init(vf[4], gain=1)
# Combine a policy and a value function into a single model
model = pfrl.nn.Branched(policy, vf)
opt = torch.optim.Adam(model.parameters(), lr=3e-4, eps=1e-5)
agent = PPO(
model,
opt,
obs_normalizer=obs_normalizer,
gpu=args.gpu,
update_interval=args.update_interval,
minibatch_size=args.batch_size,
epochs=args.epochs,
clip_eps_vf=None,
entropy_coef=0,
standardize_advantages=True,
gamma=0.998,
lambd=0.95,
)
if args.load or args.load_pretrained:
if args.load_pretrained:
raise Exception("Pretrained models are currently unsupported.")
# either load or load_pretrained must be false
assert not args.load or not args.load_pretrained
if args.load:
agent.load(args.load)
else:
agent.load(
utils.download_model("PPO", args.env, model_type="final")[0])
if args.demo:
env = make_env(0, True)
eval_stats = experiments.eval_performance(
env=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"],
))
elif args.num_envs == 1:
experiments.train_agent_with_evaluation(
agent=agent,
env=make_env(0, False),
eval_env=make_env(0, True),
outdir=args.outdir,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
save_best_so_far_agent=True,
use_tensorboard=True)
else:
experiments.train_agent_batch_with_evaluation(
agent=agent,
env=make_batch_env(False),
eval_env=make_batch_env(True),
outdir=args.outdir,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
log_interval=args.log_interval,
max_episode_len=timestep_limit,
save_best_so_far_agent=True,
)
def main():
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="Hopper-v2",
help="OpenAI Gym MuJoCo env to perform algorithm on.",
)
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 32)")
parser.add_argument("--gpu",
type=int,
default=0,
help="GPU to use, set to -1 if no GPU.")
parser.add_argument("--load",
type=str,
default="",
help="Directory to load agent from.")
parser.add_argument(
"--steps",
type=int,
default=10**6,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--eval-n-runs",
type=int,
default=10,
help="Number of episodes run for each evaluation.",
)
parser.add_argument(
"--eval-interval",
type=int,
default=5000,
help="Interval in timesteps between evaluations.",
)
parser.add_argument(
"--replay-start-size",
type=int,
default=10000,
help="Minimum replay buffer size before " +
"performing gradient updates.",
)
parser.add_argument("--batch-size",
type=int,
default=100,
help="Minibatch size")
parser.add_argument("--render",
action="store_true",
help="Render env states in a GUI window.")
parser.add_argument("--demo",
action="store_true",
help="Just run evaluation, not training.")
parser.add_argument("--load-pretrained",
action="store_true",
default=False)
parser.add_argument("--pretrained-type",
type=str,
default="best",
choices=["best", "final"])
parser.add_argument("--monitor",
action="store_true",
help="Wrap env with gym.wrappers.Monitor.")
parser.add_argument("--log-level",
type=int,
default=logging.INFO,
help="Level of the root logger.")
args = parser.parse_args()
logging.basicConfig(level=args.log_level)
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 PFRL
utils.set_random_seed(args.seed)
def make_env(test):
env = gym.make(args.env)
# Unwrap TimeLimit wrapper
assert isinstance(env, gym.wrappers.TimeLimit)
env = env.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 = pfrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = pfrl.wrappers.Monitor(env, args.outdir)
if args.render and not test:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.max_episode_steps
obs_space = env.observation_space
action_space = env.action_space
print("Observation space:", obs_space)
print("Action space:", action_space)
obs_size = obs_space.low.size
action_size = action_space.low.size
policy = nn.Sequential(
nn.Linear(obs_size, 400),
nn.ReLU(),
nn.Linear(400, 300),
nn.ReLU(),
nn.Linear(300, action_size),
nn.Tanh(),
pfrl.policies.DeterministicHead(),
)
policy_optimizer = torch.optim.Adam(policy.parameters())
def make_q_func_with_optimizer():
q_func = nn.Sequential(
pfrl.nn.ConcatObsAndAction(),
nn.Linear(obs_size + action_size, 400),
nn.ReLU(),
nn.Linear(400, 300),
nn.ReLU(),
nn.Linear(300, 1),
)
q_func_optimizer = torch.optim.Adam(q_func.parameters())
return q_func, q_func_optimizer
q_func1, q_func1_optimizer = make_q_func_with_optimizer()
q_func2, q_func2_optimizer = make_q_func_with_optimizer()
rbuf = replay_buffers.ReplayBuffer(10**6)
explorer = explorers.AdditiveGaussian(scale=0.1,
low=action_space.low,
high=action_space.high)
def burnin_action_func():
"""Select random actions until model is updated one or more times."""
