def main():
import logging
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=-1)
parser.add_argument('--env', type=str, default='Hopper-v2')
parser.add_argument('--num-envs', type=int, default=1)
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**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('--logger-level', type=int, default=logging.DEBUG)
parser.add_argument('--monitor', action='store_true')
parser.add_argument('--window-size', type=int, default=100)
parser.add_argument('--update-interval', type=int, default=2048)
parser.add_argument('--log-interval', type=int, default=1000)
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 ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
# 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 = chainerrl.wrappers.CastObservationToFloat32(env)
# if args.monitor:
# env = chainerrl.wrappers.Monitor(env, args.outdir)
# if not test:
# # Scale rewards (and thus returns) to a reasonable range so that
# # training is easier
# env = chainerrl.wrappers.ScaleReward(env, args.reward_scale_factor)
# if args.render:
# env = chainerrl.wrappers.Render(env)
# return env
def make_env(test):
env = gym.make(
"DaktyPushingSimulationEnv-v0",
level=5,
simulation_backend="mujoco",
control_frequency_in_hertz=100,
state_space_components_to_be_used=None,
alternate_env_object=None,
discretization_factor_torque_control_space=None,
model_as_function_for_pixel_to_latent_space_parsing=(None, None))
# print('\n############\n', env, '\n############\n')
env.unwrapped.finger.set_resolution_quality('low')
# print('\n############\n', env, '\n############\n')
env = gym.wrappers.TimeLimit(env)
# print('\n############\n', env, '\n############\n')
# 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)
process_seed = 420
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 = chainerrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = chainerrl.wrappers.Monitor(env, args.outdir)
if args.render and not test:
env = chainerrl.wrappers.Render(env)
return env
def make_batch_env(test):
return chainerrl.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 = make_env(0)
timestep_limit = sample_env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_space = sample_env.observation_space
action_space = sample_env.action_space
print('\n\n------------------- obs_space: ', obs_space.shape, '\n\n\n')
# Normalize observations based on their empirical mean and variance
obs_normalizer = chainerrl.links.EmpiricalNormalization(obs_space.low.size,
clip_threshold=5)
winit_last = chainer.initializers.LeCunNormal(1e-2)
action_size = action_space.low.size
policy = chainer.Sequential(
L.Linear(None, 64),
F.tanh,
L.Linear(None, 64),
F.tanh,
L.Linear(None, action_size, initialW=winit_last),
chainerrl.policies.GaussianHeadWithStateIndependentCovariance(
action_size=action_size,
var_type='diagonal',
var_func=lambda x: F.exp(2 * x), # Parameterize log std
var_param_init=0, # log std = 0 => std = 1
))
vf = chainer.Sequential(
concat_obs_and_action,
L.Linear(None, 64),
F.tanh,
L.Linear(None, 64),
F.tanh,
L.Linear(None, 1),
)
# Combine a policy and a value function into a single model
model = chainerrl.links.Branched(policy, vf)
opt = chainer.optimizers.Adam(alpha=args.lr, eps=1e-5)
opt.setup(model)
if args.weight_decay > 0:
opt.add_hook(NonbiasWeightDecay(args.weight_decay))
agent = PPO(
model,
opt,
obs_normalizer=obs_normalizer,
gpu=args.gpu,
update_interval=args.update_interval,
minibatch_size=args.batchsize,
epochs=args.epochs,
clip_eps_vf=None,
entropy_coef=args.entropy_coef,
standardize_advantages=args.standardize_advantages,
)
if args.load:
agent.load(args.load)
if args.demo:
env = make_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:
env = make_env(False)
n_episodes = 10000
# pbar = tqdm(total=n_episodes)
max_episode_len = 1000
for i in range(1, n_episodes + 1):
# pbar.update(1)
obs = env.reset()
# print('obs inital..............', obs.shape)
reward = 0
done = False
R = 0 # return (sum of rewards)
t = 0 # time step
# pbar = tqdm(total=max_episode_len)
while not done and t < max_episode_len:
# pbar.update(1)
# Uncomment to watch the behaviour
# env.render()
action = agent.act_and_train(obs, reward)
# print('action..................', action)
obs, reward, done, _ = env.step(action)
# print('obs.....................', obs)
# print('reward..................', reward)
R += reward
t += 1
if i % 10 == 0:
print('episode:', i, 'R:', R, 'statistics:',
agent.get_statistics())
agent.stop_episode_and_train(obs, reward, done)
print('Finished.')
# Linearly decay the learning rate to zero
def lr_setter(env, agent, value):
agent.optimizer.alpha = value
lr_decay_hook = experiments.LinearInterpolationHook(
args.steps, args.lr, 0, lr_setter)
experiments.train_agent_batch_with_evaluation(
agent=agent,
env=make_env(False),
eval_env=make_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,
return_window_size=args.window_size,
max_episode_len=timestep_limit,
save_best_so_far_agent=False,
step_hooks=[
lr_decay_hook,
],
)
def main():
import logging
parser = argparse.ArgumentParser()
parser.add_argument('processes', type=int)
parser.add_argument('--env', type=str, default='CartPole-v0')
parser.add_argument('--seed', type=int, default=0,
help='Random seed [0, 2 ** 32)')
parser.add_argument('--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=50)
parser.add_argument('--n-times-replay', type=int, default=4)
parser.add_argument('--n-hidden-channels', type=int, default=100)
parser.add_argument('--n-hidden-layers', type=int, default=2)
parser.add_argument('--replay-capacity', type=int, default=5000)
parser.add_argument('--replay-start-size', type=int, default=10 ** 3)
parser.add_argument('--disable-online-update', action='store_true')
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('--eval-interval', type=int, default=10 ** 5)
parser.add_argument('--eval-n-runs', type=int, default=10)
parser.add_argument('--reward-scale-factor', type=float, default=1e-2)
parser.add_argument('--rmsprop-epsilon', type=float, default=1e-2)
parser.add_argument('--render', action='store_true', default=False)
parser.add_argument('--lr', type=float, default=7e-4)
parser.add_argument('--demo', action='store_true', default=False)
parser.add_argument('--load', type=str, default='')
parser.add_argument('--logger-level', type=int, default=logging.DEBUG)
parser.add_argument('--monitor', action='store_true')
parser.add_argument('--truncation-threshold', type=float, default=5)
parser.add_argument('--trust-region-delta', type=float, default=0.1)
args = parser.parse_args()
logging.basicConfig(level=args.logger_level)
# Set a random seed used in ChainerRL.
# If you use more than one processes, the results will be no longer
# deterministic even with the same random seed.
misc.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 ** 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)
if args.monitor and process_idx == 0:
env = gym.wrappers.Monitor(env, args.outdir)
# Scale rewards observed by agents
if not test:
misc.env_modifiers.make_reward_filtered(
env, lambda x: x * args.reward_scale_factor)
if args.render and process_idx == 0 and not test:
misc.env_modifiers.make_rendered(env)
return env
sample_env = gym.make(args.env)
timestep_limit = sample_env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_space = sample_env.observation_space
action_space = sample_env.action_space
if isinstance(action_space, spaces.Box):
model = acer.ACERSDNSeparateModel(
pi=policies.FCGaussianPolicy(
obs_space.low.size, action_space.low.size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
bound_mean=True,
min_action=action_space.low,
max_action=action_space.high),
v=v_functions.FCVFunction(
obs_space.low.size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers),
adv=q_functions.FCSAQFunction(
obs_space.low.size, action_space.low.size,
n_hidden_channels=args.n_hidden_channels // 4,
n_hidden_layers=args.n_hidden_layers),
)
else:
model = acer.ACERSeparateModel(
pi=links.Sequence(
L.Linear(obs_space.low.size, args.n_hidden_channels),
F.relu,
L.Linear(args.n_hidden_channels, action_space.n,
initialW=LeCunNormal(1e-3)),
SoftmaxDistribution),
q=links.Sequence(
L.Linear(obs_space.low.size, args.n_hidden_channels),
F.relu,
L.Linear(args.n_hidden_channels, action_space.n,
initialW=LeCunNormal(1e-3)),
DiscreteActionValue),
)
opt = rmsprop_async.RMSpropAsync(
lr=args.lr, eps=args.rmsprop_epsilon, alpha=0.99)
opt.setup(model)
opt.add_hook(chainer.optimizer.GradientClipping(40))
replay_buffer = EpisodicReplayBuffer(args.replay_capacity)
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,
disable_online_update=args.disable_online_update,
use_trust_region=True,
trust_region_delta=args.trust_region_delta,
truncation_threshold=args.truncation_threshold,
beta=args.beta, phi=phi)
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_runs=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_async(
agent=agent,
outdir=args.outdir,
processes=args.processes,
make_env=make_env,
profile=args.profile,
steps=args.steps,
eval_n_runs=args.eval_n_runs,
eval_interval=args.eval_interval,
max_episode_len=timestep_limit)
def main():
import logging
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--env', type=str, default='Hopper-v2')
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='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**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('--logger-level', type=int, default=logging.DEBUG)
parser.add_argument('--monitor', action='store_true')
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 ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
args.outdir = experiments.prepare_output_dir(args, args.outdir)
def make_env(test):
env = gym.make(args.env)
# Use different random seeds for train and test envs
env_seed = 2**32 - 1 - args.seed if test else args.seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = chainerrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = chainerrl.wrappers.Monitor(env, args.outdir)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = chainerrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if args.render:
env = chainerrl.wrappers.Render(env)
return env
sample_env = gym.make(args.env)
timestep_limit = sample_env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_space = sample_env.observation_space
action_space = sample_env.action_space
# Normalize observations based on their empirical mean and variance
obs_normalizer = chainerrl.links.EmpiricalNormalization(obs_space.low.size,
clip_threshold=5)
# Switch policy types accordingly to action space types
if args.arch == 'FFSoftmax':
model = A3CFFSoftmax(obs_space.low.size, action_space.n)
elif args.arch == 'FFMellowmax':
model = A3CFFMellowmax(obs_space.low.size, action_space.n)
elif args.arch == 'FFGaussian':
model = A3CFFGaussian(obs_space.low.size,
action_space,
bound_mean=args.bound_mean)
opt = chainer.optimizers.Adam(alpha=args.lr, eps=1e-5)
opt.setup(model)
if args.weight_decay > 0:
opt.add_hook(NonbiasWeightDecay(args.weight_decay))
agent = PPO(
model,
opt,
obs_normalizer=obs_normalizer,
gpu=args.gpu,
update_interval=args.update_interval,
minibatch_size=args.batchsize,
epochs=args.epochs,
clip_eps_vf=None,
entropy_coef=args.entropy_coef,
standardize_advantages=args.standardize_advantages,
)
if args.load:
agent.load(args.load)
if args.demo:
env = make_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:
# Linearly decay the learning rate to zero
def lr_setter(env, agent, value):
agent.optimizer.alpha = value
lr_decay_hook = experiments.LinearInterpolationHook(
args.steps, args.lr, 0, lr_setter)
# Linearly decay the clipping parameter to zero
def clip_eps_setter(env, agent, value):
agent.clip_eps = max(value, 1e-8)
clip_eps_decay_hook = experiments.LinearInterpolationHook(
args.steps, 0.2, 0, clip_eps_setter)
experiments.train_agent_with_evaluation(
agent=agent,
env=make_env(False),
eval_env=make_env(True),
outdir=args.outdir,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
train_max_episode_len=timestep_limit,
save_best_so_far_agent=False,
step_hooks=[
lr_decay_hook,
clip_eps_decay_hook,
],
)
def main():
import logging
logging.basicConfig(level=logging.DEBUG)
parser = argparse.ArgumentParser()
parser.add_argument('--outdir',
type=str,
default='results',
help='Directory path to save output files.'
