def run_task(snapshot_config, *_):
"""Run task."""
with LocalRunner(snapshot_config=snapshot_config) as runner:
n_epochs = 100
n_epoch_cycles = 20
sampler_batch_size = 500
num_timesteps = n_epochs * n_epoch_cycles * sampler_batch_size
env = gym.make('PongNoFrameskip-v4')
env = Noop(env, noop_max=30)
env = MaxAndSkip(env, skip=4)
env = EpisodicLife(env)
if 'FIRE' in env.unwrapped.get_action_meanings():
env = FireReset(env)
env = Grayscale(env)
env = Resize(env, 84, 84)
env = ClipReward(env)
env = StackFrames(env, 4)
env = TfEnv(env)
replay_buffer = SimpleReplayBuffer(
env_spec=env.spec, size_in_transitions=int(5e4), time_horizon=1)
qf = DiscreteCNNQFunction(
env_spec=env.spec,
filter_dims=(8, 4, 3),
num_filters=(32, 64, 64),
strides=(4, 2, 1),
dueling=False)
policy = DiscreteQfDerivedPolicy(env_spec=env.spec, qf=qf)
epilson_greedy_strategy = EpsilonGreedyStrategy(
env_spec=env.spec,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
algo = DQN(
env_spec=env.spec,
policy=policy,
qf=qf,
exploration_strategy=epilson_greedy_strategy,
replay_buffer=replay_buffer,
qf_lr=1e-4,
discount=0.99,
min_buffer_size=int(1e4),
double_q=False,
n_train_steps=500,
n_epoch_cycles=n_epoch_cycles,
target_network_update_freq=2,
buffer_batch_size=32)
runner.setup(algo, env)
runner.train(
n_epochs=n_epochs,
n_epoch_cycles=n_epoch_cycles,
batch_size=sampler_batch_size)
def test_dqn_cartpole_pickle(self):
"""Test DQN with CartPole environment."""
deterministic.set_seed(100)
with TFTrainer(snapshot_config, sess=self.sess) as trainer:
n_epochs = 10
steps_per_epoch = 10
sampler_batch_size = 500
num_timesteps = n_epochs * steps_per_epoch * sampler_batch_size
env = GymEnv('CartPole-v0')
replay_buffer = PathBuffer(capacity_in_transitions=int(1e4))
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(64, 64))
policy = DiscreteQFArgmaxPolicy(env_spec=env.spec, qf=qf)
epilson_greedy_policy = EpsilonGreedyPolicy(
env_spec=env.spec,
policy=policy,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
sampler = LocalSampler(
agents=epilson_greedy_policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
is_tf_worker=True,
worker_class=FragmentWorker)
algo = DQN(env_spec=env.spec,
policy=policy,
qf=qf,
exploration_policy=epilson_greedy_policy,
replay_buffer=replay_buffer,
sampler=sampler,
qf_lr=1e-4,
discount=1.0,
min_buffer_size=int(1e3),
double_q=False,
n_train_steps=500,
grad_norm_clipping=5.0,
steps_per_epoch=steps_per_epoch,
target_network_update_freq=1,
buffer_batch_size=32)
trainer.setup(algo, env)
with tf.compat.v1.variable_scope(
'DiscreteMLPQFunction/mlp/hidden_0', reuse=True):
bias = tf.compat.v1.get_variable('bias')
# assign it to all one
old_bias = tf.ones_like(bias).eval()
bias.load(old_bias)
h = pickle.dumps(algo)
with tf.compat.v1.Session(graph=tf.Graph()):
pickle.loads(h)
with tf.compat.v1.variable_scope(
'DiscreteMLPQFunction/mlp/hidden_0', reuse=True):
new_bias = tf.compat.v1.get_variable('bias')
new_bias = new_bias.eval()
assert np.array_equal(old_bias, new_bias)
env.close()
def test_dqn_cartpole_pickle(self):
"""Test DQN with CartPole environment."""
