def _test_load_rainbow(self, gpu):
from pfrl.q_functions import DistributionalDuelingDQN
q_func = DistributionalDuelingDQN(4, 51, -10, 10)
pnn.to_factorized_noisy(q_func, sigma_scale=0.5)
explorer = explorers.Greedy()
opt = torch.optim.Adam(q_func.parameters(), 6.25e-5, eps=1.5 * 10**-4)
rbuf = replay_buffers.ReplayBuffer(100)
agent = agents.CategoricalDoubleDQN(
q_func,
opt,
rbuf,
gpu=gpu,
gamma=0.99,
explorer=explorer,
minibatch_size=32,
replay_start_size=50,
target_update_interval=32000,
update_interval=4,
batch_accumulator="mean",
phi=lambda x: x,
)
downloaded_model, exists = download_model(
"Rainbow",
"BreakoutNoFrameskip-v4",
model_type=self.pretrained_type)
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def _test_load_td3(self, gpu):
obs_size = 11
action_size = 3
def make_q_func_with_optimizer():
q_func = nn.Sequential(
pnn.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()
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())
rbuf = replay_buffers.ReplayBuffer(100)
explorer = explorers.AdditiveGaussian(scale=0.1,
low=[-1.0, -1.0, -1.0],
high=[1.0, 1.0, 1.0])
agent = 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=1000,
gpu=gpu,
minibatch_size=100,
burnin_action_func=None,
)
downloaded_model, exists = download_model(
"TD3", "Hopper-v2", model_type=self.pretrained_type)
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def _test_load_ddpg(self, gpu):
obs_size = 11
action_size = 3
from pfrl.nn import ConcatObsAndAction
q_func = nn.Sequential(
ConcatObsAndAction(),
nn.Linear(obs_size + action_size, 400),
nn.ReLU(),
nn.Linear(400, 300),
nn.ReLU(),
nn.Linear(300, 1),
)
from pfrl.nn import BoundByTanh
from pfrl.policies import DeterministicHead
policy = nn.Sequential(
nn.Linear(obs_size, 400),
nn.ReLU(),
nn.Linear(400, 300),
nn.ReLU(),
nn.Linear(300, action_size),
BoundByTanh(low=[-1.0, -1.0, -1.0], high=[1.0, 1.0, 1.0]),
DeterministicHead(),
)
opt_a = torch.optim.Adam(policy.parameters())
opt_c = torch.optim.Adam(q_func.parameters())
explorer = explorers.AdditiveGaussian(scale=0.1,
low=[-1.0, -1.0, -1.0],
high=[1.0, 1.0, 1.0])
agent = agents.DDPG(
policy,
q_func,
opt_a,
opt_c,
replay_buffers.ReplayBuffer(100),
gamma=0.99,
explorer=explorer,
replay_start_size=1000,
target_update_method="soft",
target_update_interval=1,
update_interval=1,
soft_update_tau=5e-3,
n_times_update=1,
gpu=gpu,
minibatch_size=100,
burnin_action_func=None,
)
downloaded_model, exists = download_model(
"DDPG", "Hopper-v2", model_type=self.pretrained_type)
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def _test_load_iqn(self, gpu):
n_actions = 4
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 / 32)
rbuf = replay_buffers.ReplayBuffer(100)
explorer = explorers.LinearDecayEpsilonGreedy(
start_epsilon=1.0,
end_epsilon=0.1,
decay_steps=10**6,
random_action_func=lambda: np.random.randint(4),
)
agent = agents.IQN(
q_func,
opt,
rbuf,
gpu=gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=50,
target_update_interval=10**4,
update_interval=4,
batch_accumulator="mean",
phi=lambda x: x,
quantile_thresholds_N=64,
quantile_thresholds_N_prime=64,
quantile_thresholds_K=32,
)
downloaded_model, exists = download_model(
"IQN", "BreakoutNoFrameskip-v4", model_type=self.pretrained_type)
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def _test_load_ppo(self, gpu):
obs_size = 11
action_size = 3
from pfrl.policies import GaussianHeadWithStateIndependentCovariance
policy = torch.nn.Sequential(
nn.Linear(obs_size, 64),
nn.Tanh(),
nn.Linear(64, 64),
nn.Tanh(),
nn.Linear(64, action_size),
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),
)
model = pnn.Branched(policy, vf)
opt = torch.optim.Adam(model.parameters(), lr=3e-4, eps=1e-5)
agent = agents.PPO(
model,
opt,
obs_normalizer=None,
gpu=gpu,
update_interval=2048,
minibatch_size=64,
epochs=10,
clip_eps_vf=None,