' If it does not exist, it will be created.')
parser.add_argument('--env', type=str, default='CartPole-v1')
parser.add_argument('--seed', type=int, default=0)
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--final-exploration-steps', type=int, default=1000)
parser.add_argument('--start-epsilon', type=float, default=1.0)
parser.add_argument('--end-epsilon', type=float, default=0.1)
parser.add_argument('--demo', action='store_true', default=False)
parser.add_argument('--load', type=str, default=None)
parser.add_argument('--steps', type=int, default=10**8)
parser.add_argument('--prioritized-replay', action='store_true')
parser.add_argument('--episodic-replay', action='store_true')
parser.add_argument('--replay-start-size', type=int, default=50)
parser.add_argument('--target-update-interval', type=int, default=100)
parser.add_argument('--target-update-method', type=str, default='hard')
parser.add_argument('--soft-update-tau', type=float, default=1e-2)
parser.add_argument('--update-interval', type=int, default=1)
parser.add_argument('--eval-n-runs', type=int, default=100)
parser.add_argument('--eval-interval', type=int, default=1000)
parser.add_argument('--n-hidden-channels', type=int, default=12)
parser.add_argument('--n-hidden-layers', type=int, default=3)
parser.add_argument('--gamma', type=float, default=0.95)
parser.add_argument('--minibatch-size', type=int, default=None)
parser.add_argument('--render-train', action='store_true')
parser.add_argument('--render-eval', action='store_true')
parser.add_argument('--monitor', action='store_true')
parser.add_argument('--reward-scale-factor', type=float, default=1.0)
args = parser.parse_args()
# Set a random seed used in ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print('Output files are saved in {}'.format(args.outdir))
def make_env(test):
env = gym.make(args.env)
env_seed = 2**32 - 1 - args.seed if test else args.seed
env.seed(env_seed)
if args.monitor:
env = gym.wrappers.Monitor(env, args.outdir)
if not test:
misc.env_modifiers.make_reward_filtered(
env, lambda x: x * args.reward_scale_factor)
if ((args.render_eval and test) or (args.render_train and not test)):
misc.env_modifiers.make_rendered(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_size = env.observation_space.low.size
action_space = env.action_space
n_atoms = 51
v_max = 500
v_min = 0
n_actions = action_space.n
q_func = q_functions.DistributionalFCStateQFunctionWithDiscreteAction(
obs_size,
n_actions,
n_atoms,
v_min,
v_max,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers)
# Use epsilon-greedy for exploration
explorer = explorers.LinearDecayEpsilonGreedy(args.start_epsilon,
args.end_epsilon,
args.final_exploration_steps,
action_space.sample)
opt = optimizers.Adam(1e-3)
opt.setup(q_func)
rbuf_capacity = 50000 # 5 * 10 ** 5
if args.episodic_replay:
if args.minibatch_size is None:
args.minibatch_size = 4
if args.prioritized_replay:
betasteps = (args.steps - args.replay_start_size) \
// args.update_interval
rbuf = replay_buffer.PrioritizedEpisodicReplayBuffer(
rbuf_capacity, betasteps=betasteps)
else:
rbuf = replay_buffer.EpisodicReplayBuffer(rbuf_capacity)
else:
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_buffer.PrioritizedReplayBuffer(rbuf_capacity,
betasteps=betasteps)
else:
rbuf = replay_buffer.ReplayBuffer(rbuf_capacity)
def phi(obs):
return obs.astype(np.float32)
agent = chainerrl.agents.CategoricalDQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=args.gamma,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
phi=phi,
minibatch_size=args.minibatch_size,
target_update_method=args.target_update_method,
soft_update_tau=args.soft_update_tau,
episodic_update=args.episodic_replay,
episodic_update_len=16)
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_runs=args.eval_n_runs,
max_episode_len=timestep_limit)
print('n_runs: {} mean: {} median: {} stdev {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_runs=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
eval_env=eval_env,
max_episode_len=timestep_limit)
def main():
import logging
logging.basicConfig(level=logging.DEBUG)
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)
args = parser.parse_args()
# Set a random seed used in ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print('Output files are saved in {}'.format(args.outdir))
def clip_action_filter(a):
return np.clip(a, action_space.low, action_space.high)
def make_env(test):
env = gym.make(args.env)
# Use different random seeds for train and test envs
env_seed = 2**32 - 1 - args.seed if test else args.seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = chainerrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = chainerrl.wrappers.Monitor(env, args.outdir)
if isinstance(env.action_space, spaces.Box):
misc.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 = chainerrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if ((args.render_eval and test) or (args.render_train and not test)):
env = chainerrl.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:
links.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
# Draw the computational graph and save it in the output directory.
chainerrl.misc.draw_computational_graph(
[q_func(np.zeros_like(obs_space.low, dtype=np.float32)[None])],
os.path.join(args.outdir, 'model'))
opt = optimizers.Adam()
opt.setup(q_func)
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_buffer.PrioritizedReplayBuffer(rbuf_capacity,
betasteps=betasteps)
else:
rbuf = replay_buffer.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']))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
eval_env=eval_env,
train_max_episode_len=timestep_limit)
def main():
import logging
parser = argparse.ArgumentParser()
parser.add_argument('--env', type=str, default='Pendulum-v0')
parser.add_argument('--arch',
type=str,
default='Gaussian',
choices=('FFSoftmax', 'FFMellowmax', 'Gaussian'))
parser.add_argument('--seed',
type=int,
default=0,
help='Random seed [0, 2 ** 32)')
parser.add_argument('--outdir', type=str, default=None)
parser.add_argument('--profile', action='store_true')
parser.add_argument('--steps', type=int, default=8 * 10**7)
parser.add_argument('--update-steps', type=int, default=5)
parser.add_argument('--log-interval', type=int, default=1000)
parser.add_argument('--eval-interval', type=int, default=10**5)
parser.add_argument('--eval-n-runs', type=int, default=10)
parser.add_argument('--reward-scale-factor', type=float, default=1e-2)
parser.add_argument('--rmsprop-epsilon', type=float, default=1e-5)
parser.add_argument('--render', action='store_true', default=False)
parser.add_argument('--gamma',
type=float,
default=0.99,
help='discount factor')
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('--lr', type=float, default=7e-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('--logger-level', type=int, default=logging.DEBUG)
parser.add_argument('--monitor', action='store_true')
parser.add_argument('--max-grad-norm',
type=float,
default=0.5,
help='value loss coefficient')
parser.add_argument('--alpha',
type=float,
default=0.99,
help='RMSprop optimizer alpha')
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)
args = parser.parse_args()
logging.basicConfig(level=args.logger_level)
# Set a random seed used in ChainerRL.
# If you use more than one processes, the results will be no longer
# deterministic even with the same random seed.
misc.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 = chainerrl.wrappers.CastObservationToFloat32(env)
if args.monitor and process_idx == 0:
env = gym.wrappers.Monitor(env, args.outdir)
# Scale rewards observed by agents
if not test:
misc.env_modifiers.make_reward_filtered(
env, lambda x: x * args.reward_scale_factor)
if args.render and process_idx == 0 and not test:
env = chainerrl.wrappers.Render(env)
return env
def make_batch_env(test):
return chainerrl.envs.MultiprocessVectorEnv([
(lambda: 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.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_space = sample_env.observation_space
action_space = sample_env.action_space
# Switch policy types accordingly to action space types
if args.arch == 'Gaussian':
model = A2CGaussian(obs_space.low.size, action_space.low.size)
elif args.arch == 'FFSoftmax':
model = A2CFFSoftmax(obs_space.low.size, action_space.n)
elif args.arch == 'FFMellowmax':
model = A2CFFMellowmax(obs_space.low.size, action_space.n)
optimizer = chainer.optimizers.RMSprop(args.lr,
eps=args.rmsprop_epsilon,
alpha=args.alpha)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.GradientClipping(args.max_grad_norm))
if args.weight_decay > 0:
optimizer.add_hook(NonbiasWeightDecay(args.weight_decay))
agent = a2c.A2C(model,
optimizer,
gamma=args.gamma,
gpu=args.gpu,
num_processes=args.num_envs,
update_steps=args.update_steps,
use_gae=args.use_gae,
tau=args.tau)
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_runs=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),
steps=args.steps,
log_interval=args.log_interval,
eval_n_runs=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
)
sample_env.close()
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('--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('--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('--logging-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.logging_level)