return np.random.uniform(action_space.low,
action_space.high).astype(np.float32)
# Hyperparameters in http://arxiv.org/abs/1802.09477
agent = pfrl.agents.TD3(
policy,
q_func1,
q_func2,
policy_optimizer,
q_func1_optimizer,
q_func2_optimizer,
rbuf,
gamma=0.99,
soft_update_tau=5e-3,
explorer=explorer,
replay_start_size=args.replay_start_size,
gpu=args.gpu,
minibatch_size=args.batch_size,
burnin_action_func=burnin_action_func,
)
if len(args.load) > 0 or args.load_pretrained:
# either load or load_pretrained must be false
assert not len(args.load) > 0 or not args.load_pretrained
if len(args.load) > 0:
agent.load(args.load)
else:
agent.load(
utils.download_model("TD3",
args.env,
model_type=args.pretrained_type)[0])
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"],
))
import json
import os
with open(os.path.join(args.outdir, "demo_scores.json"), "w") as f:
json.dump(eval_stats, f)
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
parser = argparse.ArgumentParser()
parser.add_argument("--env", type=str, default="CartPole-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(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument("--beta", type=float, default=1e-4)
parser.add_argument("--batchsize", type=int, default=10)
parser.add_argument("--steps", type=int, default=10**5)
parser.add_argument("--eval-interval", type=int, default=10**4)
parser.add_argument("--eval-n-runs", type=int, default=100)
parser.add_argument("--reward-scale-factor", type=float, default=1e-2)
parser.add_argument("--render", action="store_true", default=False)
parser.add_argument("--lr", type=float, default=1e-3)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load", type=str, default="")
parser.add_argument("--log-level", type=int, default=logging.INFO)
parser.add_argument("--monitor", action="store_true")
args = parser.parse_args()
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
args.outdir = experiments.prepare_output_dir(args, args.outdir)
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 = pfrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = pfrl.wrappers.Monitor(env, args.outdir)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = pfrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if args.render and not test:
env = pfrl.wrappers.Render(env)
return env
train_env = make_env(test=False)
timestep_limit = train_env.spec.max_episode_steps
obs_space = train_env.observation_space
action_space = train_env.action_space
obs_size = obs_space.low.size
hidden_size = 200
# Switch policy types accordingly to action space types
if isinstance(action_space, gym.spaces.Box):
model = nn.Sequential(
nn.Linear(obs_size, hidden_size),
nn.LeakyReLU(0.2),
nn.Linear(hidden_size, hidden_size),
nn.LeakyReLU(0.2),
nn.Linear(hidden_size, action_space.low.size),
GaussianHeadWithFixedCovariance(0.3),
)
else:
model = nn.Sequential(
nn.Linear(obs_size, hidden_size),
nn.LeakyReLU(0.2),
nn.Linear(hidden_size, hidden_size),
nn.LeakyReLU(0.2),
nn.Linear(hidden_size, action_space.n),
SoftmaxCategoricalHead(),
)
opt = torch.optim.Adam(model.parameters(), lr=args.lr)
agent = pfrl.agents.REINFORCE(
model,
opt,
gpu=args.gpu,
beta=args.beta,
batchsize=args.batchsize,
max_grad_norm=1.0,
)
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=train_env,
eval_env=eval_env,
outdir=args.outdir,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
train_max_episode_len=timestep_limit,
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--env",
type=str,
default="BreakoutNoFrameskip-v4",
help="OpenAI Atari domain to perform algorithm on.",
)
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("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
parser.add_argument("--gpu",
type=int,
default=0,
help="GPU to use, set to -1 if no GPU.")
