def test_fit_online_pendulum_with_sac():
env = gym.make('Pendulum-v0')
eval_env = gym.make('Pendulum-v0')
algo = SAC()
buffer = ReplayBuffer(1000, env)
algo.fit_online(env,
buffer,
n_epochs=1,
eval_env=eval_env,
logdir='test_data',
tensorboard=False)
def test_fit_online_pendulum_with_sac():
env = gym.make("Pendulum-v0")
eval_env = gym.make("Pendulum-v0")
algo = SAC()
buffer = ReplayBuffer(1000, env)
algo.fit_online(
env,
buffer,
n_steps=500,
eval_env=eval_env,
logdir="test_data",
)
def test_timelimit_aware(timelimit_aware):
env = gym.make("Pendulum-v0")
algo = SAC()
buffer = ReplayBuffer(1000, env)
algo.fit_online(
env,
buffer,
n_steps=500,
logdir="test_data",
timelimit_aware=timelimit_aware,
)
terminal_count = 0
for i in range(len(buffer)):
terminal_count += int(buffer.transitions[i].terminal)
if timelimit_aware:
assert terminal_count == 0
else:
assert terminal_count > 0
import gym
from d3rlpy.algos import SAC
from d3rlpy.online.buffers import ReplayBuffer
env = gym.make('Pendulum-v0')
eval_env = gym.make('Pendulum-v0')
# setup algorithm
sac = SAC(batch_size=100, use_gpu=False)
# replay buffer for experience replay
buffer = ReplayBuffer(maxlen=100000, env=env)
# start training
# probablistic policies does not need explorers
sac.fit_online(env,
buffer,
n_epochs=100,
eval_env=eval_env,
n_steps_per_epoch=1000,
n_updates_per_epoch=100)
import gym
from d3rlpy.algos import SAC
from d3rlpy.online.buffers import ReplayBuffer
env = gym.make('Pendulum-v0')
eval_env = gym.make('Pendulum-v0')
# setup algorithm
sac = SAC(batch_size=100, use_gpu=False)
# replay buffer for experience replay
buffer = ReplayBuffer(maxlen=100000, env=env)
# start training
# probablistic policies does not need explorers
sac.fit_online(env,
buffer,
n_steps=100000,
eval_env=eval_env,
n_steps_per_epoch=1000,
update_start_step=1000)