# Set a random seed used in ChainerRL.
# If you use more than one processes, the results will be no longer
# deterministic even with the same random seed.
misc.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 = chainerrl.wrappers.Monitor(
env, args.outdir,
mode='evaluation' if test else 'training')
if args.render:
env = chainerrl.wrappers.Render(env)
return env
def make_batch_env(test):
return chainerrl.envs.MultiprocessVectorEnv(
[functools.partial(make_env, idx, test)
for idx, env in enumerate(range(args.num_envs))])
sample_env = make_env(0, test=False)
n_actions = sample_env.action_space.n
model = A2CFF(n_actions)
optimizer = rmsprop_async.RMSpropAsync(lr=args.lr,
eps=args.rmsprop_epsilon,
alpha=args.alpha)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.GradientClipping(args.max_grad_norm))
if args.weight_decay > 0:
optimizer.add_hook(NonbiasWeightDecay(args.weight_decay))
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,
)
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
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--env', type=str, default='Hopper-v2')
parser.add_argument('--num-envs', type=int, default=1)
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**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('--logger-level', type=int, default=logging.DEBUG)
parser.add_argument('--monitor', action='store_true')
parser.add_argument('--window-size', type=int, default=100)
parser.add_argument('--update-interval', type=int, default=2048)
parser.add_argument('--log-interval', type=int, default=1000)
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 ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
# 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 = chainerrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = chainerrl.wrappers.Monitor(env, args.outdir)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = chainerrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if args.render:
env = chainerrl.wrappers.Render(env)
return env
def make_batch_env(test):
return chainerrl.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.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_space = sample_env.observation_space
action_space = sample_env.action_space
# Normalize observations based on their empirical mean and variance
obs_normalizer = chainerrl.links.EmpiricalNormalization(obs_space.low.size,
clip_threshold=5)
winit_last = chainer.initializers.LeCunNormal(1e-2)
# Switch policy types accordingly to action space types
if isinstance(action_space, gym.spaces.Discrete):
n_actions = action_space.n
policy = chainer.Sequential(
L.Linear(None, 64),
F.tanh,
L.Linear(None, 64),
F.tanh,
L.Linear(None, n_actions, initialW=winit_last),
chainerrl.distribution.SoftmaxDistribution,
)
elif isinstance(action_space, gym.spaces.Box):
action_size = action_space.low.size
policy = chainer.Sequential(
L.Linear(None, 64),
F.tanh,
L.Linear(None, 64),
F.tanh,
L.Linear(None, action_size, initialW=winit_last),
chainerrl.policies.GaussianHeadWithStateIndependentCovariance(
action_size=action_size,
var_type='diagonal',
var_func=lambda x: F.exp(2 * x), # Parameterize log std
var_param_init=0, # log std = 0 => std = 1
),
)
else:
print("""\
This example only supports gym.spaces.Box or gym.spaces.Discrete action spaces."""
) # NOQA
return
vf = chainer.Sequential(
L.Linear(None, 64),
F.tanh,
L.Linear(None, 64),
F.tanh,
L.Linear(None, 1),
)
# Combine a policy and a value function into a single model
model = chainerrl.links.Branched(policy, vf)
opt = chainer.optimizers.Adam(alpha=args.lr, eps=1e-5)
opt.setup(model)
if args.weight_decay > 0:
opt.add_hook(NonbiasWeightDecay(args.weight_decay))
agent = PPO(
model,
opt,
obs_normalizer=obs_normalizer,
gpu=args.gpu,
update_interval=args.update_interval,
minibatch_size=args.batchsize,
epochs=args.epochs,
clip_eps_vf=None,
entropy_coef=args.entropy_coef,
standardize_advantages=args.standardize_advantages,
)
if args.load:
agent.load(args.load)
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:
# Linearly decay the learning rate to zero
def lr_setter(env, agent, value):
agent.optimizer.alpha = value
lr_decay_hook = 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,
eval_interval=args.eval_interval,
log_interval=args.log_interval,
return_window_size=args.window_size,
max_episode_len=timestep_limit,
save_best_so_far_agent=False,
step_hooks=[
lr_decay_hook,
],
)
def main():
import logging
parser = argparse.ArgumentParser()
parser.add_argument('processes', type=int)
parser.add_argument('--env', type=str, default='CartPole-v0')
parser.add_argument('--arch', type=str, default='FFSoftmax',
choices=('FFSoftmax', 'FFMellowmax', 'LSTMGaussian'))
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('--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('--eval-interval', type=int, default=10 ** 5)
parser.add_argument('--eval-n-runs', type=int, default=10)
parser.add_argument('--reward-scale-factor', type=float, default=1e-2)
parser.add_argument('--rmsprop-epsilon', type=float, default=1e-1)
parser.add_argument('--render', action='store_true', default=False)
parser.add_argument('--lr', type=float, default=7e-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('--logger-level', type=int, default=logging.DEBUG)
parser.add_argument('--monitor', action='store_true')
args = parser.parse_args()
logging.basicConfig(level=args.logger_level)
# Set a random seed used in ChainerRL.
# If you use more than one processes, the results will be no longer
# deterministic even with the same random seed.
misc.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 ** 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)
if args.monitor and process_idx == 0:
env = gym.wrappers.Monitor(env, args.outdir)
# Scale rewards observed by agents
if not test:
misc.env_modifiers.make_reward_filtered(
env, lambda x: x * args.reward_scale_factor)
if args.render and process_idx == 0 and not test:
misc.env_modifiers.make_rendered(env)
return env
sample_env = gym.make(args.env)
timestep_limit = sample_env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_space = sample_env.observation_space
action_space = sample_env.action_space
# Switch policy types accordingly to action space types
if args.arch == 'LSTMGaussian':
model = A3CLSTMGaussian(obs_space.low.size, action_space.low.size)
elif args.arch == 'FFSoftmax':
model = A3CFFSoftmax(obs_space.low.size, action_space.n)
elif args.arch == 'FFMellowmax':
model = A3CFFMellowmax(obs_space.low.size, action_space.n)
opt = rmsprop_async.RMSpropAsync(
lr=args.lr, eps=args.rmsprop_epsilon, alpha=0.99)
opt.setup(model)
opt.add_hook(chainer.optimizer.GradientClipping(40))
if args.weight_decay > 0:
opt.add_hook(NonbiasWeightDecay(args.weight_decay))
agent = a3c.A3C(model, opt, t_max=args.t_max, gamma=0.99,
beta=args.beta, phi=phi)
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_runs=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_async(
agent=agent,
outdir=args.outdir,
processes=args.processes,
make_env=make_env,
profile=args.profile,
steps=args.steps,
eval_n_runs=args.eval_n_runs,
eval_interval=args.eval_interval,
max_episode_len=timestep_limit)
def main():
import logging
logging.basicConfig(level=logging.DEBUG)
parser = argparse.ArgumentParser()
parser.add_argument('processes', type=int)
parser.add_argument('rom', type=str)
parser.add_argument('--seed',
type=int,
default=0,
help='Random seed [0, 2 ** 31)')
parser.add_argument('--outdir', type=str, default=None)
parser.add_argument('--use-sdl', action='store_true')
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('--max-episode-len', type=int, default=10000)
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('--lr', type=float, default=7e-4)
parser.add_argument('--eval-interval', type=int, default=10**6)
parser.add_argument('--eval-n-runs', type=int, default=10)
parser.add_argument('--weight-decay', type=float, default=0.0)
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.set_defaults(use_sdl=False)
parser.set_defaults(use_lstm=False)
args = parser.parse_args()
# Set a random seed used in ChainerRL.
# If you use more than one processes, the results will be no longer
# deterministic even with the same random seed.
misc.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 = ale.ALE(args.rom).number_of_actions
if args.use_lstm:
model = acer.ACERSharedModel(
shared=links.Sequence(links.NIPSDQNHead(), L.LSTM(256, 256)),
pi=links.Sequence(L.Linear(256, n_actions), SoftmaxDistribution),
q=links.Sequence(L.Linear(256, n_actions), DiscreteActionValue),
)
else:
model = acer.ACERSharedModel(
shared=links.NIPSDQNHead(),
pi=links.Sequence(L.Linear(256, n_actions), SoftmaxDistribution),
q=links.Sequence(L.Linear(256, n_actions), DiscreteActionValue),
)
opt = rmsprop_async.RMSpropAsync(lr=7e-4, eps=4e-3, alpha=0.99)
opt.setup(model)
opt.add_hook(chainer.optimizer.GradientClipping(40))
if args.weight_decay > 0:
opt.add_hook(NonbiasWeightDecay(args.weight_decay))
replay_buffer = EpisodicReplayBuffer(10**6 // args.processes)
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=dqn_phi)
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 = ale.ALE(args.rom,
use_sdl=args.use_sdl,
treat_life_lost_as_terminal=not test,
seed=env_seed)
if not test:
misc.env_modifiers.make_reward_clipped(env, -1, 1)
return env
if args.demo:
env = make_env(0, True)
eval_stats = experiments.eval_performance(env=env,
agent=agent,
n_runs=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):
agent.optimizer.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_runs=args.eval_n_runs,
eval_interval=args.eval_interval,
max_episode_len=args.max_episode_len,
global_step_hooks=[lr_decay_hook])
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('--use-sdl', action='store_true', default=False)
parser.add_argument('--eval-epsilon', type=float, default=0.0)
parser.add_argument('--noisy-net-sigma', type=float, default=0.5)
parser.add_argument('--steps', type=int, default=5 * 10**7)
parser.add_argument(
'--max-frames',
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help='Maximum number of frames for each episode.')
