def _test_abc_batch(self,
steps=100000,
require_success=True,
gpu=-1,
load_model=False,
num_envs=4):
env, _ = self.make_vec_env_and_successful_return(test=False,
num_envs=num_envs)
test_env, successful_return = self.make_vec_env_and_successful_return(
test=True, num_envs=num_envs)
agent = self.make_agent(env, gpu)
max_episode_len = None if self.episodic else 2
if load_model:
print("Load agent from", self.agent_dirname)
agent.load(self.agent_dirname)
# Train
train_agent_batch_with_evaluation(
agent=agent,
env=env,
steps=steps,
outdir=self.tmpdir,
eval_interval=200,
eval_n_steps=None,
eval_n_episodes=40,
successful_score=successful_return,
eval_env=test_env,
log_interval=100,
max_episode_len=max_episode_len,
)
env.close()
# Test
n_test_runs = 10
eval_returns, _ = batch_run_evaluation_episodes(
test_env,
agent,
n_steps=None,
n_episodes=n_test_runs,
max_episode_len=max_episode_len,
)
test_env.close()
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)
def _test_batch_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_vec_env_and_successful_return(test=False)
test_env, successful_return = self.make_vec_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_batch_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,
)
env.close()
# Test
n_test_runs = 5
eval_returns, _ = batch_run_evaluation_episodes(
test_env,
agent,
n_steps=None,
n_episodes=n_test_runs,
)
test_env.close()
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 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="RoboschoolAtlasForwardWalk-v1",
help="OpenAI Gym env to perform algorithm on.",
)
parser.add_argument("--num-envs",
type=int,
default=4,
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**7,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--eval-n-runs",
type=int,
default=20,
help="Number of episodes run for each evaluation.",
)
parser.add_argument(
"--eval-interval",
type=int,
default=100000,
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(
"--update-interval",
type=int,
default=1,
help="Interval in timesteps between model 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("--monitor",
action="store_true",
help="Wrap env with Monitor to write videos.")
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(
"--n-hidden-channels",
type=int,
default=1024,
help="Number of hidden channels of NN models.",
)
parser.add_argument("--discount",
type=float,
default=0.98,
help="Discount factor.")
parser.add_argument("--n-step-return",
type=int,
default=3,
help="N-step return.")
parser.add_argument("--lr",
type=float,
default=3e-4,
help="Learning rate.")
parser.add_argument("--adam-eps",
type=float,
default=1e-1,
help="Adam eps.")
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_batch_env(test):
return pfrl.envs.MultiprocessVectorEnv([
functools.partial(make_env, args, process_seeds[idx], test)
for idx, env in enumerate(range(args.num_envs))
])
sample_env = make_env(args, process_seeds[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)
del sample_env
action_size = action_space.low.size
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)])
policy = nn.Sequential(
nn.Linear(obs_space.low.size, args.n_hidden_channels),
nn.ReLU(),
nn.Linear(args.n_hidden_channels, args.n_hidden_channels),
nn.ReLU(),
nn.Linear(args.n_hidden_channels, 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)
policy_optimizer = torch.optim.Adam(policy.parameters(),
lr=args.lr,
eps=args.adam_eps)
def make_q_func_with_optimizer():
q_func = nn.Sequential(
pfrl.nn.ConcatObsAndAction(),
nn.Linear(obs_space.low.size + action_size,
args.n_hidden_channels),
nn.ReLU(),
nn.Linear(args.n_hidden_channels, args.n_hidden_channels),
nn.ReLU(),
nn.Linear(args.n_hidden_channels, 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=args.lr,
eps=args.adam_eps)
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, num_steps=args.n_step_return)
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=args.discount,
update_interval=args.update_interval,
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=args.lr,
)
if len(args.load) > 0:
agent.load(args.load)
if args.demo:
eval_env = make_env(args, seed=0, test=True)
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_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(
"--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():
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, help="GPU to use, set to -1 if no GPU."
