def parse_arch(arch, n_actions):
if arch == "nature":
return nn.Sequential(
pnn.LargeAtariCNN(),
init_chainer_default(nn.Linear(512, n_actions)),
DiscreteActionValueHead(),
)
elif arch == "doubledqn":
# raise NotImplementedError("Single shared bias not implemented yet")
return nn.Sequential(
pnn.LargeAtariCNN(),
init_chainer_default(nn.Linear(512, n_actions, bias=False)),
SingleSharedBias(),
DiscreteActionValueHead(),
)
elif arch == "nips":
return nn.Sequential(
pnn.SmallAtariCNN(),
init_chainer_default(nn.Linear(256, n_actions)),
DiscreteActionValueHead(),
)
elif arch == "dueling":
return DuelingDQN(n_actions)
else:
raise RuntimeError("Not supported architecture: {}".format(arch))
def __init__(self):
super().__init__()
self.l1 = nn.Conv2d(1, 32, 8, stride=4)
self.l2 = nn.ReLU()
self.l3 = nn.Conv2d(32, 64, 4, stride=2)
self.l4 = nn.ReLU()
self.l5 = nn.Conv2d(64, 64, 3, stride=1)
self.l6 = nn.Flatten()
self.l7 = nn.ReLU()
self.r8 = nn.LSTM(input_size=3136, hidden_size=512)
self.l9 = init_chainer_default(torch.nn.Linear(512, n_actions))
self.l10 = DiscreteActionValueHead()
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 __init__(self, n_actions, max_episode_steps):
super().__init__()
self.embed = nn.Embedding(max_episode_steps + 1, 3136)
self.image2hidden = nn.Sequential(
nn.Conv2d(3, 32, 8, stride=4),
nn.ReLU(),
nn.Conv2d(32, 64, 4, stride=2),
nn.ReLU(),
nn.Conv2d(64, 64, 3, stride=1),
nn.Flatten(),
)
self.hidden2out = nn.Sequential(
nn.Linear(3136, 512),
nn.ReLU(),
nn.Linear(512, n_actions),
DiscreteActionValueHead(),
)
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 _test_abc(
self,
t_max,
use_lstm,
discrete=True,
episodic=True,
steps=100000,
require_success=True,
):
nproc = 8
def make_env(process_idx, test):
size = 2
return ABC(
size=size,
discrete=discrete,
episodic=episodic or test,
partially_observable=self.use_lstm,
deterministic=test,
)
sample_env = make_env(0, False)
action_space = sample_env.action_space
obs_space = sample_env.observation_space
replay_buffer = EpisodicReplayBuffer(10**4)
obs_size = obs_space.low.size
hidden_size = 20
if discrete:
n_actions = action_space.n
head = acer.ACERDiscreteActionHead(
pi=nn.Sequential(
nn.Linear(hidden_size, n_actions),
SoftmaxCategoricalHead(),
),
q=nn.Sequential(
nn.Linear(hidden_size, n_actions),
DiscreteActionValueHead(),
),
)
else:
action_size = action_space.low.size
head = acer.ACERContinuousActionHead(
pi=nn.Sequential(
nn.Linear(hidden_size, action_size * 2),
GaussianHeadWithDiagonalCovariance(),
),
v=nn.Sequential(nn.Linear(hidden_size, 1), ),
adv=nn.Sequential(
ConcatObsAndAction(),
nn.Linear(hidden_size + action_size, 1),
),
)
if use_lstm:
model = pfrl.nn.RecurrentSequential(
nn.Linear(obs_size, hidden_size),
nn.LeakyReLU(),
nn.LSTM(num_layers=1,
input_size=hidden_size,
hidden_size=hidden_size),
head,
)
else:
model = nn.Sequential(
nn.Linear(obs_size, hidden_size),
nn.LeakyReLU(),
head,
)
eps = 1e-8
opt = pfrl.optimizers.SharedRMSpropEpsInsideSqrt(model.parameters(),
lr=1e-3,
eps=eps,
alpha=0.99)
gamma = 0.5
beta = 1e-5
if self.n_times_replay == 0 and self.disable_online_update:
# At least one of them must be enabled
pytest.skip()
agent = acer.ACER(
model,
opt,
replay_buffer=replay_buffer,
t_max=t_max,
gamma=gamma,
beta=beta,
n_times_replay=self.n_times_replay,
act_deterministically=True,
disable_online_update=self.disable_online_update,
replay_start_size=100,
use_trust_region=self.use_trust_region,
recurrent=use_lstm,
)
max_episode_len = None if episodic else 2
with warnings.catch_warnings(record=True) as warns:
train_agent_async(
outdir=self.outdir,
processes=nproc,
make_env=make_env,
agent=agent,
steps=steps,
max_episode_len=max_episode_len,
eval_interval=500,
eval_n_steps=None,
eval_n_episodes=5,
successful_score=1,
)
assert len(warns) == 0, warns[0]