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load-pretrained",
action="store_true",
default=False)
parser.add_argument("--pretrained-type",
type=str,
default="best",
choices=["best", "final"])
parser.add_argument("--load", type=str, default=None)
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument(
"--steps",
type=int,
default=5 * 10**7,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--replay-start-size",
type=int,
default=5 * 10**4,
help="Minimum replay buffer size before " +
"performing gradient updates.",
)
parser.add_argument("--eval-n-steps", type=int, default=125000)
parser.add_argument("--eval-interval", type=int, default=250000)
parser.add_argument("--n-best-episodes", type=int, default=30)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=None),
episode_life=not test,
clip_rewards=not test,
)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, 0.05)
if args.monitor:
env = pfrl.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training")
if args.render:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
n_actions = env.action_space.n
q_func = nn.Sequential(
pnn.LargeAtariCNN(),
init_chainer_default(nn.Linear(512, n_actions)),
DiscreteActionValueHead(),
)
# Use the same hyperparameters as the Nature paper
opt = pfrl.optimizers.RMSpropEpsInsideSqrt(
q_func.parameters(),
lr=2.5e-4,
alpha=0.95,
momentum=0.0,
eps=1e-2,
centered=True,
)
rbuf = replay_buffers.ReplayBuffer(10**6)
explorer = explorers.LinearDecayEpsilonGreedy(
start_epsilon=1.0,
end_epsilon=0.1,
decay_steps=10**6,
random_action_func=lambda: np.random.randint(n_actions),
)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
Agent = agents.DQN
agent = Agent(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=10**4,
clip_delta=True,
update_interval=4,
batch_accumulator="sum",
phi=phi,
)
if args.load or args.load_pretrained:
# either load or load_pretrained must be false
assert not args.load or not args.load_pretrained
if args.load:
agent.load(args.load)
else:
agent.load(
utils.download_model("DQN",
args.env,
model_type=args.pretrained_type)[0])
if args.demo:
eval_stats = experiments.eval_performance(env=eval_env,
agent=agent,
n_steps=args.eval_n_steps,
n_episodes=None)
print("n_episodes: {} mean: {} median: {} stdev {}".format(
eval_stats["episodes"],
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=args.eval_n_steps,
eval_n_episodes=None,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=True,
eval_env=eval_env,
)
dir_of_best_network = os.path.join(args.outdir, "best")
agent.load(dir_of_best_network)
# run 30 evaluation episodes, each capped at 5 mins of play
stats = experiments.evaluator.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.n_best_episodes,
max_episode_len=4500,
logger=None,
)
with open(os.path.join(args.outdir, "bestscores.json"), "w") as f:
json.dump(stats, f)
print("The results of the best scoring network:")
for stat in stats:
print(str(stat) + ":" + str(stats[stat]))
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--env", type=str, default="BreakoutNoFrameskip-v4")
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("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
parser.add_argument("--gpu", type=int, default=0)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load-pretrained",
action="store_true",
default=False)
parser.add_argument("--pretrained-type",
type=str,
default="best",
choices=["best", "final"])
parser.add_argument("--load", type=str, default=None)
parser.add_argument("--final-exploration-frames", type=int, default=10**6)
parser.add_argument("--final-epsilon", type=float, default=0.01)
parser.add_argument("--eval-epsilon", type=float, default=0.001)
parser.add_argument("--steps", type=int, default=5 * 10**7)
parser.add_argument(
"--max-frames",
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help="Maximum number of frames for each episode.",
)
parser.add_argument("--replay-start-size", type=int, default=5 * 10**4)
parser.add_argument("--target-update-interval", type=int, default=10**4)
parser.add_argument("--eval-interval", type=int, default=250000)
parser.add_argument("--eval-n-steps", type=int, default=125000)
parser.add_argument("--update-interval", type=int, default=4)
parser.add_argument("--batch-size", type=int, default=32)
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument("--batch-accumulator",
type=str,
default="mean",
choices=["mean", "sum"])
parser.add_argument("--quantile-thresholds-N", type=int, default=64)
parser.add_argument("--quantile-thresholds-N-prime", type=int, default=64)
parser.add_argument("--quantile-thresholds-K", type=int, default=32)
parser.add_argument("--n-best-episodes", type=int, default=200)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test,