)
parser.add_argument(
"--demo",
action="store_true",
default=False,
help="Evaluate the agent without training.",
)
parser.add_argument(
"--load",
type=str,
default=None,
help="Load a saved agent from a given directory.",
)
parser.add_argument(
"--final-exploration-steps",
type=int,
default=5 * 10 ** 5,
help="Timesteps after which we stop annealing exploration rate",
)
parser.add_argument(
"--final-epsilon",
type=float,
default=0.2,
help="Final value of epsilon during training.",
)
parser.add_argument(
"--steps",
type=int,
default=2 * 10 ** 6,
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(
"--target-update-interval",
type=int,
default=1 * 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=1,
help="Frequency (in timesteps) of network updates.",
)
parser.add_argument(
"--eval-n-runs",
type=int,
default=100,
help="Number of episodes used for evaluation.",
)
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("--lr", type=float, default=6.25e-5, help="Learning rate")
parser.add_argument(
"--num-envs", type=int, default=1, help="Number of envs run in parallel."
)
parser.add_argument(
"--batch-size", type=int, default=32, help="Batch size used for training."
)
parser.add_argument(
"--record",
action="store_true",
default=False,
help="Record videos of evaluation envs. --render should also be specified.",
)
parser.add_argument("--gamma", type=float, default=0.99, help="Discount factor.")
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 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)
print("Output files are saved in {}".format(args.outdir))
max_episode_steps = 8
def make_env(idx, test):
from pybullet_envs.bullet.kuka_diverse_object_gym_env import (
KukaDiverseObjectEnv,
) # NOQA
# 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
# Set a random seed for this subprocess
utils.set_random_seed(env_seed)
env = KukaDiverseObjectEnv(
isDiscrete=True,
renders=args.render and (args.demo or not test),
height=84,
width=84,
maxSteps=max_episode_steps,
isTest=test,
)
# Disable file caching to keep memory usage small
env._p.setPhysicsEngineParameter(enableFileCaching=False)
assert env.observation_space is None
env.observation_space = gym.spaces.Box(
low=0, high=255, shape=(84, 84, 3), dtype=np.uint8
)
# (84, 84, 3) -> (3, 84, 84)
env = TransposeObservation(env, (2, 0, 1))
env = ObserveElapsedSteps(env, max_episode_steps)
# KukaDiverseObjectEnv internally asserts int actions
env = CastAction(env, int)
env.seed(int(env_seed))
if test and args.record:
assert args.render, "To use --record, --render needs be specified."
video_dir = os.path.join(args.outdir, "video_{}".format(idx))
os.mkdir(video_dir)
env = RecordMovie(env, video_dir)
return env
def make_batch_env(test):
return pfrl.envs.MultiprocessVectorEnv(
[functools.partial(make_env, idx, test) for idx in range(args.num_envs)]
)
eval_env = make_batch_env(test=True)
n_actions = eval_env.action_space.n
q_func = GraspingQFunction(n_actions, max_episode_steps)
# Use the hyper parameters of the Nature paper
opt = pfrl.optimizers.RMSpropEpsInsideSqrt(
q_func.parameters(),
lr=args.lr,
alpha=0.95,
momentum=0.0,
eps=1e-2,
centered=True,
)
# 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
)
explorer = explorers.LinearDecayEpsilonGreedy(
1.0,
args.final_epsilon,
args.final_exploration_steps,
lambda: np.random.randint(n_actions),
)
def phi(x):
# Feature extractor
image, elapsed_steps = x
# Normalize RGB values: [0, 255] -> [0, 1]
norm_image = np.asarray(image, dtype=np.float32) / 255
return norm_image, elapsed_steps
agent = pfrl.agents.DoubleDQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=args.gamma,
explorer=explorer,
minibatch_size=args.batch_size,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
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_batch_with_evaluation(
agent=agent,
env=make_batch_env(test=False),
eval_env=eval_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,
log_interval=1000,
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--env",
type=str,
default="BreakoutNoFrameskip-v4",
help="Gym Env ID.")
parser.add_argument("--gpu",
type=int,
default=0,
help="GPU device ID. Set to -1 to use CPUs only.")
parser.add_argument(
"--num-envs",
type=int,
default=8,
help="Number of env instances 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=10**7,
help="Total time steps for training.")
