def _test_actor_learner_training(self, gpu, steps=100000, require_success=True):
logging.basicConfig(level=logging.DEBUG)
test_env, successful_return = self.make_env_and_successful_return(test=True)
agent = self.make_agent(test_env, gpu)
# cumulative_steps init to 0
assert agent.cumulative_steps == 0
def make_env(process_idx, test):
env, _ = self.make_env_and_successful_return(test=test)
return env
step_hook = mock.Mock()
optimizer_step_hook = mock.Mock()
# Train
if steps > 0:
(
make_actor,
learner,
poller,
exception_event,
) = agent.setup_actor_learner_training(
n_actors=2,
step_hooks=[step_hook],
optimizer_step_hooks=[optimizer_step_hook],
)
poller.start()
learner.start()
train_agent_async(
processes=2,
steps=steps,
outdir=self.tmpdir,
eval_interval=200,
eval_n_steps=None,
eval_n_episodes=5,
successful_score=successful_return,
make_env=make_env,
make_agent=make_actor,
stop_event=learner.stop_event,
exception_event=exception_event,
)
learner.stop()
learner.join()
poller.stop()
poller.join()
# Test
# Because in actor-learner traininig the model can be updated between
# evaluation and saving, it is difficult to guarantee the learned
# model successfully passes the test.
# Thus we only check if the training was successful.
# As the test can finish before running all the steps,
# we can only test the range
assert agent.cumulative_steps > 0
assert agent.cumulative_steps <= steps + 1
# Unlike the non-actor-learner cases, the step_hooks and
# optimizer_step_hooks are only called when the update happens
# when we do a fast test, the update may not be triggered due to
# limited amount of experience, the call_count can be 0 in such case
assert step_hook.call_count >= 0
assert step_hook.call_count <= steps / agent.update_interval
assert optimizer_step_hook.call_count == step_hook.call_count
for i, call in enumerate(step_hook.call_args_list):
args, kwargs = call
assert args[0] is None
assert args[1] is agent
assert args[2] == (i + 1) * agent.update_interval
for i, call in enumerate(optimizer_step_hook.call_args_list):
args, kwargs = call
assert args[0] is None
assert args[1] is agent
assert args[2] == i + 1
successful_path = os.path.join(self.tmpdir, "successful")
finished_path = os.path.join(self.tmpdir, "{}_finish".format(steps))
if require_success:
assert os.path.exists(successful_path)
else:
assert os.path.exists(successful_path) or os.path.exists(finished_path)
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("--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_pretrained:
raise Exception("Pretrained models are currently unsupported.")
if args.load:
agent.load(args.load)
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,
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("processes", type=int)
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("--replay-start-size", type=int, default=10000)
parser.add_argument("--n-times-replay", type=int, default=4)
parser.add_argument("--beta", type=float, default=1e-2)
parser.add_argument("--profile", action="store_true")
parser.add_argument("--steps", type=int, default=10**7)
parser.add_argument(
"--max-frames",
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help="Maximum number of frames for each episode.",
)
parser.add_argument("--lr", type=float, default=7e-4)
parser.add_argument("--eval-interval", type=int, default=10**5)
parser.add_argument("--eval-n-runs", type=int, default=10)
parser.add_argument("--use-lstm", action="store_true")
parser.add_argument("--demo", action="store_true", default=False)
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."