entropy_coef=0,
standardize_advantages=True,
gamma=0.995,
lambd=0.97,
)
downloaded_model, exists = download_model("PPO",
"Hopper-v2",
model_type="final")
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def _test_load_dqn(self, gpu):
from pfrl.q_functions import DiscreteActionValueHead
n_actions = 4
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(100)
explorer = explorers.LinearDecayEpsilonGreedy(
start_epsilon=1.0,
end_epsilon=0.1,
decay_steps=10**6,
random_action_func=lambda: np.random.randint(4),
)
agent = agents.DQN(
q_func,
opt,
rbuf,
gpu=gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=50,
target_update_interval=10**4,
clip_delta=True,
update_interval=4,
batch_accumulator="sum",
phi=lambda x: x,
)
downloaded_model, exists = download_model(
"DQN", "BreakoutNoFrameskip-v4", model_type=self.pretrained_type)
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def _test_load_trpo(self, gpu):
obs_size = 11
action_size = 3
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),
)
vf_opt = torch.optim.Adam(vf.parameters())
agent = agents.TRPO(
policy=policy,
vf=vf,
vf_optimizer=vf_opt,
gpu=gpu,
update_interval=5000,
max_kl=0.01,
conjugate_gradient_max_iter=20,
conjugate_gradient_damping=1e-1,
gamma=0.995,
lambd=0.97,
vf_epochs=5,
entropy_coef=0,
)
downloaded_model, exists = download_model(
"TRPO", "Hopper-v2", model_type=self.pretrained_type)
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def test_load_a3c(self):
from pfrl.policies import SoftmaxCategoricalHead
obs_size = 4
n_actions = 4
a3c_model = nn.Sequential(
nn.Conv2d(obs_size, 16, 8, stride=4),
nn.ReLU(),
nn.Conv2d(16, 32, 4, stride=2),
nn.ReLU(),
nn.Flatten(),
nn.Linear(2592, 256),
nn.ReLU(),
pfrl.nn.Branched(
nn.Sequential(
nn.Linear(256, n_actions),
SoftmaxCategoricalHead(),
),
nn.Linear(256, 1),
),
)
from pfrl.optimizers import SharedRMSpropEpsInsideSqrt
opt = SharedRMSpropEpsInsideSqrt(a3c_model.parameters(),
lr=7e-4,
eps=1e-1,
alpha=0.99)
agent = agents.A3C(a3c_model,
opt,
t_max=5,
gamma=0.99,
beta=1e-2,
phi=lambda x: x)
downloaded_model, exists = download_model(
"A3C", "BreakoutNoFrameskip-v4", model_type=self.pretrained_type)
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def _test_load_sac(self, gpu):
obs_size = 11
action_size = 3
def squashed_diagonal_gaussian_head(x):
assert x.shape[-1] == action_size * 2
mean, log_scale = torch.chunk(x, 2, dim=1)
log_scale = torch.clamp(log_scale, -20.0, 2.0)
var = torch.exp(log_scale * 2)
from torch import distributions
base_distribution = distributions.Independent(
distributions.Normal(loc=mean, scale=torch.sqrt(var)), 1)
# cache_size=1 is required for numerical stability
return distributions.transformed_distribution.TransformedDistribution(
base_distribution,
[distributions.transforms.TanhTransform(cache_size=1)],
)
from pfrl.nn.lmbda import Lambda
policy = nn.Sequential(
nn.Linear(obs_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, action_size * 2),
Lambda(squashed_diagonal_gaussian_head),
)
policy_optimizer = torch.optim.Adam(policy.parameters(), lr=3e-4)
def make_q_func_with_optimizer():
q_func = nn.Sequential(
pfrl.nn.ConcatObsAndAction(),
nn.Linear(obs_size + action_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, 1),
)
torch.nn.init.xavier_uniform_(q_func[1].weight)
torch.nn.init.xavier_uniform_(q_func[3].weight)
torch.nn.init.xavier_uniform_(q_func[5].weight)
q_func_optimizer = torch.optim.Adam(q_func.parameters(), lr=3e-4)
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()
agent = agents.SoftActorCritic(
policy,
q_func1,
q_func2,
policy_optimizer,
q_func1_optimizer,
q_func2_optimizer,
replay_buffers.ReplayBuffer(100),
gamma=0.99,
replay_start_size=1000,
gpu=gpu,
minibatch_size=256,
burnin_action_func=None,
entropy_target=-3,
temperature_optimizer_lr=3e-4,
)
downloaded_model, exists = download_model(