parser.add_argument('--replay-start-size', type=int, default=2 * 10**4)
parser.add_argument('--eval-n-steps', type=int, default=125000)
parser.add_argument('--eval-interval', type=int, default=250000)
parser.add_argument('--logging-level',
type=int,
default=20,
help='Logging level. 10:DEBUG, 20:INFO etc.')
parser.add_argument('--render',
action='store_true',
default=False,
help='Render env states in a GUI window.')
parser.add_argument('--monitor',
action='store_true',
default=False,
help='Monitor env. Videos and additional information'
' are saved as output files.')
parser.add_argument('--n-best-episodes', type=int, default=200)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.logging_level)
# Set a random seed used in ChainerRL.
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print('Output files are saved in {}'.format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(atari_wrappers.make_atari(
args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = chainerrl.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 = chainerrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
n_actions = env.action_space.n
n_atoms = 51
v_max = 10
v_min = -10
q_func = DistributionalDuelingDQN(
n_actions,
n_atoms,
v_min,
v_max,
)
# Noisy nets
links.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
# Draw the computational graph and save it in the output directory.
chainerrl.misc.draw_computational_graph(
[q_func(np.zeros((4, 84, 84), dtype=np.float32)[None])],
os.path.join(args.outdir, 'model'))
# Use the same hyper parameters as https://arxiv.org/abs/1707.06887
opt = chainer.optimizers.Adam(6.25e-5, eps=1.5 * 10**-4)
opt.setup(q_func)
# Prioritized Replay
# Anneal beta from beta0 to 1 throughout training
update_interval = 4
betasteps = args.steps / update_interval
rbuf = replay_buffer.PrioritizedReplayBuffer(10**6,
alpha=0.5,
beta0=0.4,
betasteps=betasteps,
num_steps=3)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
Agent = agents.CategoricalDoubleDQN
agent = Agent(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
minibatch_size=32,
replay_start_size=args.replay_start_size,
target_update_interval=32000,
update_interval=update_interval,
batch_accumulator='mean',
phi=phi,
)
if args.load:
agent.load(args.load)
if args.demo:
eval_stats = experiments.eval_performance(env=eval_env,
agent=agent,
n_steps=args.eval_n_steps,
n_episodes=None)
print('n_episodes: {} mean: {} median: {} stdev {}'.format(
eval_stats['episodes'], eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=args.eval_n_steps,
eval_n_episodes=None,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=True,
eval_env=eval_env,
)
dir_of_best_network = os.path.join(args.outdir, "best")
agent.load(dir_of_best_network)
# run 200 evaluation episodes, each capped at 30 mins of play
stats = experiments.evaluator.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.n_best_episodes,
max_episode_len=args.max_frames / 4,
logger=None)
with open(os.path.join(args.outdir, 'bestscores.json'), 'w') as f:
# temporary hack to handle python 2/3 support issues.
# json dumps does not support non-string literal dict keys
json_stats = json.dumps(stats)
print(str(json_stats), file=f)
print("The results of the best scoring network:")
for stat in stats:
print(str(stat) + ":" + str(stats[stat]))
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)
args = parser.parse_args()
misc.set_random_seed(args.seed, gpus=(args.gpu,))
args.outdir = experiments.prepare_output_dir(
args, args.outdir, argv=sys.argv)
print('Output files are saved in {}'.format(args.outdir))
""" initialize clientpool and environment """
client_pool = [('127.0.0.1', 10000), ('127.0.0.1', 10001), ('127.0.0.1', 10002)]
env.init(start_minecraft=False, client_pool=client_pool)
test = True
""" set the seed """
env_seed = args.seed
env.seed(env_seed)
"""Cast observations to float32 because the model needs float32"""
env = chainerrl.wrappers.CastObservationToFloat32(env)
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('--logging-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.logging_level)
# Set a random seed used in ChainerRL.
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
# 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=not args.no_frame_stack,
)
env.seed(env_seed)
if args.monitor:
env = chainerrl.wrappers.Monitor(
env, args.outdir, mode='evaluation' if test else 'training')
if args.render:
env = chainerrl.wrappers.Render(env)
return env
def make_batch_env(test):
return chainerrl.envs.MultiprocessVectorEnv([
(lambda: make_env(idx, test))
for idx, env in enumerate(range(args.num_envs))
])
sample_env = make_env(0, test=False)
print('Observation space', sample_env.observation_space)
print('Action space', sample_env.action_space)
n_actions = sample_env.action_space.n
winit_last = chainer.initializers.LeCunNormal(1e-2)
if args.recurrent:
model = chainerrl.links.StatelessRecurrentSequential(
L.Convolution2D(None, 32, 8, stride=4), F.relu,
L.Convolution2D(None, 64, 4, stride=2), F.relu,
L.Convolution2D(None, 64, 3, stride=1), F.relu,
L.Linear(None, 512), F.relu, L.NStepGRU(1, 512, 512, 0),
chainerrl.links.Branched(
chainer.Sequential(
L.Linear(None, n_actions, initialW=winit_last),
chainerrl.distribution.SoftmaxDistribution,
),
L.Linear(None, 1),
))
else:
model = chainer.Sequential(
L.Convolution2D(None, 32, 8, stride=4), F.relu,
L.Convolution2D(None, 64, 4, stride=2), F.relu,
L.Convolution2D(None, 64, 3, stride=1), F.relu,
L.Linear(None, 512), F.relu,
chainerrl.links.Branched(
chainer.Sequential(
L.Linear(None, n_actions, initialW=winit_last),
chainerrl.distribution.SoftmaxDistribution,
),
L.Linear(None, 1),
))
# Draw the computational graph and save it in the output directory.
fake_obss = np.zeros(sample_env.observation_space.shape,
dtype=np.float32)[None]
if args.recurrent:
fake_out, _ = model(fake_obss, None)
else:
fake_out = model(fake_obss)
chainerrl.misc.draw_computational_graph([fake_out],
os.path.join(args.outdir, 'model'))
opt = chainer.optimizers.Adam(alpha=args.lr, eps=1e-5)
opt.setup(model)
opt.add_hook(chainer.optimizer.GradientClipping(0.5))
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,
)
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):
agent.optimizer.alpha = 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():
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='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('--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('--logger-level',
type=int,
default=logging.INFO,
help='Level of the root logger.')
parser.add_argument('--monitor',
action='store_true',
help='Wrap env with gym.wrappers.Monitor.')
parser.add_argument('--log-interval',
type=int,
default=1000,
help='Interval in timesteps between outputting log'
' messages during training')
parser.add_argument('--update-interval',
type=int,
default=2048,
help='Interval in timesteps between model updates.')
parser.add_argument('--epochs',
type=int,
default=10,
help='Number of epochs to update model for per PPO'
' iteration.')
parser.add_argument('--batch-size',
type=int,
default=64,
help='Minibatch size')
args = parser.parse_args()
logging.basicConfig(level=args.logger_level)
# Set a random seed used in ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
# 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 = chainerrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = chainerrl.wrappers.Monitor(env, args.outdir)
if args.render:
env = chainerrl.wrappers.Render(env)
return env
def make_batch_env(test):
return chainerrl.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 = chainerrl.links.EmpiricalNormalization(obs_space.low.size,
clip_threshold=5)
# While the original paper initialized weights by normal distribution,
# we use orthogonal initialization as the latest openai/baselines does.
winit = chainerrl.initializers.Orthogonal(1.)
winit_last = chainerrl.initializers.Orthogonal(1e-2)
action_size = action_space.low.size
policy = chainer.Sequential(
L.Linear(None, 64, initialW=winit),
F.tanh,
L.Linear(None, 64, initialW=winit),
F.tanh,
L.Linear(None, action_size, initialW=winit_last),
chainerrl.policies.GaussianHeadWithStateIndependentCovariance(
action_size=action_size,
var_type='diagonal',
var_func=lambda x: F.exp(2 * x), # Parameterize log std
var_param_init=0, # log std = 0 => std = 1
),
)
vf = chainer.Sequential(
L.Linear(None, 64, initialW=winit),
F.tanh,
L.Linear(None, 64, initialW=winit),
F.tanh,
L.Linear(None, 1, initialW=winit),
)
# Combine a policy and a value function into a single model
model = chainerrl.links.Branched(policy, vf)
opt = chainer.optimizers.Adam(3e-4, eps=1e-5)
opt.setup(model)
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:
# 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(
misc.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('--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-episode-len',
type=int,
default=30 * 60 * 60 // 4, # 30 minutes with 60/4 fps
help='Maximum number of steps 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('--logging-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)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.logging_level)
# Set a random seed used in ChainerRL.
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
# 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),
episode_life=not test,
clip_rewards=not test)
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = chainerrl.wrappers.RandomizeAction(env, args.eval_epsilon)
env.seed(env_seed)
if args.monitor:
env = gym.wrappers.Monitor(
env, args.outdir, mode='evaluation' if test else 'training')
if args.render:
env = chainerrl.wrappers.Render(env)
return env
def make_batch_env(test):
return chainerrl.envs.MultiprocessVectorEnv([
(lambda: make_env(idx, test))
for idx, env in enumerate(range(args.num_envs))
])
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:
links.to_factorized_noisy(q_func)
# Turn off explorer
explorer = explorers.Greedy()
# Draw the computational graph and save it in the output directory.
chainerrl.misc.draw_computational_graph(
[q_func(np.zeros((4, 84, 84), dtype=np.float32)[None])],
os.path.join(args.outdir, 'model'))
# Use the same hyper parameters as the Nature paper's
opt = optimizers.RMSpropGraves(lr=args.lr,
alpha=0.95,
momentum=0.0,
eps=1e-2)
opt.setup(q_func)
# 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_buffer.PrioritizedReplayBuffer(10**6,
alpha=0.6,
beta0=0.4,
betasteps=betasteps)
else:
rbuf = replay_buffer.ReplayBuffer(10**6)
explorer = explorers.LinearDecayEpsilonGreedy(
1.0, args.final_epsilon, args.final_exploration_frames,
lambda: np.random.randint(n_actions))
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
Agent = 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_runs=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_runs=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=False,
max_episode_len=args.max_episode_len,
log_interval=1000,
)
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('--lr', type=float, default=7e-4)
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('--final-exploration-frames',
type=int,
default=4 * 10**6)
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('--profile', action='store_true')
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=None)
parser.add_argument('--logging-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.logging_level)