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=2.5e-4,
help="Learning rate.")
parser.add_argument(
"--eval-interval",
type=int,
default=100000,
help="Interval (in timesteps) between evaluation phases.",
)
parser.add_argument(
"--eval-n-runs",
type=int,
default=10,
help="Number of episodes ran in an evaluation phase.",
)
parser.add_argument(
"--demo",
action="store_true",
default=False,
help="Run demo episodes, not training.",
)
parser.add_argument(
"--load",
type=str,
default="",
help=("Directory path to load a saved agent data from"
" if it is a non-empty string."),
)
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(
"--update-interval",
type=int,
default=128 * 8,
help="Interval (in timesteps) between PPO iterations.",
)
parser.add_argument(
"--batchsize",
type=int,
default=32 * 8,
help="Size of minibatch (in timesteps).",
)
parser.add_argument(
"--epochs",
type=int,
default=4,
help="Number of epochs used for each PPO iteration.",
)
parser.add_argument(
"--log-interval",
type=int,
default=10000,
help="Interval (in timesteps) of printing logs.",
)
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(
"--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 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)
print("Output files are saved in {}".format(args.outdir))
def make_env(idx, test):
# 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
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=False,
)
env.seed(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
def make_batch_env(test):
vec_env = pfrl.envs.MultiprocessVectorEnv([
(lambda: make_env(idx, test))
for idx, env in enumerate(range(args.num_envs))
])
if not args.no_frame_stack:
vec_env = pfrl.wrappers.VectorFrameStack(vec_env, 4)
return vec_env
sample_env = make_batch_env(test=False)
print("Observation space", sample_env.observation_space)
print("Action space", sample_env.action_space)
n_actions = sample_env.action_space.n
obs_n_channels = sample_env.observation_space.low.shape[0]
del sample_env
def lecun_init(layer, gain=1):
if isinstance(layer, (nn.Conv2d, nn.Linear)):
pfrl.initializers.init_lecun_normal(layer.weight, gain)
nn.init.zeros_(layer.bias)
else:
pfrl.initializers.init_lecun_normal(layer.weight_ih_l0, gain)
pfrl.initializers.init_lecun_normal(layer.weight_hh_l0, gain)
nn.init.zeros_(layer.bias_ih_l0)
nn.init.zeros_(layer.bias_hh_l0)
return layer
if args.recurrent:
model = pfrl.nn.RecurrentSequential(
lecun_init(nn.Conv2d(obs_n_channels, 32, 8, stride=4)),
nn.ReLU(),
lecun_init(nn.Conv2d(32, 64, 4, stride=2)),
nn.ReLU(),
lecun_init(nn.Conv2d(64, 64, 3, stride=1)),
nn.ReLU(),
nn.Flatten(),
lecun_init(nn.Linear(3136, 512)),
nn.ReLU(),
lecun_init(nn.GRU(num_layers=1, input_size=512, hidden_size=512)),
pfrl.nn.Branched(
nn.Sequential(
lecun_init(nn.Linear(512, n_actions), 1e-2),
SoftmaxCategoricalHead(),
),
lecun_init(nn.Linear(512, 1)),
),
)
else:
model = nn.Sequential(
lecun_init(nn.Conv2d(obs_n_channels, 32, 8, stride=4)),
nn.ReLU(),
lecun_init(nn.Conv2d(32, 64, 4, stride=2)),
nn.ReLU(),
lecun_init(nn.Conv2d(64, 64, 3, stride=1)),
nn.ReLU(),
nn.Flatten(),
lecun_init(nn.Linear(3136, 512)),
nn.ReLU(),
pfrl.nn.Branched(
nn.Sequential(
lecun_init(nn.Linear(512, n_actions), 1e-2),
SoftmaxCategoricalHead(),
),
lecun_init(nn.Linear(512, 1)),
),
)
opt = torch.optim.Adam(model.parameters(), lr=args.lr, eps=1e-5)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
agent = PPO(
model,
opt,
gpu=args.gpu,
phi=phi,
update_interval=args.update_interval,
minibatch_size=args.batchsize,
epochs=args.epochs,
clip_eps=0.1,
clip_eps_vf=None,
standardize_advantages=True,
entropy_coef=1e-2,
recurrent=args.recurrent,
max_grad_norm=0.5,
)
if args.load:
agent.load(args.load)
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,
)
print("n_runs: {} mean: {} median: {} stdev: {}".format(
args.eval_n_runs,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
else:
step_hooks = []
# Linearly decay the learning rate to zero
def lr_setter(env, agent, value):
for param_group in agent.optimizer.param_groups:
param_group["lr"] = value
step_hooks.append(
experiments.LinearInterpolationHook(args.steps, args.lr, 0,
lr_setter))
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,
checkpoint_freq=args.checkpoint_frequency,
eval_interval=args.eval_interval,
log_interval=args.log_interval,
save_best_so_far_agent=False,