"SAC", "Hopper-v2", model_type=self.pretrained_type)
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def main():
import logging
torch.cuda.empty_cache()
parser = argparse.ArgumentParser()
parser.add_argument('--gpu',
type=int,
default=0,
help='GPU to use, set to -1 if no GPU')
parser.add_argument('--env',
type=str,
default='LidarBat-v0',
help='Bat simulation env')
parser.add_argument('--arch',
type=str,
default='FFGaussian',
choices=('FFSoftmax', 'FFMellowmax', 'FFGaussian'))
parser.add_argument('--bound-mean', action='store_true')
parser.add_argument('--seed',
type=int,
default=0,
help='Random seed [0, 2 ** 32)')
parser.add_argument('--outdir',
type=str,
default='data/ppo',
help='Directory path to save output files.'
' If it does not exist, it will be created.')
parser.add_argument('--steps', type=int, default=10**6)
parser.add_argument('--eval-interval', type=int, default=10000)
parser.add_argument('--eval-n-runs', type=int, default=10)
parser.add_argument('--reward-scale-factor', type=float, default=1e-2)
parser.add_argument('--standardize-advantages', action='store_true')
parser.add_argument('--render', action='store_true', default=False)
parser.add_argument('--lr', type=float, default=3e-4)
parser.add_argument('--weight-decay', type=float, default=0.0)
parser.add_argument('--demo', action='store_true', default=False)
parser.add_argument('--load', type=str, default='')
parser.add_argument("--load-pretrained",
action="store_true",
default=False)
parser.add_argument('--logger-level', type=int, default=logging.DEBUG)
parser.add_argument('--monitor', action='store_true')
parser.add_argument(
"--log-interval",
type=int,
default=1000,
help=
"Interval in timesteps between outputting log messages during training",
)
parser.add_argument("--num-envs",
type=int,
default=1,
help="Number of envs run in parallel.")
parser.add_argument("--batch-size",
type=int,
default=64,
help="Minibatch size")
parser.add_argument('--update-interval', type=int, default=2048)
parser.add_argument('--batchsize', type=int, default=64)
parser.add_argument('--epochs', type=int, default=10)
parser.add_argument('--entropy-coef', type=float, default=0.0)
args = parser.parse_args()
logging.basicConfig(level=args.logger_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**32
args.outdir = experiments.prepare_output_dir(args, args.outdir)
def make_env(process_idx, test):
env = gym.make(args.env)
# Use different random seeds for train and test envs
process_seed = int(process_seeds[process_idx])
env_seed = 2**32 - 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)
if args.monitor:
env = pfrl.wrappers.Monitor(env, args.outdir)
# TODO
# if not test is not here
if args.render:
env = pfrl.wrappers.Render(env)
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 = gym.make(args.env)
timestep_limit = 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)
# pulicy here magic number must be concidered again
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=args.lr, 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=args.entropy_coef,
standardize_advantages=True,
gamma=0.995,
lambd=0.97,
)
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_batch_env(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"],
))
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=False,
)
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():
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("--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("--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=256,
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-interval",
type=int,
default=1000,
help=
"Interval in timesteps between outputting log messages during training",
)
parser.add_argument("--log-level",
type=int,
default=logging.INFO,
help="Level of the root logger.")