# Set a random seed used in ChainerRL.
# If you use more than one processes, the results will be no longer
# deterministic even with the same random seed.
misc.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 test:
# Randomize actions like epsilon-greedy in evaluation as well
env = chainerrl.wrappers.RandomizeAction(env, 0.05)
if args.monitor:
env = chainerrl.wrappers.Monitor(
env, args.outdir, mode='evaluation' if test else 'training')
if args.render:
env = chainerrl.wrappers.Render(env)
return env
sample_env = make_env(0, test=False)
action_space = sample_env.action_space
assert isinstance(action_space, spaces.Discrete)
# Define a model and its optimizer
q_func = links.Sequence(links.NIPSDQNHead(), L.Linear(256, action_space.n),
DiscreteActionValue)
opt = rmsprop_async.RMSpropAsync(lr=args.lr, eps=1e-1, alpha=0.99)
opt.setup(q_func)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
# Make process-specific agents to diversify exploration
def make_agent(process_idx):
# Random epsilon assignment described in the original paper
rand = random.random()
if rand < 0.4:
epsilon_target = 0.1
elif rand < 0.7:
epsilon_target = 0.01
else:
epsilon_target = 0.5
explorer = explorers.LinearDecayEpsilonGreedy(
1, epsilon_target, args.final_exploration_frames,
action_space.sample)
# Suppress the explorer logger
explorer.logger.setLevel(logging.INFO)
return nsq.NSQ(q_func,
opt,
t_max=5,
gamma=0.99,
i_target=40000,
explorer=explorer,
phi=phi)
if args.demo:
env = make_env(0, True)
agent = make_agent(0)
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):
agent.optimizer.lr = value
lr_decay_hook = experiments.LinearInterpolationHook(
args.steps, args.lr, 0, lr_setter)
experiments.train_agent_async(
outdir=args.outdir,
processes=args.processes,
make_env=make_env,
make_agent=make_agent,
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(self):
import logging
logging.basicConfig(level=logging.INFO)
# Set a random seed used in ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print('Output files are saved in {}'.format(args.outdir))
env = self.env_make(test=False)
timestep_limit = env.total_time
obs_size = env.observation.size
action_space = env.action_space
# Q function
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:
links.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
# Draw the computational graph and save it in the output directory.
# chainerrl.misc.draw_computational_graph([q_func(np.zeros_like(obs_space.low, dtype=np.float32)[None])],
# os.path.join(args.outdir, 'model'))
opt = optimizers.Adam()
opt.setup(q_func)
rbuf = self.buffer()
agent = DQN(q_func,
opt,
rbuf,
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 = self.env_make(test=True)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
max_episode_len=timestep_limit)
print('n_runs: {} mean: {} median: {} stdev: {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
eval_env=eval_env,
train_max_episode_len=timestep_limit)
pass
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--outdir', type=str, default='results',
help='Directory path to save output files.'
' If it does not exist, it will be created.')
parser.add_argument('--env', type=str, default='Hopper-v2',
help='OpenAI Gym MuJoCo env to perform algorithm on.')
parser.add_argument('--seed', type=int, default=0,
help='Random seed [0, 2 ** 32)')
parser.add_argument('--gpu', type=int, default=0,
help='GPU to use, set to -1 if no GPU.')
parser.add_argument('--load', type=str, default='',
help='Directory to load agent from.')
parser.add_argument('--steps', type=int, default=10 ** 6,
help='Total number of timesteps to train the agent.')
parser.add_argument('--eval-n-runs', type=int, default=10,
help='Number of episodes run for each evaluation.')
parser.add_argument('--eval-interval', type=int, default=5000,
help='Interval in timesteps between evaluations.')
parser.add_argument('--replay-start-size', type=int, default=10000,
help='Minimum replay buffer size before ' +
'performing gradient updates.')
parser.add_argument('--batch-size', type=int, default=100,
help='Minibatch size')
parser.add_argument('--render', action='store_true',
help='Render env states in a GUI window.')
parser.add_argument('--demo', action='store_true',
help='Just run evaluation, not training.')
parser.add_argument('--monitor', action='store_true',
help='Wrap env with gym.wrappers.Monitor.')
parser.add_argument('--logger-level', type=int, default=logging.INFO,
help='Level of the root logger.')
args = parser.parse_args()
logging.basicConfig(level=args.logger_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 ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu,))
def make_env(test):
env = gym.make(args.env)
# Unwrap TimeLimit wrapper
assert isinstance(env, gym.wrappers.TimeLimit)
env = env.env
# Use different random seeds for train and test envs
env_seed = 2 ** 32 - 1 - args.seed if test else args.seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = chainerrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = chainerrl.wrappers.Monitor(env, args.outdir)
if args.render and not test:
env = chainerrl.wrappers.Render(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_space = env.observation_space
action_space = env.action_space
print('Observation space:', obs_space)
print('Action space:', action_space)
action_size = action_space.low.size
winit = chainer.initializers.LeCunUniform(3 ** -0.5)
q_func = chainer.Sequential(
concat_obs_and_action,
L.Linear(None, 400, initialW=winit),
F.relu,
L.Linear(None, 300, initialW=winit),
F.relu,
L.Linear(None, 1, initialW=winit),
)
policy = chainer.Sequential(
L.Linear(None, 400, initialW=winit),
F.relu,
L.Linear(None, 300, initialW=winit),
F.relu,
L.Linear(None, action_size, initialW=winit),
F.tanh,
chainerrl.distribution.ContinuousDeterministicDistribution,
)
model = DDPGModel(q_func=q_func, policy=policy)
# Draw the computational graph and save it in the output directory.
fake_obs = chainer.Variable(
model.xp.zeros_like(obs_space.low, dtype=np.float32)[None],
name='observation')
fake_action = chainer.Variable(
model.xp.zeros_like(action_space.low, dtype=np.float32)[None],
name='action')
chainerrl.misc.draw_computational_graph(
[policy(fake_obs)], os.path.join(args.outdir, 'policy'))
chainerrl.misc.draw_computational_graph(
[q_func(fake_obs, fake_action)], os.path.join(args.outdir, 'q_func'))
opt_a = optimizers.Adam()
opt_c = optimizers.Adam()
opt_a.setup(model['policy'])
opt_c.setup(model['q_function'])
rbuf = replay_buffer.ReplayBuffer(10 ** 6)
explorer = explorers.AdditiveGaussian(
scale=0.1, low=action_space.low, high=action_space.high)
def burnin_action_func():
"""Select random actions until model is updated one or more times."""
return np.random.uniform(
action_space.low, action_space.high).astype(np.float32)
# Hyperparameters in http://arxiv.org/abs/1802.09477
agent = DDPG(
model,
opt_a,
opt_c,
rbuf,
gamma=0.99,
explorer=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=burnin_action_func,
)
if len(args.load) > 0:
agent.load(args.load)
eval_env = make_env(test=True)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
max_episode_len=timestep_limit)
print('n_runs: {} mean: {} median: {} stdev {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
experiments.train_agent_with_evaluation(
agent=agent, env=env, steps=args.steps,
eval_env=eval_env, eval_n_steps=None,
eval_n_episodes=args.eval_n_runs, eval_interval=args.eval_interval,
outdir=args.outdir,
train_max_episode_len=timestep_limit)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('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('--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=10**6)
parser.add_argument('--eval-n-runs', type=int, default=10)
parser.add_argument('--weight-decay', type=float, default=0.0)
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('--logging-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.logging_level)
# Set a random seed used in ChainerRL.
# If you use more than one processes, the results will be no longer
# deterministic even with the same random seed.
misc.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
if args.use_lstm:
model = A3CLSTM(n_actions)
else:
model = A3CFF(n_actions)
# Draw the computational graph and save it in the output directory.
fake_obs = chainer.Variable(np.zeros((4, 84, 84), dtype=np.float32)[None],
name='observation')
with chainerrl.recurrent.state_reset(model):
# The state of the model is reset again after drawing the graph
chainerrl.misc.draw_computational_graph([model(fake_obs)],
os.path.join(
args.outdir, 'model'))
opt = rmsprop_async.RMSpropAsync(lr=7e-4, eps=1e-1, alpha=0.99)
opt.setup(model)
opt.add_hook(chainer.optimizer.GradientClipping(40))
if args.weight_decay > 0:
opt.add_hook(NonbiasWeightDecay(args.weight_decay))
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)
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 = gym.wrappers.Monitor(
env, args.outdir, mode='evaluation' if test else 'training')
if args.render:
env = chainerrl.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):
agent.optimizer.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_runs=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
parser = argparse.ArgumentParser()
parser.add_argument('--processes', type=int, default=8)
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--env', type=str, default='CartPole-v0')
parser.add_argument('--seed', type=int, default=0,
help='Random seed [0, 2 ** 32)')
parser.add_argument('--outdir', type=str, default='results',
help='Directory path to save output files.'
