def set_env(config: Config, expand: ImageSetting) -> None:
# ppo parameters
config.nworkers = 32
config.nsteps = 125
config.value_loss_weight = 0.5
config.entropy_weight = 0.01
config.gae_tau = 0.95
config.use_gae = True
config.ppo_minibatch_size = 400
config.ppo_clip = 0.1
config.lr_decay = False
config.set_parallel_env(lambda _env_gen, _num_w: ParallelRogueEnvExt(
StairRewardParallel(
[CONFIG] * config.nworkers,
max_steps=500,
stair_reward=50.0,
image_setting=expand,
)))
config.eval_env = RogueEnvExt(
StairRewardEnv(
RogueEnv(config_dict=CONFIG,
mex_steps=500,
stair_reward=50.0,
image_setting=expand), 100.0))
config.max_steps = int(2e7) * 2
config.eval_freq = None
config.save_freq = int(2e6)
def main(
envname: str = "Hopper",
tau: float = 12 * 20,
update_freq: int = 10,
) -> Config:
c = Config()
c.max_steps = int(4e5)
c.nworkers = 12
c.nsteps = 20
c.set_env(lambda: PyBullet(envname))
c.set_net_fn("actor-critic",
net.actor_critic.fc_shared(policy=SeparateStdGaussianDist))
c.set_parallel_env(pybullet_parallel())
c.set_optimizer(kfac.default_sgd(eta_max=0.1))
c.set_preconditioner(lambda net: kfac.KfacPreConditioner(
net,
tau=tau,
update_freq=update_freq,
norm_scaler=kfac.SquaredFisherScaler(eta_max=0.1, delta=0.001),
))
c.gae_lambda = 0.95
c.use_gae = True
c.eval_deterministic = False
c.value_loss_weight = 0.5
c.entropy_weight = 0.0
c.eval_freq = None
return c
def config() -> Config:
c = Config()
c.max_steps = int(4e5)
c.nworkers = 12
c.nsteps = 20
c.set_parallel_env(MultiProcEnv)
c.set_optimizer(kfac.default_sgd(eta_max=0.1))
c.set_preconditioner(
lambda net: kfac.KfacPreConditioner(net, **KFAC_KWARGS))
c.gae_lambda = 0.95
c.use_gae = False
c.lr_min = 0.0
c.value_loss_weight = 0.1
c.entropy_weight = 0.01
c.eval_freq = None
return c
def config() -> Config:
c = Config()
c.set_env(lambda: PyBullet('Hopper'))
c.set_net_fn('actor-critic',
net.actor_critic.fc_shared(policy=SeparateStdGaussinanHead))
c.set_parallel_env(pybullet_parallel())
c.max_steps = int(1e6)
c.nworkers = 12
c.nsteps = 5
c.set_optimizer(lambda params: Adam(params, lr=0.001))
c.grad_clip = 0.5
c.gae_lambda = 0.95
c.value_loss_weight = 0.5
c.entropy_weight = 0.0
c.use_gae = False
c.eval_deterministic = False
c.eval_freq = None
return c
def config() -> Config:
c = Config()
c.max_steps = int(4e5)
c.nworkers = 12
c.nsteps = 20
c.set_env(lambda: PyBullet('Hopper'))
c.set_net_fn('actor-critic',
net.actor_critic.fc_shared(policy=SeparateStdGaussinanHead))
c.set_parallel_env(pybullet_parallel())
c.set_optimizer(kfac.default_sgd(eta_max=0.1))
c.set_preconditioner(
lambda net: kfac.KfacPreConditioner(net, **KFAC_KWARGS))
c.gae_lambda = 0.95
c.use_gae = True
c.eval_deterministic = False
c.value_loss_weight = 0.5
c.entropy_weight = 0.0
c.eval_freq = None
return c
def config() -> Config:
c = Config()
c.set_env(lambda: PyBullet('Hopper'))
c.set_net_fn('actor-critic',
net.actor_critic.fc_shared(policy=SeparateStdGaussinanHead))
c.set_parallel_env(pybullet_parallel())
c.set_optimizer(lambda params: Adam(params, lr=3.0e-4, eps=1.0e-4))
c.max_steps = int(2e6)
c.grad_clip = 0.5
# ppo params
c.value_loss_weight = 0.5
c.entropy_weight = 0.0
c.gae_lambda = 0.95
c.nworkers = 4
c.nsteps = 512
c.ppo_minibatch_size = (4 * 512) // 8
c.ppo_clip = 0.2
c.use_gae = True
c.use_reward_monitor = True
c.eval_freq = None
return c