parser.add_argument(
"--policy-output-scale",
type=float,
default=1.0,
help="Weight initialization scale of policy output.",
)
parser.add_argument(
"--optimizer",
type=str,
default="AdaBelief",
)
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)
# 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 make_env(process_idx, test):
env = gym.make(args.env)
# Unwrap TimiLimit wrapper
assert isinstance(env, gym.wrappers.TimeLimit)
env = env.env
# Use different random seeds for train and test envs
process_seed = int(process_seeds[process_idx])
env_seed = 2**32 - 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)
# Normalize action space to [-1, 1]^n
env = pfrl.wrappers.NormalizeActionSpace(env)
if args.monitor:
env = gym.wrappers.Monitor(env, args.outdir)
if args.render:
env = pfrl.wrappers.Render(env)
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))
])
sample_env = make_env(process_idx=0, test=False)
timestep_limit = 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)
obs_size = obs_space.low.size
action_size = action_space.low.size
if LooseVersion(torch.__version__) < LooseVersion("1.5.0"):
raise Exception("This script requires a PyTorch version >= 1.5.0")
def squashed_diagonal_gaussian_head(x):
assert x.shape[-1] == action_size * 2
mean, log_scale = torch.chunk(x, 2, dim=1)
log_scale = torch.clamp(log_scale, -20.0, 2.0)
var = torch.exp(log_scale * 2)
base_distribution = distributions.Independent(
distributions.Normal(loc=mean, scale=torch.sqrt(var)), 1)
# cache_size=1 is required for numerical stability
return distributions.transformed_distribution.TransformedDistribution(
base_distribution,
[distributions.transforms.TanhTransform(cache_size=1)])
def make_optimizer(parameters):
if args.optimizer == "OfficialAdaBelief":
import adabelief_pytorch
optim_class = adabelief_pytorch.AdaBelief
optim = optim_class(parameters, betas=(0.9, 0.999), eps=1e-12)
else:
optim_class = getattr(
torch_optimizer,
args.optimizer,
getattr(torch.optim, args.optimizer, None),
)
optim = optim_class(parameters)
assert optim_class is not None
print(str(optim_class), "with default hyperparameters")
return optim
policy = nn.Sequential(
nn.Linear(obs_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, action_size * 2),
Lambda(squashed_diagonal_gaussian_head),
)
torch.nn.init.xavier_uniform_(policy[0].weight)
torch.nn.init.xavier_uniform_(policy[2].weight)
torch.nn.init.xavier_uniform_(policy[4].weight,
gain=args.policy_output_scale)
policy_optimizer = make_optimizer(policy.parameters())
def make_q_func_with_optimizer():
q_func = nn.Sequential(
pfrl.nn.ConcatObsAndAction(),
nn.Linear(obs_size + action_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, 1),
)
torch.nn.init.xavier_uniform_(q_func[1].weight)
torch.nn.init.xavier_uniform_(q_func[3].weight)
torch.nn.init.xavier_uniform_(q_func[5].weight)
q_func_optimizer = make_optimizer(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)
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.SoftActorCritic(
policy,
q_func1,
q_func2,
policy_optimizer,
q_func1_optimizer,
q_func2_optimizer,
rbuf,
gamma=0.99,
replay_start_size=args.replay_start_size,
gpu=args.gpu,
minibatch_size=args.batch_size,
burnin_action_func=burnin_action_func,
entropy_target=-action_size,
temperature_optimizer_lr=3e-4,
)
if len(args.load) > 0 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 len(args.load) > 0 or not args.load_pretrained
if len(args.load) > 0:
agent.load(args.load)
else:
agent.load(
utils.download_model("SAC",
args.env,
model_type=args.pretrained_type)[0])
if args.demo:
eval_stats = experiments.eval_performance(
env=make_batch_env(test=True),
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_batch_with_evaluation(
agent=agent,
env=make_batch_env(test=False),
eval_env=make_batch_env(test=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,
)
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(
"--env",
type=str,
default="reach_target-ee-vision-v0",
help="OpenAI Gym MuJoCo env to perform algorithm on.",
)
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(
"--action-size",
type=int,
default=3,
help="Action size (needs to match env.action_space)",
)
parser.add_argument("--batch-size", type=int, default=64, help="Minibatch size")
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 ** 32
args.outdir = experiments.prepare_output_dir(args, args.outdir)
def make_env(process_idx, test):
render_mode = 'human' if args.render else None
env = NormalizeAction(GraspActionWrapper(FlattenObservation(ResizeObservation(WristObsWrapper(gym.make(args.env, render_mode=render_mode)), (64, 64))), args.action_size))