step_hooks=step_hooks,
)
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=0.01)
parser.add_argument("--eval-epsilon", type=float, default=0.001)
parser.add_argument("--noisy-net-sigma", type=float, default=None)
parser.add_argument(
"--arch",
type=str,
default="doubledqn",
choices=["nature", "nips", "dueling", "doubledqn"],
)
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=3 * 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("--no-clip-delta",
dest="clip_delta",
action="store_false")
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("--num-envs", type=int, default=1)
parser.add_argument("--n-step-return", type=int, default=1)
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 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)
print("Output files are saved in {}".format(args.outdir))
def make_env(idx, test):
# 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
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test,
frame_stack=False,
)
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, args.eval_epsilon)
env.seed(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
def make_batch_env(test):
vec_env = pfrl.envs.MultiprocessVectorEnv([
functools.partial(make_env, idx, test)
for idx, env in enumerate(range(args.num_envs))
])
vec_env = pfrl.wrappers.VectorFrameStack(vec_env, 4)
return vec_env
sample_env = make_env(0, test=False)
n_actions = sample_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()
# Use the same hyper parameters as the Nature paper's
opt = optim.RMSprop(
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.n_step_return,
)
else:
rbuf = replay_buffers.ReplayBuffer(10**6, num_steps=args.n_step_return)
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 = 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=make_batch_env(test=True),
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_batch_with_evaluation(
agent=agent,
env=make_batch_env(test=False),
eval_env=make_batch_env(test=True),
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,
log_interval=1000,
)
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("--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")
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("--steps", type=int, default=8 * 10**7)
parser.add_argument("--update-steps", type=int, default=5)
parser.add_argument("--lr", type=float, default=7e-4)
parser.add_argument("--gamma",
type=float,
default=0.99,
help="discount factor")
parser.add_argument("--rmsprop-epsilon", type=float, default=1e-5)
parser.add_argument(
"--use-gae",
action="store_true",
default=False,
help="use generalized advantage estimation",
)
parser.add_argument("--tau",
type=float,
default=0.95,
help="gae parameter")
parser.add_argument("--alpha",
type=float,
default=0.99,
help="RMSprop optimizer alpha")
parser.add_argument("--eval-interval", type=int, default=10**6)
parser.add_argument("--eval-n-runs", type=int, default=10)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load", type=str, default="")
parser.add_argument("--max-grad-norm",
type=float,
default=40,
help="value loss coefficient")
parser.add_argument(
"--gpu",
"-g",
type=int,
default=-1,
help="GPU ID (negative value indicates CPU)",
)
parser.add_argument("--num-envs", type=int, default=1)
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.set_defaults(use_lstm=False)
args = parser.parse_args()
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.num_envs) + args.seed * args.num_envs
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
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(0, test=False)
obs_channel_size = sample_env.observation_space.low.shape[0]
n_actions = sample_env.action_space.n
model = nn.Sequential(
nn.Conv2d(obs_channel_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),
),
)
optimizer = pfrl.optimizers.RMSpropEpsInsideSqrt(
model.parameters(),
lr=args.lr,
eps=args.rmsprop_epsilon,
alpha=args.alpha,
)
agent = a2c.A2C(
model,
optimizer,
gamma=args.gamma,
gpu=args.gpu,
num_processes=args.num_envs,
update_steps=args.update_steps,
phi=phi,
use_gae=args.use_gae,
tau=args.tau,
max_grad_norm=args.max_grad_norm,
)
if args.load:
agent.load(args.load)
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,
)
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),
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,
log_interval=1000,
)