)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, args.eval_epsilon)
if args.monitor:
env = pfrl.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training")
if args.render:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
n_actions = env.action_space.n
q_func = pfrl.agents.iqn.ImplicitQuantileQFunction(
psi=nn.Sequential(
nn.Conv2d(4, 32, 8, stride=4),
nn.ReLU(),
nn.Conv2d(32, 64, 4, stride=2),
nn.ReLU(),
nn.Conv2d(64, 64, 3, stride=1),
nn.ReLU(),
nn.Flatten(),
),
phi=nn.Sequential(
pfrl.agents.iqn.CosineBasisLinear(64, 3136),
nn.ReLU(),
),
f=nn.Sequential(
nn.Linear(3136, 512),
nn.ReLU(),
nn.Linear(512, n_actions),
),
)
# Use the same hyper parameters as https://arxiv.org/abs/1710.10044
opt = torch.optim.Adam(q_func.parameters(),
lr=5e-5,
eps=1e-2 / args.batch_size)
rbuf = replay_buffers.ReplayBuffer(10**6)
explorer = explorers.LinearDecayEpsilonGreedy(
1.0,
args.final_epsilon,
args.final_exploration_frames,
lambda: np.random.randint(n_actions),
)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
agent = pfrl.agents.IQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
batch_accumulator=args.batch_accumulator,
phi=phi,
quantile_thresholds_N=args.quantile_thresholds_N,
quantile_thresholds_N_prime=args.quantile_thresholds_N_prime,
quantile_thresholds_K=args.quantile_thresholds_K,
)
if args.load or args.load_pretrained:
# either load or load_pretrained must be false
assert not args.load or not args.load_pretrained
if args.load:
agent.load(args.load)
else:
agent.load(
utils.download_model("IQN",
args.env,
model_type=args.pretrained_type)[0])
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=args.eval_n_steps,
n_episodes=None,
)
print("n_steps: {} mean: {} median: {} stdev {}".format(
args.eval_n_steps,
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=args.eval_n_steps,
eval_n_episodes=None,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=True,
eval_env=eval_env,
)
dir_of_best_network = os.path.join(args.outdir, "best")
agent.load(dir_of_best_network)
# run 200 evaluation episodes, each capped at 30 mins of play
stats = experiments.evaluator.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.n_best_episodes,
max_episode_len=args.max_frames / 4,
logger=None,
)
with open(os.path.join(args.outdir, "bestscores.json"), "w") as f:
json.dump(stats, f)
print("The results of the best scoring network:")
for stat in stats:
print(str(stat) + ":" + str(stats[stat]))
def main():
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("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
parser.add_argument("--gpu", type=int, default=0)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load-pretrained",
action="store_true",
default=False)
parser.add_argument("--pretrained-type",
type=str,
default="best",
choices=["best", "final"])
parser.add_argument("--load", type=str, default=None)
parser.add_argument("--eval-epsilon", type=float, default=0.0)
parser.add_argument("--noisy-net-sigma", type=float, default=0.2)
parser.add_argument("--steps", type=int, default=5 * 10**7)
parser.add_argument(
"--max-frames",
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help="Maximum number of frames for each episode.",
)
parser.add_argument("--replay-start-size", type=int, default=2 * 10**3)
parser.add_argument("--eval-n-steps", type=int, default=125000)
parser.add_argument("--eval-interval", type=int, default=250000)
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument("--n-best-episodes", type=int, default=200)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
test_ID = datetime.datetime.now().strftime("%Y-%m-%d_%H_%M_%S")
args.outdir = experiments.prepare_output_dir(args, args.outdir, test_ID)
print("Output files are saved in {}".format(args.outdir))
env = MapRootEnv()
eval_env = MapRootEnv()
n_actions = env.action_space.n
input_shape = env.input_shape
n_atoms = 51
v_max = 10
v_min = -10
q_func = MyDistributionalDuelingDQN(n_actions, n_atoms, v_min, v_max,
input_shape[2])
# Noisy nets
pnn.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
# Use the same hyper parameters as https://arxiv.org/abs/1710.02298
opt = torch.optim.Adam(q_func.parameters(), 6.25e-5, eps=1.5 * 10**-4)
# Prioritized Replay
# Anneal beta from beta0 to 1 throughout training
update_interval = 4
betasteps = args.steps / update_interval
rbuf = replay_buffers.PrioritizedReplayBuffer(
5 * 10**4,
alpha=0.5,
beta0=0.4,
betasteps=betasteps,
num_steps=3,
normalize_by_max="batch",
)
Agent = agents.CategoricalDoubleDQN
agent = Agent(q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.80,
explorer=explorer,
minibatch_size=32,
replay_start_size=args.replay_start_size,
target_update_interval=32000,
update_interval=update_interval,
batch_accumulator="mean")
if args.demo:
eval_stats = experiments.eval_performance(env=eval_env,