# env = GraspActionWrapper(RescaleAction(FlattenObservation(ResizeObservation(WristObsWrapper(gym.make(args.env)), (64, 64))), -0.5, 0.5))
# Use different random seeds for train and test envs
process_seed = int(process_seeds[process_idx])
env_seed = 2 ** 32 - 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)
if args.monitor:
env = pfrl.wrappers.Monitor(env, args.outdir)
if args.render:
env = pfrl.wrappers.Render(env)
return env
def make_batch_env(test):
return 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 = RescaleAction(GraspActionWrapper(FlattenObservation(ResizeObservation(WristObsWrapper(gym.make(args.env)), (64, 64))), args.action_size), -0.5, 0.5)
# timestep_limit = sample_env.spec.max_episode_steps
timestep_limit = 200
# obs_space = sample_env.observation_space
obs_space = spaces.Box(low=0, high=1, shape=(64 * 64 * 3,))
# action_space = sample_env.action_space
action_space = spaces.Box(low=-1.0, high=1.0, shape=(args.action_size,))
print("Observation space:", obs_space)
print("Action space:", action_space)
# assert obs_space == spaces.Box(low=0, high=1, shape=(64 * 64 * 3,))
# assert action_space == spaces.Box(low=-1.0, high=1.0, shape=(args.action_size,))
# sample_env.close()
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=0.995,
lambd=0.97,
)
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_batch_env(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"],
)
)
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(
"--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("--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("--processes", type=int, default=16)
parser.add_argument("--env", type=str, default="BreakoutNoFrameskip-v4")
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
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("--t-max", type=int, default=5)
parser.add_argument("--beta", type=float, default=1e-2)
parser.add_argument("--profile", action="store_true")
parser.add_argument("--steps", type=int, default=8 * 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("--lr", type=float, default=7e-4)
parser.add_argument("--eval-interval", type=int, default=250000)
parser.add_argument("--eval-n-steps", type=int, default=125000)
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="")
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.
# If you use more than one processes, the results will be no longer
# deterministic even with the same random seed.
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.processes) + args.seed * args.processes
assert process_seeds.max() < 2**31
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(process_idx, test):
# Use different random seeds for train and test envs
process_seed = process_seeds[process_idx]
env_seed = 2**31 - 1 - process_seed if test else process_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 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
sample_env = make_env(0, False)
obs_size = sample_env.observation_space.low.shape[0]
n_actions = sample_env.action_space.n
model = nn.Sequential(
nn.Conv2d(obs_size, 16, 8, stride=4),
nn.ReLU(),
nn.Conv2d(16, 32, 4, stride=2),
nn.ReLU(),
nn.Flatten(),
nn.Linear(2592, 256),
nn.ReLU(),
pfrl.nn.Branched(
nn.Sequential(
nn.Linear(256, n_actions),
SoftmaxCategoricalHead(),
),
nn.Linear(256, 1),
),
)
# SharedRMSprop is same as torch.optim.RMSprop except that it initializes
# its state in __init__, allowing it to be moved to shared memory.
opt = SharedRMSpropEpsInsideSqrt(model.parameters(),
lr=7e-4,
eps=1e-1,
alpha=0.99)
assert opt.state_dict()["state"], (
"To share optimizer state across processes, the state must be"
" initialized before training.")
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
agent = a3c.A3C(
model,
opt,
t_max=args.t_max,
gamma=0.99,
beta=args.beta,
phi=phi,
max_grad_norm=40.0,
)
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("A3C",
args.env,
model_type=args.pretrained_type)[0])
if args.demo:
env = make_env(0, True)
eval_stats = experiments.eval_performance(env=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:
# Linearly decay the learning rate to zero
def lr_setter(env, agent, value):
for pg in agent.optimizer.param_groups:
assert "lr" in pg
pg["lr"] = value
lr_decay_hook = experiments.LinearInterpolationHook(
args.steps, args.lr, 0, lr_setter)
experiments.train_agent_async(
agent=agent,
outdir=args.outdir,
processes=args.processes,
make_env=make_env,
profile=args.profile,
steps=args.steps,
eval_n_steps=args.eval_n_steps,
eval_n_episodes=None,
eval_interval=args.eval_interval,
global_step_hooks=[lr_decay_hook],
save_best_so_far_agent=True,
)
