def test_F3_oh_value():
for i in range(3):
ll_runs = 1
steps = 10000
ep_s = ExponentialDecay(steps / 16, 0.5, 0.05, steps)
lr_s = ConstantSched(0.05)
device = 'cuda'
actions = 3
obs_shape = (1, )
batch_size = 32
env = gym.make('F3-v0')
#env = RewardPerStep(env, reward_per_step=-0.01)
env = TimeLimit(env, max_episode_steps=20)
env = NormalizeFunctional(env,
obs_f=normalize_obs,
reward_f=normalize_reward)
env = LookAhead(env)
env = Reset(env)
#env = Monitor(env)
env = BatchTensor(env, device='cuda')
#critic = FixupV(obs_shape, 4).to(device)
critic = OneHotV(obs_shape, 12).to(device)
policy = VPolicy(critic,
actions,
EpsilonGreedyProperDiscreteDist,
epsilon=1.0).to(device)
exp_buffer = ExpBuffer(max_timesteps=steps // 10,
ll_runs=ll_runs,
batch_size=batch_size,
observation_shape=obs_shape)
stepper = td_value.Stepper(env, OneObsToState(), exp_buffer)
run_on(stepper=stepper,
learner=td_value.train_one_value,
env=env,
critic=critic,
policy=policy,
ll_runs=ll_runs,
eps_sched=ep_s,
actions=actions,
exp_buffer=exp_buffer,
batch_size=batch_size,
discount=0.8,
lr_sched=lr_s,
rendermode='episodic',
steps=steps,
logging_freq=1,
run_id=f'f5_value_{i}',
warmup_steps=0)
def __init__(self,
env,
critic,
behaviour_policy,
greedy_policy,
exp_buffer,
join=OneObsToState(),
device='cuda',
viewer=None,
plot=None):
self.env = env
self.state = self.env.reset()
self.join = join
self.exp_buffer = exp_buffer
self.v = viewer
self.tb = None
self.critic = critic
self.behaviour_policy = behaviour_policy
self.greedy_policy = greedy_policy
self.device = device
self.plotter = plot