agent=agent,
n_steps=args.eval_n_steps,
n_episodes=None)
print("n_episodes: {} mean: {} median: {} stdev {}".format(
eval_stats["episodes"],
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=args.eval_n_steps,
eval_n_episodes=None,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=True,
eval_env=eval_env,
logger=TBLogger(args.outdir))
# dir_of_best_network = os.path.join(args.outdir, "best")
# agent.load(dir_of_best_network)
# run 200 evaluation episodes, each capped at 30 mins of play
stats = experiments.evaluator.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.n_best_episodes,
max_episode_len=args.max_frames / 4,
logger=None)
with open(os.path.join(args.outdir, "bestscores.json"), "w") as f:
json.dump(stats, f)
print("The results of the best scoring network:")
for stat in stats:
print(str(stat) + ":" + str(stats[stat]))
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--env", type=str, default="BreakoutNoFrameskip-v4")
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("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
parser.add_argument("--gpu", type=int, default=0)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load-pretrained",
action="store_true",
default=False)
parser.add_argument("--pretrained-type",
type=str,
default="best",
choices=["best", "final"])
parser.add_argument("--load", type=str, default=None)
parser.add_argument("--eval-epsilon", type=float, default=0.0)
parser.add_argument("--noisy-net-sigma", type=float, default=0.5)
parser.add_argument("--steps", type=int, default=5 * 10**7)
parser.add_argument(
"--max-frames",
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help="Maximum number of frames for each episode.",
)
parser.add_argument("--replay-start-size", type=int, default=2 * 10**4)
parser.add_argument("--eval-n-steps", type=int, default=125000)
parser.add_argument("--eval-interval", type=int, default=250000)
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument("--n-best-episodes", type=int, default=200)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test,
)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, args.eval_epsilon)
if args.monitor:
env = pfrl.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training")
if args.render:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
n_actions = env.action_space.n
n_atoms = 51
v_max = 10
v_min = -10
q_func = DistributionalDuelingDQN(
n_actions,
n_atoms,
v_min,
v_max,
)
# Noisy nets
pnn.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
# Use the same hyper parameters as https://arxiv.org/abs/1710.02298
opt = torch.optim.Adam(q_func.parameters(), 6.25e-5, eps=1.5 * 10**-4)
# Prioritized Replay
# Anneal beta from beta0 to 1 throughout training
update_interval = 4
betasteps = args.steps / update_interval
rbuf = replay_buffers.PrioritizedReplayBuffer(
10**6,
alpha=0.5,
beta0=0.4,
betasteps=betasteps,
num_steps=3,
normalize_by_max="memory",
)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
Agent = agents.CategoricalDoubleDQN
agent = Agent(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
minibatch_size=32,
replay_start_size=args.replay_start_size,
target_update_interval=32000,
update_interval=update_interval,
batch_accumulator="mean",
phi=phi,
)
if args.load or args.load_pretrained:
# either load_ or load_pretrained must be false
assert not args.load or not args.load_pretrained
if args.load:
agent.load(args.load)
else:
agent.load(
utils.download_model("Rainbow",
args.env,
model_type=args.pretrained_type)[0])
if args.demo:
eval_stats = experiments.eval_performance(env=eval_env,
agent=agent,
n_steps=args.eval_n_steps,
n_episodes=None)
print("n_episodes: {} mean: {} median: {} stdev {}".format(
eval_stats["episodes"],
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=args.eval_n_steps,
eval_n_episodes=None,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=True,
eval_env=eval_env,
)
dir_of_best_network = os.path.join(args.outdir, "best")
agent.load(dir_of_best_network)
# run 200 evaluation episodes, each capped at 30 mins of play
stats = experiments.evaluator.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.n_best_episodes,
max_episode_len=args.max_frames / 4,
logger=None,
)
with open(os.path.join(args.outdir, "bestscores.json"), "w") as f:
json.dump(stats, f)
print("The results of the best scoring network:")
for stat in stats:
print(str(stat) + ":" + str(stats[stat]))
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--env",
type=str,
default="BreakoutNoFrameskip-v4",
help="OpenAI Atari domain to perform algorithm on.",
)
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("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
parser.add_argument("--gpu",
type=int,
default=0,
help="GPU to use, set to -1 if no GPU.")