' If it does not exist, it will be created.')
parser.add_argument('--batchsize', type=int, default=10)
parser.add_argument('--rollout-len', type=int, default=10)
parser.add_argument('--n-hidden-channels', type=int, default=100)
parser.add_argument('--n-hidden-layers', type=int, default=2)
parser.add_argument('--n-times-replay', type=int, default=1)
parser.add_argument('--replay-start-size', type=int, default=10000)
parser.add_argument('--t-max', type=int, default=None)
parser.add_argument('--tau', 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('--eval-interval', type=int, default=10 ** 5)
parser.add_argument('--eval-n-runs', type=int, default=10)
parser.add_argument('--reward-scale-factor', type=float, default=1e-2)
parser.add_argument('--render', action='store_true', default=False)
parser.add_argument('--lr', type=float, default=7e-4)
parser.add_argument('--demo', action='store_true', default=False)
parser.add_argument('--load', type=str, default='')
parser.add_argument('--logger-level', type=int, default=logging.DEBUG)
parser.add_argument('--monitor', action='store_true')
parser.add_argument('--train-async', action='store_true', default=False)
parser.add_argument('--prioritized-replay', action='store_true',
default=False)
parser.add_argument('--disable-online-update', action='store_true',
default=False)
parser.add_argument('--backprop-future-values', action='store_true',
default=True)
parser.add_argument('--no-backprop-future-values', action='store_false',
dest='backprop_future_values')
args = parser.parse_args()
logging.basicConfig(level=args.logger_level)
# Set a random seed used in ChainerRL.
# If you use async training (--train-async), the results will be no longer
# deterministic even with the same random seed.
misc.set_random_seed(args.seed, gpus=(args.gpu,))
if args.train_async:
# 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 ** 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
if args.train_async:
process_seed = int(process_seeds[process_idx])
env_seed = 2 ** 32 - 1 - process_seed if test else process_seed
else:
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 = chainerrl.wrappers.CastObservationToFloat32(env)
if args.monitor and process_idx == 0:
env = chainerrl.wrappers.Monitor(env, args.outdir)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = chainerrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if args.render and process_idx == 0 and not test:
env = chainerrl.wrappers.Render(env)
return env
sample_env = gym.make(args.env)
timestep_limit = sample_env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_space = sample_env.observation_space
action_space = sample_env.action_space
# Switch policy types accordingly to action space types
if isinstance(action_space, gym.spaces.Box):
model = chainerrl.agents.pcl.PCLSeparateModel(
pi=chainerrl.policies.FCGaussianPolicy(
obs_space.low.size, action_space.low.size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
bound_mean=True,
min_action=action_space.low,
max_action=action_space.high,
var_wscale=1e-3,
var_bias=1,
var_type='diagonal',
),
v=chainerrl.v_functions.FCVFunction(
obs_space.low.size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
)
)
else:
model = chainerrl.agents.pcl.PCLSeparateModel(
pi=chainerrl.policies.FCSoftmaxPolicy(
obs_space.low.size, action_space.n,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers
),
v=chainerrl.v_functions.FCVFunction(
obs_space.low.size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
),
)
if not args.train_async and args.gpu >= 0:
chainer.cuda.get_device(args.gpu).use()
model.to_gpu(args.gpu)
if args.train_async:
opt = rmsprop_async.RMSpropAsync(lr=args.lr, alpha=0.99)
else:
opt = chainer.optimizers.Adam(alpha=args.lr)
opt.setup(model)
if args.prioritized_replay:
replay_buffer = \
chainerrl.replay_buffer.PrioritizedEpisodicReplayBuffer(
capacity=5 * 10 ** 3,
uniform_ratio=0.1,
default_priority_func=exp_return_of_episode,
wait_priority_after_sampling=False,
return_sample_weights=False)
else:
replay_buffer = chainerrl.replay_buffer.EpisodicReplayBuffer(
capacity=5 * 10 ** 3)
agent = chainerrl.agents.PCL(
model, opt, replay_buffer=replay_buffer,
t_max=args.t_max, gamma=0.99,
tau=args.tau,
rollout_len=args.rollout_len,
n_times_replay=args.n_times_replay,
replay_start_size=args.replay_start_size,
batchsize=args.batchsize,
train_async=args.train_async,
disable_online_update=args.disable_online_update,
backprop_future_values=args.backprop_future_values,
)
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=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:
if args.train_async:
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,
max_episode_len=timestep_limit)
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=make_env(0, test=False),
eval_env=make_env(0, test=True),
outdir=args.outdir,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
train_max_episode_len=timestep_limit)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--env', type=str, default='BreakoutNoFrameskip-v4')
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('--steps', type=int, default=5 * 10**7)
parser.add_argument(
'--max-frames',
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help='Maximum number of frames for each episode.')
parser.add_argument('--replay-start-size', type=int, default=5 * 10**4)
parser.add_argument('--target-update-interval', type=int, default=10**4)
parser.add_argument('--eval-interval', type=int, default=10**5)
parser.add_argument('--update-interval', type=int, default=4)
parser.add_argument('--eval-n-runs', type=int, default=10)
parser.add_argument('--batch-size', type=int, default=32)
parser.add_argument('--logging-level',
type=int,
default=20,
help='Logging level. 10:DEBUG, 20:INFO etc.')
parser.add_argument('--render',
action='store_true',
default=False,
help='Render env states in a GUI window.')
parser.add_argument('--monitor',
action='store_true',
default=False,
help='Monitor env. Videos and additional information'
' are saved as output files.')
parser.add_argument('--batch-accumulator',
type=str,
default='mean',
choices=['mean', 'sum'])
parser.add_argument('--quantile-thresholds-N', type=int, default=64)
parser.add_argument('--quantile-thresholds-N-prime', type=int, default=64)
parser.add_argument('--quantile-thresholds-K', type=int, default=32)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.logging_level)
# Set a random seed used in ChainerRL.
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print('Output files are saved in {}'.format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(atari_wrappers.make_atari(
args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = chainerrl.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 = chainerrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
n_actions = env.action_space.n
q_func = chainerrl.agents.iqn.ImplicitQuantileQFunction(
psi=chainerrl.links.Sequence(
L.Convolution2D(None, 32, 8, stride=4),
F.relu,
L.Convolution2D(None, 64, 4, stride=2),
F.relu,
L.Convolution2D(None, 64, 3, stride=1),
F.relu,
functools.partial(F.reshape, shape=(-1, 3136)),
),
phi=chainerrl.links.Sequence(
chainerrl.agents.iqn.CosineBasisLinear(64, 3136),
F.relu,
),
f=chainerrl.links.Sequence(
L.Linear(None, 512),
F.relu,
L.Linear(None, n_actions),
),
)
# Draw the computational graph and save it in the output directory.
fake_obss = np.zeros((4, 84, 84), dtype=np.float32)[None]
fake_taus = np.zeros(32, dtype=np.float32)[None]
chainerrl.misc.draw_computational_graph([q_func(fake_obss)(fake_taus)],
os.path.join(args.outdir, 'model'))
# Use the same hyper parameters as https://arxiv.org/abs/1710.10044
opt = chainer.optimizers.Adam(5e-5, eps=1e-2 / args.batch_size)
opt.setup(q_func)
rbuf = replay_buffer.ReplayBuffer(10**6)
explorer = explorers.LinearDecayEpsilonGreedy(
1.0, args.final_epsilon, args.final_exploration_frames,
lambda: np.random.randint(n_actions))
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
agent = chainerrl.agents.IQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
batch_accumulator=args.batch_accumulator,
phi=phi,
quantile_thresholds_N=args.quantile_thresholds_N,
quantile_thresholds_N_prime=args.quantile_thresholds_N_prime,
quantile_thresholds_K=args.quantile_thresholds_K,
)
if args.load:
agent.load(args.load)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
)
print('n_runs: {} mean: {} median: {} stdev {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=False,
eval_env=eval_env,
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--env',
type=str,
default='BreakoutNoFrameskip-v4',
help='OpenAI Atari domain to perform algorithm on.')
parser.add_argument('--outdir',
type=str,
default='results',
help='Directory path to save output files.'
' If it does not exist, it will be created.')
parser.add_argument('--seed',
type=int,
default=0,
help='Random seed [0, 2 ** 31)')
parser.add_argument('--gpu',
type=int,
default=0,
help='GPU to use, set to -1 if no GPU.')
parser.add_argument('--demo', action='store_true', default=False)
parser.add_argument('--load', type=str, default=None)
parser.add_argument('--final-exploration-frames',
type=int,
default=10**6,
help='Timesteps after which we stop ' +
'annealing exploration rate')
parser.add_argument('--final-epsilon',
type=float,
default=0.01,
help='Final value of epsilon during training.')
parser.add_argument('--eval-epsilon',
type=float,
default=0.001,
help='Exploration epsilon used during eval episodes.')
parser.add_argument('--noisy-net-sigma', type=float, default=None)
parser.add_argument('--arch',
type=str,
default='doubledqn',
choices=['nature', 'nips', 'dueling', 'doubledqn'],
help='Network architecture to use.')
parser.add_argument('--steps',
type=int,
default=5 * 10**7,
help='Total number of timesteps to train the agent.')
parser.add_argument(
'--max-frames',
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help='Maximum number of frames for each episode.')
parser.add_argument('--replay-start-size',
type=int,
default=5 * 10**4,
help='Minimum replay buffer size before ' +
'performing gradient updates.')
parser.add_argument('--target-update-interval',
type=int,
default=3 * 10**4,
help='Frequency (in timesteps) at which ' +
'the target network is updated.')
parser.add_argument('--eval-interval',
type=int,
default=10**5,
help='Frequency (in timesteps) of evaluation phase.')
parser.add_argument('--update-interval',
type=int,
default=4,
help='Frequency (in timesteps) of network updates.')
parser.add_argument('--eval-n-runs', type=int, default=10)
parser.add_argument('--no-clip-delta',
dest='clip_delta',
action='store_false')
parser.add_argument('--num-step-return', type=int, default=1)
parser.set_defaults(clip_delta=True)
parser.add_argument('--agent',
type=str,
default='DoubleDQN',
choices=['DQN', 'DoubleDQN', 'PAL'])
parser.add_argument('--logging-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.')