# The agent returned by train_agent_async is not guaranteed to be
# successful because parameters could be modified by other processes
# after success. Thus here the successful model is loaded explicitly.
if require_success:
agent.load(os.path.join(self.outdir, "successful"))
# Test
env = make_env(0, True)
n_test_runs = 5
eval_returns = run_evaluation_episodes(
env,
agent,
n_steps=None,
n_episodes=n_test_runs,
max_episode_len=max_episode_len,
)
successful_return = 1
if require_success:
n_succeeded = np.sum(np.asarray(eval_returns) >= successful_return)
assert n_succeeded == n_test_runs
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(
"--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="'DClawTurnFixed-v0'",
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=-1, 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(
"--max-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=64, 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."
)
parser.add_argument("--gamma", type=float, default=0.9)
parser.add_argument("--ddpg-training-steps", type=int, default=int(1e3))
parser.add_argument("--adversary-training-steps", type=int,default=int(1e3))
args = parser.parse_args()
logging.basicConfig(level=args.log_level)
args.outdir = './results'
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('DClawTurnFixed-v0')
# 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
q_func = nn.Sequential(
ConcatObsAndAction(),
nn.Linear(obs_size + action_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256,256),
nn.ReLU(),
nn.Linear(256, 1),
)
policy = nn.Sequential(
nn.Linear(obs_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256,256),
nn.ReLU(),
nn.Linear(256, action_size),
BoundByTanh(low=action_space.low, high=action_space.high),
DeterministicHead(),
)
ddpg_opt_a = torch.optim.Adam(policy.parameters())
ddpg_opt_c = torch.optim.Adam(q_func.parameters())
ddpg_rbuf = replay_buffers.ReplayBuffer(10 ** 6)
ddpg_explorer = explorers.AdditiveGaussian(
scale=0.1, low=action_space.low, high=action_space.high
)
def ddpg_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
ddpg_agent = DDPG(
policy,
q_func,
ddpg_opt_a,
ddpg_opt_c,
ddpg_rbuf,
gamma=args.gamma,
explorer=ddpg_explorer,
replay_start_size=args.replay_start_size,
target_update_method="soft",
target_update_interval=1,
update_interval=1,
soft_update_tau=5e-3,
n_times_update=1,
gpu=args.gpu,
minibatch_size=args.batch_size,
burnin_action_func=ddpg_burnin_action_func,
)
def adversary_random_func():
return np.random.randint(0,9)
# adversary_q = Critic(obs_size, 1, hidden_size=adversary_hidden_size)
# adversary_action_space = gym.spaces.discrete.Discrete(9)
# adversary_q = q_functions.FCQuadraticStateQFunction(
# obs_size, 1, n_hidden_channels = 256, n_hidden_layers = 2,action_space = adversary_action_space
# )
adversary_q = nn.Sequential(
nn.Linear(obs_size, 256),
nn.Linear(256,256),
nn.Linear(256,256),
nn.Linear(256,1),
DiscreteActionValueHead(),
)
adversary_optimizer = torch.optim.Adam(adversary_q.parameters(), lr=1e-3)
adversary_rbuf_capacity = int(1e6)
adversary_rbuf = replay_buffers.ReplayBuffer(adversary_rbuf_capacity)
adversary_explorer = explorers.LinearDecayEpsilonGreedy(
1.0, 0.1, 10**4, adversary_random_func
)
adversary_agent = DQN(
adversary_q,
adversary_optimizer,
adversary_rbuf,
gpu=args.gpu,
gamma=args.gamma,
explorer=adversary_explorer,
replay_start_size=args.replay_start_size,
target_update_interval=1,
minibatch_size=args.batch_size,
target_update_method='soft',
soft_update_tau=5e-3
)
logger = logging.getLogger(__name__)
eval_env = make_env(test=True)
evaluator = Evaluator(
agent=ddpg_agent,
n_steps=None,
n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