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load", type=str, default=None)
parser.add_argument(
"--final-exploration-frames",
type=int,
default=10**6,
help="Timesteps after which we stop " + "annealing exploration rate",
)
parser.add_argument(
"--final-epsilon",
type=float,
default=0.01,
help="Final value of epsilon during training.",
)
parser.add_argument(
"--eval-epsilon",
type=float,
default=0.001,
help="Exploration epsilon used during eval episodes.",
)
parser.add_argument("--noisy-net-sigma", type=float, default=None)
parser.add_argument(
"--arch",
type=str,
default="doubledqn",
choices=["nature", "nips", "dueling", "doubledqn"],
help="Network architecture to use.",
)
parser.add_argument(
"--steps",
type=int,
default=5 * 10**7,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--max-frames",
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help="Maximum number of frames for each episode.",
)
parser.add_argument(
"--replay-start-size",
type=int,
default=5 * 10**4,
help="Minimum replay buffer size before " +
"performing gradient updates.",
)
parser.add_argument(
"--target-update-interval",
type=int,
default=3 * 10**4,
help="Frequency (in timesteps) at which " +
"the target network is updated.",
)
parser.add_argument(
"--eval-interval",
type=int,
default=10**5,
help="Frequency (in timesteps) of evaluation phase.",
)
parser.add_argument(
"--update-interval",
type=int,
default=4,
help="Frequency (in timesteps) of network updates.",
)
parser.add_argument("--eval-n-runs", type=int, default=10)
parser.add_argument("--no-clip-delta",
dest="clip_delta",
action="store_false")
parser.add_argument("--num-step-return", type=int, default=1)
parser.set_defaults(clip_delta=True)
parser.add_argument("--agent",
type=str,
default="DoubleDQN",
choices=["DQN", "DoubleDQN", "PAL"])
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument("--lr",
type=float,
default=2.5e-4,
help="Learning rate.")