args = parser.parse_args()
import logging
logging.basicConfig(level=args.logging_level)
# Set a random seed used in ChainerRL.
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print('Output files are saved in {}'.format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(atari_wrappers.make_atari(
args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = chainerrl.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 = chainerrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
n_actions = env.action_space.n
q_func = parse_arch(args.arch, n_actions)
if args.noisy_net_sigma is not None:
links.to_factorized_noisy(q_func)
# Turn off explorer
explorer = explorers.Greedy()
# Draw the computational graph and save it in the output directory.
chainerrl.misc.draw_computational_graph(
[q_func(np.zeros((4, 84, 84), dtype=np.float32)[None])],
os.path.join(args.outdir, 'model'))
# Use the Nature paper's hyperparameters
opt = optimizers.RMSpropGraves(lr=args.lr,
alpha=0.95,
momentum=0.0,
eps=1e-2)
opt.setup(q_func)
# 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_buffer.PrioritizedReplayBuffer(
10**6,
alpha=0.6,
beta0=0.4,
betasteps=betasteps,
num_steps=args.num_step_return)
else:
rbuf = replay_buffer.ReplayBuffer(10**6, args.num_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=eval_env,
agent=agent,
n_runs=args.eval_n_runs)
print('n_runs: {} mean: {} median: {} stdev {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=False,
eval_env=eval_env,
)
def main():
import logging
logging.basicConfig(level=logging.DEBUG)
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='Humanoid-v1')
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**6)
parser.add_argument('--actor-lr', type=float, default=1e-4)
parser.add_argument('--critic-lr', type=float, default=1e-3)
parser.add_argument('--load', type=str, default='')
parser.add_argument('--steps', type=int, default=10**7)
parser.add_argument('--n-hidden-channels', type=int, default=300)
parser.add_argument('--n-hidden-layers', type=int, default=3)
parser.add_argument('--replay-start-size', type=int, default=5000)
parser.add_argument('--n-update-times', type=int, default=1)
parser.add_argument('--target-update-interval', type=int, default=1)
parser.add_argument('--target-update-method',
type=str,
default='soft',
choices=['hard', 'soft'])
parser.add_argument('--soft-update-tau', type=float, default=1e-2)
parser.add_argument('--update-interval', type=int, default=4)
parser.add_argument('--eval-n-runs', type=int, default=100)
parser.add_argument('--eval-interval', type=int, default=10**5)
parser.add_argument('--gamma', type=float, default=0.995)
parser.add_argument('--minibatch-size', type=int, default=200)
parser.add_argument('--render', action='store_true')
parser.add_argument('--demo', action='store_true')
parser.add_argument('--use-bn', action='store_true', default=False)
parser.add_argument('--monitor', action='store_true')
parser.add_argument('--reward-scale-factor', type=float, default=1e-2)
args = parser.parse_args()
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 ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
def clip_action_filter(a):
return np.clip(a, action_space.low, action_space.high)
def reward_filter(r):
return r * args.reward_scale_factor
def make_env(test):
env = gym.make(args.env)
# Use different random seeds for train and test envs
env_seed = 2**32 - 1 - args.seed if test else args.seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = chainerrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = gym.wrappers.Monitor(env, args.outdir)
if isinstance(env.action_space, spaces.Box):
misc.env_modifiers.make_action_filtered(env, clip_action_filter)
if not test:
misc.env_modifiers.make_reward_filtered(env, reward_filter)
if args.render and not test:
misc.env_modifiers.make_rendered(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_size = np.asarray(env.observation_space.shape).prod()
action_space = env.action_space
action_size = np.asarray(action_space.shape).prod()
if args.use_bn:
q_func = q_functions.FCBNLateActionSAQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
normalize_input=True)
pi = policy.FCBNDeterministicPolicy(
obs_size,
action_size=action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
min_action=action_space.low,
max_action=action_space.high,
bound_action=True,
normalize_input=True)
else:
q_func = q_functions.FCSAQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers)
pi = policy.FCDeterministicPolicy(
obs_size,
action_size=action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
min_action=action_space.low,
max_action=action_space.high,
bound_action=True)
model = DDPGModel(q_func=q_func, policy=pi)
opt_a = optimizers.Adam(alpha=args.actor_lr)
opt_c = optimizers.Adam(alpha=args.critic_lr)
opt_a.setup(model['policy'])
opt_c.setup(model['q_function'])
opt_a.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_a')
opt_c.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_c')
rbuf = replay_buffer.ReplayBuffer(5 * 10**5)
def random_action():
a = action_space.sample()
if isinstance(a, np.ndarray):
a = a.astype(np.float32)
return a
ou_sigma = (action_space.high - action_space.low) * 0.2
explorer = explorers.AdditiveOU(sigma=ou_sigma)
agent = DDPG(model,
opt_a,
opt_c,
rbuf,
gamma=args.gamma,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_method=args.target_update_method,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
soft_update_tau=args.soft_update_tau,
n_times_update=args.n_update_times,
gpu=args.gpu,
minibatch_size=args.minibatch_size)
if len(args.load) > 0:
agent.load(args.load)
eval_env = make_env(test=True)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_runs=args.eval_n_runs,
max_episode_len=timestep_limit)
print('n_runs: {} mean: {} median: {} stdev {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_env=eval_env,
eval_n_runs=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
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('--logging-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.logging_level)
# Set a random seed used in ChainerRL.
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
# 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
misc.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,
)
# (84, 84, 3) -> (3, 84, 84)
env = TransposeObservation(env, (2, 0, 1))
env = ObserveElapsedSteps(env, max_episode_steps)
# KukaDiverseObjectEnv internally asserts int actions and does not
# accept python-future's newint.
env = CastAction(env, __builtins__.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 chainerrl.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)
# Draw the computational graph and save it in the output directory.
fake_obs = (
np.zeros((3, 84, 84), dtype=np.float32)[None],
np.zeros((), dtype=np.int32)[None],
)
chainerrl.misc.draw_computational_graph([q_func(fake_obs)],
os.path.join(args.outdir, 'model'))
# Use the hyper parameters of the Nature paper
opt = optimizers.RMSpropGraves(lr=args.lr,
alpha=0.95,
momentum=0.0,
eps=1e-2)
opt.setup(q_func)
# Anneal beta from beta0 to 1 throughout training
betasteps = args.steps / args.update_interval
rbuf = replay_buffer.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 = chainerrl.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():
import logging
logging.basicConfig(level=logging.DEBUG)
parser = argparse.ArgumentParser()
parser.add_argument('--outdir', type=str, default='dqn_out')
parser.add_argument('--env', type=str, default='Pendulum-v0')
parser.add_argument('--seed', type=int, default=None)
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('--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('--episodic-replay', action='store_true')
parser.add_argument('--replay-start-size', type=int, default=None)
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)
args = parser.parse_args()
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print('Output files are saved in {}'.format(args.outdir))
if args.seed is not None:
misc.set_random_seed(args.seed)
def clip_action_filter(a):
return np.clip(a, action_space.low, action_space.high)
def make_env(for_eval):
env = gym.make(args.env)
if args.monitor:
env = gym.wrappers.Monitor(env, args.outdir)
if isinstance(env.action_space, spaces.Box):
misc.env_modifiers.make_action_filtered(env, clip_action_filter)
if not for_eval:
misc.env_modifiers.make_reward_filtered(
env, lambda x: x * args.reward_scale_factor)
if ((args.render_eval and for_eval)
or (args.render_train and not for_eval)):
misc.env_modifiers.make_rendered(env)
return env
env = make_env(for_eval=False)
timestep_limit = env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_size = env.observation_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)
opt = optimizers.Adam()
opt.setup(q_func)
rbuf_capacity = 5 * 10**5
if args.episodic_replay:
if args.minibatch_size is None:
args.minibatch_size = 4
if args.replay_start_size is None:
args.replay_start_size = 10
if args.prioritized_replay:
betasteps = \
(args.steps - timestep_limit * args.replay_start_size) \
// args.update_interval
rbuf = replay_buffer.PrioritizedEpisodicReplayBuffer(
rbuf_capacity, betasteps=betasteps)
else:
rbuf = replay_buffer.EpisodicReplayBuffer(rbuf_capacity)
else:
if args.minibatch_size is None:
args.minibatch_size = 32
if args.replay_start_size is None:
args.replay_start_size = 1000
if args.prioritized_replay:
betasteps = (args.steps - args.replay_start_size) \
// args.update_interval
rbuf = replay_buffer.PrioritizedReplayBuffer(rbuf_capacity,
betasteps=betasteps)
else:
rbuf = replay_buffer.ReplayBuffer(rbuf_capacity)
def phi(obs):
return obs.astype(np.float32)
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,
phi=phi,
minibatch_size=args.minibatch_size,
target_update_method=args.target_update_method,
soft_update_tau=args.soft_update_tau,
episodic_update=args.episodic_replay,
episodic_update_len=16)
if args.load:
agent.load(args.load)
eval_env = make_env(for_eval=True)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_runs=args.eval_n_runs,
max_episode_len=timestep_limit)
print('n_runs: {} mean: {} median: {} stdev {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_runs=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
eval_env=eval_env,
max_episode_len=timestep_limit)
def main():
# This prevents numpy from using multiple threads
os.environ['OMP_NUM_THREADS'] = '1'
import logging
# logging.basicConfig(level=logging.DEBUG)
parser = argparse.ArgumentParser()
parser.add_argument('processes', type=int)
parser.add_argument('rom', type=str)
parser.add_argument('--seed', type=int, default=None)
parser.add_argument('--steps', type=int, default=8 * 10**7)
parser.add_argument('--use-sdl', action='store_true', default=False)
parser.add_argument('--final-exploration-frames',
type=int,
default=4 * 10**6)
parser.add_argument('--outdir', type=str, default='nsq_output')
parser.add_argument('--profile', action='store_true')
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=None)
args = parser.parse_args()
if args.seed is not None:
misc.set_random_seed(args.seed)
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print('Output files are saved in {}'.format(args.outdir))
def make_env(process_idx, test):
env = ale.ALE(args.rom,
use_sdl=args.use_sdl,
treat_life_lost_as_terminal=not test)
if not test:
misc.env_modifiers.make_reward_clipped(env, -1, 1)
return env
sample_env = make_env(0, test=False)
action_space = sample_env.action_space
assert isinstance(action_space, spaces.Discrete)
# Define a model and its optimizer
q_func = links.Sequence(links.NIPSDQNHead(), L.Linear(256, action_space.n),
DiscreteActionValue)
opt = rmsprop_async.RMSpropAsync(lr=7e-4, eps=1e-1, alpha=0.99)
opt.setup(q_func)
# Make process-specific agents to diversify exploration
def make_agent(process_idx):
# Random epsilon assignment described in the original paper
rand = random.random()
if rand < 0.4:
epsilon_target = 0.1
elif rand < 0.7:
epsilon_target = 0.01
else:
epsilon_target = 0.5
explorer = explorers.LinearDecayEpsilonGreedy(
1, epsilon_target, args.final_exploration_frames,
action_space.sample)
# Suppress the explorer logger
explorer.logger.setLevel(logging.INFO)
return nsq.NSQ(q_func,
opt,
t_max=5,
gamma=0.99,
i_target=40000,
explorer=explorer,
phi=dqn_phi)
if args.demo:
env = make_env(0, True)
agent = make_agent(0)
eval_stats = experiments.eval_performance(env=env,
agent=agent,
n_runs=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:
explorer = explorers.ConstantEpsilonGreedy(0.05, action_space.sample)
experiments.train_agent_async(outdir=args.outdir,
processes=args.processes,
make_env=make_env,
make_agent=make_agent,
profile=args.profile,
steps=args.steps,
eval_n_runs=args.eval_n_runs,
eval_interval=args.eval_interval,
eval_explorer=explorer)
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('--logger-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.logger_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 ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
# 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 chainerrl.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.tags.get(
'wrapper_config.TimeLimit.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
winit = chainer.initializers.GlorotUniform()
winit_policy_output = chainer.initializers.GlorotUniform()
def squashed_diagonal_gaussian_head(x):
assert x.shape[-1] == action_size * 2
mean, log_scale = F.split_axis(x, 2, axis=1)
log_scale = F.clip(log_scale, -20., 2.)