),
)
parser.set_defaults(use_lstm=False)
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))
n_actions = gym.make(args.env).action_space.n
input_to_hidden = nn.Sequential(
nn.Conv2d(4, 16, 8, stride=4),
nn.ReLU(),
nn.Conv2d(16, 32, 4, stride=2),
nn.ReLU(),
nn.Flatten(),
nn.Linear(2592, 256),
nn.ReLU(),
)
head = acer.ACERDiscreteActionHead(
pi=nn.Sequential(
nn.Linear(256, n_actions),
SoftmaxCategoricalHead(),
),
q=nn.Sequential(
nn.Linear(256, n_actions),
DiscreteActionValueHead(),
),
)
if args.use_lstm:
model = pfrl.nn.RecurrentSequential(
input_to_hidden,
nn.LSTM(num_layers=1, input_size=256, hidden_size=256),
head,
)
else:
model = nn.Sequential(input_to_hidden, head)
model.apply(pfrl.initializers.init_chainer_default)
opt = pfrl.optimizers.SharedRMSpropEpsInsideSqrt(model.parameters(),
lr=args.lr,
eps=4e-3,
alpha=0.99)
replay_buffer = EpisodicReplayBuffer(10**6 // args.processes)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
agent = acer.ACER(
model,
opt,
t_max=args.t_max,
gamma=0.99,
replay_buffer=replay_buffer,
n_times_replay=args.n_times_replay,
replay_start_size=args.replay_start_size,
beta=args.beta,
phi=phi,
max_grad_norm=40,
recurrent=args.use_lstm,
)
if args.load:
agent.load(args.load)
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
if args.demo:
env = make_env(0, True)
eval_stats = experiments.eval_performance(env=env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs)
print("n_runs: {} mean: {} median: {} stdev {}".format(
args.eval_n_runs,
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=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
global_step_hooks=[lr_decay_hook],
save_best_so_far_agent=False,
)
def main():
import logging
logging.basicConfig(level=logging.INFO)
parser = argparse.ArgumentParser()
parser.add_argument(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument("--env", type=str, default="Pendulum-v0")
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 32)")
parser.add_argument("--gpu", type=int, default=0)
parser.add_argument("--final-exploration-steps", type=int, default=10**4)
parser.add_argument("--start-epsilon", type=float, default=1.0)
parser.add_argument("--end-epsilon", type=float, default=0.1)
parser.add_argument("--noisy-net-sigma", type=float, default=None)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load", type=str, default=None)
parser.add_argument("--steps", type=int, default=10**5)
parser.add_argument("--prioritized-replay", action="store_true")
parser.add_argument("--replay-start-size", type=int, default=1000)
parser.add_argument("--target-update-interval", type=int, default=10**2)
parser.add_argument("--target-update-method", type=str, default="hard")
parser.add_argument("--soft-update-tau", type=float, default=1e-2)
parser.add_argument("--update-interval", type=int, default=1)
parser.add_argument("--eval-n-runs", type=int, default=100)
parser.add_argument("--eval-interval", type=int, default=10**4)
parser.add_argument("--n-hidden-channels", type=int, default=100)
parser.add_argument("--n-hidden-layers", type=int, default=2)
parser.add_argument("--gamma", type=float, default=0.99)
parser.add_argument("--minibatch-size", type=int, default=None)
parser.add_argument("--render-train", action="store_true")
parser.add_argument("--render-eval", action="store_true")
parser.add_argument("--monitor", action="store_true")
parser.add_argument("--reward-scale-factor", type=float, default=1e-3)
parser.add_argument(
"--actor-learner",
action="store_true",
help="Enable asynchronous sampling with asynchronous actor(s)",
) # NOQA
parser.add_argument(
"--num-envs",
type=int,
default=1,
help=("The number of environments for sampling (only effective with"
" --actor-learner enabled)"),
) # NOQA
args = parser.parse_args()
# Set a random seed used in PFRL
utils.set_random_seed(args.seed)
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print("Output files are saved in {}".format(args.outdir))