parser.add_argument(
"--prioritized",
action="store_true",
default=False,
help="Use prioritized experience replay.",
)
parser.add_argument(
"--checkpoint-frequency",
type=int,
default=None,
help="Frequency at which agents are stored.",
)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test,
)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, args.eval_epsilon)
if args.monitor:
env = pfrl.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training")
if args.render:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
n_actions = env.action_space.n
q_func = parse_arch(args.arch, n_actions)
if args.noisy_net_sigma is not None:
pnn.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
else:
explorer = explorers.LinearDecayEpsilonGreedy(
1.0,
args.final_epsilon,
args.final_exploration_frames,
lambda: np.random.randint(n_actions),
)
# Use the Nature paper's hyperparameters
opt = pfrl.optimizers.RMSpropEpsInsideSqrt(
q_func.parameters(),
lr=args.lr,
alpha=0.95,
momentum=0.0,
eps=1e-2,
centered=True,
)
# Select a replay buffer to use
if args.prioritized:
# Anneal beta from beta0 to 1 throughout training
betasteps = args.steps / args.update_interval
rbuf = replay_buffers.PrioritizedReplayBuffer(
10**6,
alpha=0.6,
beta0=0.4,
betasteps=betasteps,
num_steps=args.num_step_return,
)
else:
rbuf = replay_buffers.ReplayBuffer(10**6, args.num_step_return)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
Agent = parse_agent(args.agent)
agent = Agent(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
clip_delta=args.clip_delta,
update_interval=args.update_interval,
batch_accumulator="sum",
phi=phi,
)
if args.load:
agent.load(args.load)
if args.demo:
eval_stats = experiments.eval_performance(env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs)
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,
checkpoint_freq=args.checkpoint_frequency,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=False,
eval_env=eval_env,
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--env", type=str, default="SlimeVolley-v0")
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("--seed", type=int, default=0, help="Random seed [0, 2 ** 32)")
parser.add_argument("--gpu", type=int, default=0)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load", type=str, default=None)
parser.add_argument("--noisy-net-sigma", type=float, default=0.1)
parser.add_argument("--steps", type=int, default=2 * 10 ** 6)
parser.add_argument("--replay-start-size", type=int, default=1600)
parser.add_argument("--eval-n-episodes", type=int, default=1000)
parser.add_argument("--eval-interval", type=int, default=250000)
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=(
"Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument("--gamma", type=float, default=0.98)
parser.add_argument("--v-max", type=float, default=1)
parser.add_argument("--n-step-return", type=int, default=3)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2 ** 31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
if "SlimeVolley" in args.env:
# You need to install slimevolleygym
import slimevolleygym # NOQA
env = gym.make(args.env)
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env.seed(int(env_seed))
if args.monitor:
env = pfrl.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training"
)
if args.render:
env = pfrl.wrappers.Render(env)
if isinstance(env.action_space, gym.spaces.MultiBinary):
env = MultiBinaryAsDiscreteAction(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
obs_size = env.observation_space.low.size
n_actions = env.action_space.n
n_atoms = 51
v_max = args.v_max
v_min = -args.v_max
hidden_size = 512
q_func = nn.Sequential(
nn.Linear(obs_size, hidden_size),
nn.ReLU(),
nn.Linear(hidden_size, hidden_size),
nn.ReLU(),
DistributionalDuelingHead(hidden_size, n_actions, n_atoms, v_min, v_max),
)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32)
# Noisy nets
pnn.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
# Use the same eps as https://arxiv.org/abs/1710.02298
opt = torch.optim.Adam(q_func.parameters(), 1e-4, eps=1.5e-4)
# Prioritized Replay
# Anneal beta from beta0 to 1 throughout training
update_interval = 1
betasteps = args.steps / update_interval
rbuf = replay_buffers.PrioritizedReplayBuffer(
10 ** 6,
alpha=0.5,
beta0=0.4,
betasteps=betasteps,
num_steps=args.n_step_return,
normalize_by_max="memory",
)
agent = agents.CategoricalDoubleDQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=args.gamma,
explorer=explorer,
minibatch_size=32,
replay_start_size=args.replay_start_size,
target_update_interval=2000,
update_interval=update_interval,
batch_accumulator="mean",
phi=phi,
max_grad_norm=10,
)
if args.load:
agent.load(args.load)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env, agent=agent, n_steps=None, n_episodes=args.eval_n_episodes,
)
print(
"n_episodes: {} mean: {} median: {} stdev {}".format(
eval_stats["episodes"],
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_episodes,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=True,
eval_env=eval_env,
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--env",
type=str,
default="BreakoutNoFrameskip-v4",
help="OpenAI Atari domain to perform algorithm on.",
)
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("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
parser.add_argument("--gpu",
type=int,
default=0,
help="GPU to use, set to -1 if no GPU.")