max_episode_len=timestep_limit,
env=eval_env,
step_offset=0,
save_best_so_far_agent=True,
use_tensorboard=True,
logger=logger,
)
episode_reward = 0
ddpg_episode_idx = 0
adversary_episode_idx = 0
# o_0, r_0
current_state = env.reset()
t = 0
ddpg_t = 0
adversary_t = 0
episode_len = 0
try:
while t < args.max_steps:
for i in range(args.ddpg_training_steps):
t += 1
ddpg_t += 1
ddpg_action = ddpg_agent.act(current_state)
adversary_action = adversary_agent.act(current_state)
ddpg_action[adversary_action] = 0
next_state, reward, done, info = env.step(ddpg_action)
episode_reward += reward
episode_len += 1
reset = episode_len == timestep_limit or info.get("needs_reset", False)
ddpg_agent.observe(next_state, reward, done, reset)
current_state = next_state
if done or reset or t == args.max_steps:
logger.info(
"ddpg phase: outdir:%s step:%s episode:%s R:%s",
args.outdir,
ddpg_t,
ddpg_episode_idx,
episode_reward,
)
logger.info("statistics:%s", ddpg_agent.get_statistics())
if evaluator is not None:
evaluator.evaluate_if_necessary(t=t, episodes=ddpg_episode_idx + 1)
if t == args.max_steps:
break
episode_reward = 0
ddpg_episode_idx += 1
episode_len = 0
current_state = env.reset()
episode_reward = 0
episode_len = 0
current_state = env.reset()
print("start adversary training ")
for i in range(args.adversary_training_steps):
t += 1
adversary_t += 1
ddpg_action = ddpg_agent.act(current_state)
adversary_action = adversary_agent.act(current_state)
ddpg_action[adversary_action] = 0
next_state, reward, done, info = env.step(ddpg_action)
reward = -reward
episode_len += 1
reset = episode_len == timestep_limit or info.get("needs_reset", False)
adversary_agent.observe(next_state, reward, done, reset)
current_state = next_state
if done or reset or t == args.max_steps:
if t == args.max_steps:
break
episode_reward = 0
adversary_episode_idx += 1
episode_len = 0
current_state = env.reset()
except (Exception, KeyboardInterrupt):
# Save the current model before being killed
save_agent(ddpg_agent, t, args.outdir, logger, suffix="_ddpg_except")
save_agent(adversary_agent, t, args.outdir, logger, suffix="_adversary_except" )
raise
# Save the final model
save_agent(ddpg_agent, t, args.outdir, logger, suffix="_ddpg_finish")
save_agent(adversary_agent, t, args.outdir, logger, suffix="_adversary_finish" )
# if args.demo:
# eval_env.render()
# eval_stats = experiments.eval_performance(
# env=eval_env,
# agent=ddpg_agent,
# n_steps=None,
# n_episodes=args.eval_n_runbase_envs,
# max_episode_len=timestep_limit,
# )
# print(
# "n_runs: {} mean: {} median: {} stdev {}".format(
# args.eval_n_runs,
# eval_stats["mean"],
# eval_stats["median"],
# eval_stats["stdev"],
# )
# )
# else:
# experiments.train_agent_with_evaluation(
# agent=ddpg_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,
# )
print("finish")
self.bandit2 = 0.8
else:
self.bandit2 = 0.2
self.bandit1 = 0.8
if ns.two:
env = TwoArmedBanditEnv()
else:
env = MultiBanditEnv()
n_actions = env.action_space.n
q_func = pfrl.nn.RecurrentSequential(
nn.LSTM(input_size=1, hidden_size=512),
nn.Linear(512, n_actions),
DiscreteActionValueHead(),
)
replay_buffer = pfrl.replay_buffers.EpisodicReplayBuffer(10**6)
explorer = explorers.LinearDecayEpsilonGreedy(
1.0,
.01,
1000000,
lambda: numpy.random.randint(n_actions),
)
def phi(x):
return numpy.asarray(x, dtype=numpy.float32) / 1.0
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--env",
type=str,
default="BreakoutNoFrameskip-v4",
help="OpenAI Atari domain to perform algorithm on.",
)
parser.add_argument(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
parser.add_argument("--gpu",
type=int,
default=0,
help="GPU to use, set to -1 if no GPU.")