# Set different random seeds for different subprocesses.
# If seed=0 and processes=4, subprocess seeds are [0, 1, 2, 3].
# If seed=1 and processes=4, subprocess seeds are [4, 5, 6, 7].
process_seeds = np.arange(args.num_envs) + args.seed * args.num_envs
assert process_seeds.max() < 2**32
def clip_action_filter(a):
return np.clip(a, action_space.low, action_space.high)
def make_env(idx=0, test=False):
env = gym.make(args.env)
# Use different random seeds for train and test envs
process_seed = int(process_seeds[idx])
env_seed = 2**32 - 1 - process_seed if test else process_seed
utils.set_random_seed(env_seed)
# Cast observations to float32 because our model uses float32
env = pfrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = pfrl.wrappers.Monitor(env, args.outdir)
if isinstance(env.action_space, spaces.Box):
utils.env_modifiers.make_action_filtered(env, clip_action_filter)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = pfrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if (args.render_eval and test) or (args.render_train and not test):
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.max_episode_steps
obs_space = env.observation_space
obs_size = obs_space.low.size
action_space = env.action_space
if isinstance(action_space, spaces.Box):
action_size = action_space.low.size
# Use NAF to apply DQN to continuous action spaces
q_func = q_functions.FCQuadraticStateQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
action_space=action_space,
)
# Use the Ornstein-Uhlenbeck process for exploration
ou_sigma = (action_space.high - action_space.low) * 0.2
explorer = explorers.AdditiveOU(sigma=ou_sigma)
else:
n_actions = action_space.n
q_func = q_functions.FCStateQFunctionWithDiscreteAction(
obs_size,
n_actions,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
)
# Use epsilon-greedy for exploration
explorer = explorers.LinearDecayEpsilonGreedy(
args.start_epsilon,
args.end_epsilon,
args.final_exploration_steps,
action_space.sample,
)
if args.noisy_net_sigma is not None:
pnn.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
opt = optim.Adam(q_func.parameters())
rbuf_capacity = 5 * 10**5
if args.minibatch_size is None:
args.minibatch_size = 32
if args.prioritized_replay:
betasteps = (args.steps -
args.replay_start_size) // args.update_interval
rbuf = replay_buffers.PrioritizedReplayBuffer(rbuf_capacity,
betasteps=betasteps)
else:
rbuf = replay_buffers.ReplayBuffer(rbuf_capacity)
agent = DQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=args.gamma,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
minibatch_size=args.minibatch_size,
target_update_method=args.target_update_method,
soft_update_tau=args.soft_update_tau,
)
if args.load:
agent.load(args.load)
eval_env = make_env(test=True)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
max_episode_len=timestep_limit,
)
print("n_runs: {} mean: {} median: {} stdev {}".format(
args.eval_n_runs,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
elif not args.actor_learner:
print(
"WARNING: Since https://github.com/pfnet/pfrl/pull/112 we have started"
" setting `eval_during_episode=True` in this script, which affects the"
" timings of evaluation phases.")
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
eval_env=eval_env,
train_max_episode_len=timestep_limit,
eval_during_episode=True,
)
else:
# using impala mode when given num of envs
# When we use multiple envs, it is critical to ensure each env
# can occupy a CPU core to get the best performance.
# Therefore, we need to prevent potential CPU over-provision caused by
# multi-threading in Openmp and Numpy.
# Disable the multi-threading on Openmp and Numpy.
os.environ["OMP_NUM_THREADS"] = "1" # NOQA
(
make_actor,
learner,
poller,
exception_event,
) = agent.setup_actor_learner_training(args.num_envs)
poller.start()
learner.start()
experiments.train_agent_async(
processes=args.num_envs,
make_agent=make_actor,
make_env=make_env,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
stop_event=learner.stop_event,
exception_event=exception_event,
)
poller.stop()
learner.stop()
poller.join()
learner.join()
def _test_actor_learner_training(self, gpu, steps=100000, require_success=True):
logging.basicConfig(level=logging.DEBUG)
test_env, successful_return = self.make_env_and_successful_return(test=True)
agent = self.make_agent(test_env, gpu)
# cumulative_steps init to 0
assert agent.cumulative_steps == 0
def make_env(process_idx, test):
env, _ = self.make_env_and_successful_return(test=test)
return env
# Train
if steps > 0:
(
make_actor,
learner,
poller,
exception_event,
) = agent.setup_actor_learner_training(n_actors=2)
poller.start()
learner.start()
train_agent_async(
processes=2,
steps=steps,
outdir=self.tmpdir,
eval_interval=200,
eval_n_steps=None,
eval_n_episodes=5,
successful_score=successful_return,
make_env=make_env,
make_agent=make_actor,
stop_event=learner.stop_event,
exception_event=exception_event,
)
learner.stop()
learner.join()
poller.stop()
poller.join()
# Test
# Because in actor-learner traininig the model can be updated between
# evaluation and saving, it is difficult to guarantee the learned
# model successfully passes the test.
# Thus we only check if the training was successful.
# As the test can finish before running all the steps,
# we can only test the range
assert agent.cumulative_steps > 0
assert agent.cumulative_steps <= steps + 1
successful_path = os.path.join(self.tmpdir, "successful")
finished_path = os.path.join(self.tmpdir, "{}_finish".format(steps))
if require_success:
assert os.path.exists(successful_path)
else:
assert os.path.exists(successful_path) or os.path.exists(finished_path)