var = F.exp(log_scale * 2)
return chainerrl.distribution.SquashedGaussianDistribution(mean,
var=var)
policy = chainer.Sequential(
L.Linear(None, args.n_hidden_channels, initialW=winit),
F.relu,
L.Linear(None, args.n_hidden_channels, initialW=winit),
F.relu,
L.Linear(None, action_size * 2, initialW=winit_policy_output),
squashed_diagonal_gaussian_head,
)
policy_optimizer = optimizers.Adam(args.lr,
eps=args.adam_eps).setup(policy)
def make_q_func_with_optimizer():
q_func = chainer.Sequential(
concat_obs_and_action,
L.Linear(None, args.n_hidden_channels, initialW=winit),
F.relu,
L.Linear(None, args.n_hidden_channels, initialW=winit),
F.relu,
L.Linear(None, 1, initialW=winit),
)
q_func_optimizer = optimizers.Adam(args.lr,
eps=args.adam_eps).setup(q_func)
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()
# Draw the computational graph and save it in the output directory.
fake_obs = chainer.Variable(policy.xp.zeros_like(obs_space.low,
dtype=np.float32)[None],
name='observation')
fake_action = chainer.Variable(policy.xp.zeros_like(
action_space.low, dtype=np.float32)[None],
name='action')
chainerrl.misc.draw_computational_graph([policy(fake_obs)],
os.path.join(
args.outdir, 'policy'))
chainerrl.misc.draw_computational_graph([q_func1(fake_obs, fake_action)],
os.path.join(
args.outdir, 'q_func1'))
chainerrl.misc.draw_computational_graph([q_func2(fake_obs, fake_action)],
os.path.join(
args.outdir, 'q_func2'))
rbuf = replay_buffer.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 = chainerrl.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=chainer.optimizers.Adam(args.lr,
eps=args.adam_eps),
)
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():
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('--logging-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.logging_level)
# Set a random seed used in ChainerRL.
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
# 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 = chainerrl.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 = chainerrl.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_actions = env.action_space.n
if args.recurrent:
# Q-network with LSTM
q_func = chainerrl.links.StatelessRecurrentSequential(
L.Convolution2D(None, 32, 8, stride=4),
F.relu,
L.Convolution2D(None, 64, 4, stride=2),
F.relu,
L.Convolution2D(None, 64, 3, stride=1),
functools.partial(F.reshape, shape=(-1, 3136)),
F.relu,
L.NStepLSTM(1, 3136, 512, 0),
L.Linear(None, n_actions),
DiscreteActionValue,
)
# Replay buffer that stores whole episodes
rbuf = replay_buffer.EpisodicReplayBuffer(10**6)
else:
# Q-network without LSTM
q_func = chainer.Sequential(
L.Convolution2D(None, 32, 8, stride=4),
F.relu,
L.Convolution2D(None, 64, 4, stride=2),
F.relu,
L.Convolution2D(None, 64, 3, stride=1),
functools.partial(F.reshape, shape=(-1, 3136)),
L.Linear(None, 512),
F.relu,
L.Linear(None, n_actions),
DiscreteActionValue,
)
# Replay buffer that stores transitions separately
rbuf = replay_buffer.ReplayBuffer(10**6)
# Draw the computational graph and save it in the output directory.
fake_obss = np.zeros(env.observation_space.shape, dtype=np.float32)[None]
if args.recurrent:
fake_out, _ = q_func(fake_obss, None)
else:
fake_out = q_func(fake_obss)
chainerrl.misc.draw_computational_graph([fake_out],
os.path.join(args.outdir, 'model'))
explorer = explorers.LinearDecayEpsilonGreedy(
1.0, args.final_epsilon, args.final_exploration_frames,
lambda: np.random.randint(n_actions))
opt = chainer.optimizers.Adam(1e-4, eps=1e-4)
opt.setup(q_func)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
agent = chainerrl.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,
)
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('--use-sdl', action='store_true', default=False)
parser.add_argument('--final-exploration-frames', type=int, default=10**6)
parser.add_argument('--final-epsilon', type=float, default=0.1)
parser.add_argument('--eval-epsilon', type=float, default=0.05)
parser.add_argument('--steps', type=int, default=10**7)
parser.add_argument(
'--max-frames',
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help='Maximum number of frames for each episode.')
parser.add_argument('--replay-start-size', type=int, default=5 * 10**4)
parser.add_argument('--target-update-interval', type=int, default=10**4)
parser.add_argument('--eval-interval', type=int, default=10**5)
parser.add_argument('--update-interval', type=int, default=4)
parser.add_argument('--eval-n-runs', type=int, default=10)
parser.add_argument('--batch-size', type=int, default=32)
parser.add_argument('--logging-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.logging_level)
# Set a random seed used in ChainerRL.
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print('Output files are saved in {}'.format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(atari_wrappers.make_atari(
args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = chainerrl.wrappers.RandomizeAction(env, args.eval_epsilon)
if args.monitor:
env = chainerrl.wrappers.Monitor(
env, args.outdir, mode='evaluation' if test else 'training')
if args.render:
env = chainerrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
n_actions = env.action_space.n
n_atoms = 51
v_max = 10
v_min = -10
q_func = chainerrl.links.Sequence(
chainerrl.links.NatureDQNHead(),
chainerrl.q_functions.DistributionalFCStateQFunctionWithDiscreteAction(
None,
n_actions,
n_atoms,
v_min,
v_max,
n_hidden_channels=0,
n_hidden_layers=0),
)
# Draw the computational graph and save it in the output directory.
chainerrl.misc.draw_computational_graph(
[q_func(np.zeros((4, 84, 84), dtype=np.float32)[None])],
os.path.join(args.outdir, 'model'))
# Use the same hyper parameters as https://arxiv.org/abs/1707.06887
opt = chainer.optimizers.Adam(2.5e-4, eps=1e-2 / args.batch_size)
opt.setup(q_func)
rbuf = replay_buffer.ReplayBuffer(10**6)
explorer = explorers.LinearDecayEpsilonGreedy(
1.0, args.final_epsilon, args.final_exploration_frames,
lambda: np.random.randint(n_actions))
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
agent = chainerrl.agents.CategoricalDQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
batch_accumulator='mean',
phi=phi,
)
if args.load:
agent.load(args.load)
if args.demo:
eval_stats = experiments.eval_performance(env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs)
print('n_runs: {} mean: {} median: {} stdev {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=False,
eval_env=eval_env,
)
def main():
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('--load', type=str, default=None, required=True)
parser.add_argument('--logging-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 demo timesteps to collect')
args = parser.parse_args()
import logging
logging.basicConfig(level=args.logging_level)
# Set a random seed used in ChainerRL.
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print('Output files are saved in {}'.format(args.outdir))
def make_env():
env = atari_wrappers.wrap_deepmind(atari_wrappers.make_atari(
args.env, max_frames=None),
episode_life=False,
clip_rewards=False)
env.seed(int(args.seed))
# Randomize actions like epsilon-greedy
env = chainerrl.wrappers.RandomizeAction(env, 0.01)
if args.monitor:
env = gym.wrappers.Monitor(env, args.outdir, mode='evaluation')
if args.render:
env = chainerrl.wrappers.Render(env)
return env
env = make_env()
n_actions = env.action_space.n
q_func = links.Sequence(links.NatureDQNHead(), L.Linear(512, n_actions),
DiscreteActionValue)
# Draw the computational graph and save it in the output directory.
chainerrl.misc.draw_computational_graph(
[q_func(np.zeros((4, 84, 84), dtype=np.float32)[None])],
os.path.join(args.outdir, 'model'))
# The optimizer and replay buffer are dummy variables required by agent
opt = optimizers.RMSpropGraves()
opt.setup(q_func)
rbuf = replay_buffer.ReplayBuffer(1)
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=None,
replay_start_size=1,
minibatch_size=1,
target_update_interval=None,
clip_delta=True,
update_interval=4,
phi=phi)
agent.load(args.load)
# saves demos to outdir/demos.pickle
experiments.collect_demonstrations(agent=agent,
env=env,
steps=args.steps,
episodes=None,
outdir=args.outdir,
max_episode_len=None)