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load", type=str, default=None)
parser.add_argument(
"--final-exploration-frames",
type=int,
default=10**6,
help="Timesteps after which we stop " + "annealing exploration rate",
)
parser.add_argument(
"--final-epsilon",
type=float,
default=0.01,
help="Final value of epsilon during training.",
)
parser.add_argument(
"--eval-epsilon",
type=float,
default=0.001,
help="Exploration epsilon used during eval episodes.",
)
parser.add_argument(
"--steps",
type=int,
default=5 * 10**7,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--max-frames",
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help="Maximum number of frames for each episode.",
)
parser.add_argument(
"--replay-start-size",
type=int,
default=5 * 10**4,
help="Minimum replay buffer size before " +
"performing gradient updates.",
)
parser.add_argument(
"--target-update-interval",
type=int,
default=3 * 10**4,
help="Frequency (in timesteps) at which " +
"the target network is updated.",
)
parser.add_argument("--demo-n-episodes", type=int, default=30)
parser.add_argument("--eval-n-steps", type=int, default=125000)
parser.add_argument(
"--eval-interval",
type=int,
default=250000,
help="Frequency (in timesteps) of evaluation phase.",
)
parser.add_argument(
"--update-interval",
type=int,
default=4,
help="Frequency (in timesteps) of network updates.",
)
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument("--lr",
type=float,
default=2.5e-4,
help="Learning rate.")
parser.add_argument(
"--recurrent",
action="store_true",
default=False,
help="Use a recurrent model. See the code for the model definition.",
)
parser.add_argument(
"--flicker",
action="store_true",
default=False,
help=("Use so-called flickering Atari, where each"
" screen is blacked out with probability 0.5."),
)
parser.add_argument(
"--no-frame-stack",
action="store_true",
default=False,
help=
("Disable frame stacking so that the agent can only see the current screen."
),
)
parser.add_argument(
"--episodic-update-len",
type=int,
default=10,
help="Maximum length of sequences for updating recurrent models",
)
parser.add_argument(
"--batch-size",
type=int,
default=32,
help=("Number of transitions (in a non-recurrent case)"
" or sequences (in a recurrent case) used for an"
" update."),
)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test,
flicker=args.flicker,
frame_stack=not args.no_frame_stack,
)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, args.eval_epsilon)
if args.monitor:
env = gym.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training")
if args.render:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
print("Observation space", env.observation_space)
print("Action space", env.action_space)
n_frames = env.observation_space.shape[0]
n_actions = env.action_space.n
if args.recurrent:
# Q-network with LSTM
q_func = pfrl.nn.RecurrentSequential(
nn.Conv2d(n_frames, 32, 8, stride=4),
nn.ReLU(),
nn.Conv2d(32, 64, 4, stride=2),
nn.ReLU(),
nn.Conv2d(64, 64, 3, stride=1),
nn.Flatten(),
nn.ReLU(),
nn.LSTM(input_size=3136, hidden_size=512),
nn.Linear(512, n_actions),
DiscreteActionValueHead(),
)
# Replay buffer that stores whole episodes
rbuf = replay_buffers.EpisodicReplayBuffer(10**6)
else:
# Q-network without LSTM
q_func = nn.Sequential(
nn.Conv2d(n_frames, 32, 8, stride=4),
nn.ReLU(),
nn.Conv2d(32, 64, 4, stride=2),
nn.ReLU(),
nn.Conv2d(64, 64, 3, stride=1),
nn.Flatten(),
nn.Linear(3136, 512),
nn.ReLU(),
nn.Linear(512, n_actions),
DiscreteActionValueHead(),
)
# Replay buffer that stores transitions separately
rbuf = replay_buffers.ReplayBuffer(10**6)
explorer = explorers.LinearDecayEpsilonGreedy(
1.0,
args.final_epsilon,
args.final_exploration_frames,
lambda: np.random.randint(n_actions),
)
opt = torch.optim.Adam(q_func.parameters(), lr=1e-4, eps=1e-4)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
agent = pfrl.agents.DoubleDQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
batch_accumulator="mean",
phi=phi,
minibatch_size=args.batch_size,
episodic_update_len=args.episodic_update_len,
recurrent=args.recurrent,
)
if args.load:
agent.load(args.load)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.demo_n_episodes,
)
print("n_runs: {} mean: {} median: {} stdev {}".format(
args.demo_n_episodes,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=args.eval_n_steps,
eval_n_episodes=None,
eval_interval=args.eval_interval,
outdir=args.outdir,
eval_env=eval_env,
)