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load", type=str, default=None)
parser.add_argument(
"--final-exploration-frames",
type=int,
default=10**6,
help="Timesteps after which we stop " + "annealing exploration rate",
)
parser.add_argument(
"--final-epsilon",
type=float,
default=0.01,
help="Final value of epsilon during training.",
)
parser.add_argument(
"--eval-epsilon",
type=float,
default=0.001,
help="Exploration epsilon used during eval episodes.",
)
parser.add_argument(
"--steps",
type=int,
default=5 * 10**7,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--max-frames",
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help="Maximum number of frames for each episode.",
)
parser.add_argument(
"--replay-start-size",
type=int,
default=5 * 10**4,
help="Minimum replay buffer size before " +
"performing gradient updates.",
)
parser.add_argument(
"--target-update-interval",
type=int,
default=3 * 10**4,
help="Frequency (in timesteps) at which " +
"the target network is updated.",
)
parser.add_argument("--demo-n-episodes", type=int, default=30)
parser.add_argument("--eval-n-steps", type=int, default=125000)
parser.add_argument(
"--eval-interval",
type=int,
default=250000,
help="Frequency (in timesteps) of evaluation phase.",
)
parser.add_argument(
"--update-interval",
type=int,
default=4,
help="Frequency (in timesteps) of network updates.",
)
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument("--lr",
type=float,
default=2.5e-4,
help="Learning rate.")
parser.add_argument(
"--recurrent",
action="store_true",
default=False,
help="Use a recurrent model. See the code for the model definition.",
)
parser.add_argument(
"--flicker",
action="store_true",
default=False,
help=("Use so-called flickering Atari, where each"
" screen is blacked out with probability 0.5."),
)
parser.add_argument(
"--no-frame-stack",
action="store_true",
default=False,
help=
("Disable frame stacking so that the agent can only see the current screen."
),
)
parser.add_argument(
"--episodic-update-len",
type=int,
default=10,
help="Maximum length of sequences for updating recurrent models",
)
parser.add_argument(
"--batch-size",
type=int,
default=32,
help=("Number of transitions (in a non-recurrent case)"
" or sequences (in a recurrent case) used for an"
" update."),
)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test,
flicker=args.flicker,
frame_stack=not args.no_frame_stack,
)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, args.eval_epsilon)
if args.monitor:
env = gym.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training")
if args.render:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
print("Observation space", env.observation_space)
print("Action space", env.action_space)
n_frames = env.observation_space.shape[0]
n_actions = env.action_space.n
if args.recurrent:
# Q-network with LSTM
q_func = pfrl.nn.RecurrentSequential(
nn.Conv2d(n_frames, 32, 8, stride=4),
nn.ReLU(),
nn.Conv2d(32, 64, 4, stride=2),
nn.ReLU(),
nn.Conv2d(64, 64, 3, stride=1),
nn.Flatten(),
nn.ReLU(),
nn.LSTM(input_size=3136, hidden_size=512),
nn.Linear(512, n_actions),
DiscreteActionValueHead(),
)
# Replay buffer that stores whole episodes
rbuf = replay_buffers.EpisodicReplayBuffer(10**6)
else:
# Q-network without LSTM
q_func = nn.Sequential(
nn.Conv2d(n_frames, 32, 8, stride=4),
nn.ReLU(),
nn.Conv2d(32, 64, 4, stride=2),
nn.ReLU(),
nn.Conv2d(64, 64, 3, stride=1),
nn.Flatten(),
nn.Linear(3136, 512),
nn.ReLU(),
nn.Linear(512, n_actions),
DiscreteActionValueHead(),
)
# Replay buffer that stores transitions separately
rbuf = replay_buffers.ReplayBuffer(10**6)
explorer = explorers.LinearDecayEpsilonGreedy(
1.0,
args.final_epsilon,
args.final_exploration_frames,
lambda: np.random.randint(n_actions),
)
opt = torch.optim.Adam(q_func.parameters(), lr=1e-4, eps=1e-4)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
agent = pfrl.agents.DoubleDQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
batch_accumulator="mean",
phi=phi,
minibatch_size=args.batch_size,
episodic_update_len=args.episodic_update_len,
recurrent=args.recurrent,
)
if args.load:
agent.load(args.load)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.demo_n_episodes,
)
print("n_runs: {} mean: {} median: {} stdev {}".format(
args.demo_n_episodes,
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=args.eval_n_steps,
eval_n_episodes=None,
eval_interval=args.eval_interval,
outdir=args.outdir,
eval_env=eval_env,
)