with LocalRunner(self.sess) as runner:
n_epochs = 10
n_epoch_cycles = 10
sampler_batch_size = 500
num_timesteps = n_epochs * n_epoch_cycles * sampler_batch_size
env = TfEnv(gym.make('CartPole-v0'))
replay_buffer = SimpleReplayBuffer(
env_spec=env.spec,
size_in_transitions=int(1e4),
time_horizon=1)
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(64, 64))
policy = DiscreteQfDerivedPolicy(env_spec=env.spec, qf=qf)
epilson_greedy_strategy = EpsilonGreedyStrategy(
env_spec=env.spec,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
algo = DQN(
env_spec=env.spec,
policy=policy,
qf=qf,
exploration_strategy=epilson_greedy_strategy,
replay_buffer=replay_buffer,
qf_lr=1e-4,
discount=1.0,
min_buffer_size=int(1e3),
double_q=False,
n_train_steps=500,
grad_norm_clipping=5.0,
n_epoch_cycles=n_epoch_cycles,
target_network_update_freq=1,
buffer_batch_size=32)
runner.setup(algo, env)
with tf.variable_scope(
'DiscreteMLPQFunction/MLPModel/mlp/hidden_0', reuse=True):
bias = tf.get_variable('bias')
# assign it to all one
old_bias = tf.ones_like(bias).eval()
bias.load(old_bias)
h = pickle.dumps(algo)
with tf.Session(graph=tf.Graph()):
pickle.loads(h)
with tf.variable_scope(
'DiscreteMLPQFunction/MLPModel/mlp/hidden_0',
reuse=True):
new_bias = tf.get_variable('bias')
new_bias = new_bias.eval()
assert np.array_equal(old_bias, new_bias)
env.close()
def dqn_cartpole(ctxt=None, seed=1):
"""Train TRPO with CubeCrash-v0 environment.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by Trainer to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
"""
set_seed(seed)
with TFTrainer(ctxt) as trainer:
n_epochs = 10
steps_per_epoch = 10
sampler_batch_size = 500
num_timesteps = n_epochs * steps_per_epoch * sampler_batch_size
env = GymEnv('CartPole-v0')
replay_buffer = PathBuffer(capacity_in_transitions=int(1e4))
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(64, 64))
policy = DiscreteQFArgmaxPolicy(env_spec=env.spec, qf=qf)
exploration_policy = EpsilonGreedyPolicy(env_spec=env.spec,
policy=policy,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
sampler = LocalSampler(agents=exploration_policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
is_tf_worker=True,
worker_class=FragmentWorker)
algo = DQN(env_spec=env.spec,
policy=policy,
qf=qf,
exploration_policy=exploration_policy,
replay_buffer=replay_buffer,
sampler=sampler,
steps_per_epoch=steps_per_epoch,
qf_lr=1e-4,
discount=1.0,
min_buffer_size=int(1e3),
double_q=True,
n_train_steps=500,
target_network_update_freq=1,
buffer_batch_size=32)
trainer.setup(algo, env)
trainer.train(n_epochs=n_epochs, batch_size=sampler_batch_size)
def test_dqn_cartpole_grad_clip(self):
"""Test DQN with CartPole environment."""
deterministic.set_seed(100)
with TFTrainer(snapshot_config, sess=self.sess) as trainer:
n_epochs = 10
steps_per_epoch = 10
sampler_batch_size = 500
num_timesteps = n_epochs * steps_per_epoch * sampler_batch_size
env = GymEnv('CartPole-v0')
replay_buffer = PathBuffer(capacity_in_transitions=int(1e4))
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(64, 64))
policy = DiscreteQFArgmaxPolicy(env_spec=env.spec, qf=qf)
epilson_greedy_policy = EpsilonGreedyPolicy(
env_spec=env.spec,
policy=policy,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
sampler = LocalSampler(
agents=epilson_greedy_policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
is_tf_worker=True,
worker_class=FragmentWorker)
algo = DQN(env_spec=env.spec,
policy=policy,
qf=qf,
exploration_policy=epilson_greedy_policy,
replay_buffer=replay_buffer,
sampler=sampler,
qf_lr=1e-4,
discount=1.0,
min_buffer_size=int(1e3),
double_q=False,
n_train_steps=500,
grad_norm_clipping=5.0,
steps_per_epoch=steps_per_epoch,
target_network_update_freq=1,
buffer_batch_size=32)
trainer.setup(algo, env)
last_avg_ret = trainer.train(n_epochs=n_epochs,
batch_size=sampler_batch_size)
assert last_avg_ret > 8.8
env.close()
def dqn_cartpole(ctxt=None, seed=1):
"""Train TRPO with CubeCrash-v0 environment.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
"""
set_seed(seed)
with LocalTFRunner(ctxt) as runner:
n_epochs = 10
steps_per_epoch = 10
sampler_batch_size = 500
num_timesteps = n_epochs * steps_per_epoch * sampler_batch_size
env = TfEnv(gym.make('CartPole-v0'))
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e4),
time_horizon=1)
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(64, 64))
policy = DiscreteQfDerivedPolicy(env_spec=env.spec, qf=qf)
exploration_policy = EpsilonGreedyPolicy(env_spec=env.spec,
policy=policy,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
algo = DQN(env_spec=env.spec,
policy=policy,
qf=qf,
exploration_policy=exploration_policy,
replay_buffer=replay_buffer,
steps_per_epoch=steps_per_epoch,
qf_lr=1e-4,
discount=1.0,
min_buffer_size=int(1e3),
double_q=True,
n_train_steps=500,
target_network_update_freq=1,
buffer_batch_size=32)
runner.setup(algo, env)
runner.train(n_epochs=n_epochs, batch_size=sampler_batch_size)
def test_dqn_cartpole_grad_clip(self):
"""Test DQN with CartPole environment."""
with LocalRunner(self.sess) as runner:
n_epochs = 10
n_epoch_cycles = 10
sampler_batch_size = 500
num_timesteps = n_epochs * n_epoch_cycles * sampler_batch_size
env = TfEnv(gym.make('CartPole-v0'))
replay_buffer = SimpleReplayBuffer(
env_spec=env.spec,
size_in_transitions=int(1e4),
time_horizon=1)
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(64, 64))
policy = DiscreteQfDerivedPolicy(env_spec=env.spec, qf=qf)
epilson_greedy_strategy = EpsilonGreedyStrategy(
env_spec=env.spec,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
algo = DQN(
env_spec=env.spec,
policy=policy,
qf=qf,
exploration_strategy=epilson_greedy_strategy,
replay_buffer=replay_buffer,
qf_lr=1e-4,
discount=1.0,
min_buffer_size=int(1e3),
double_q=False,
n_train_steps=500,
grad_norm_clipping=5.0,
n_epoch_cycles=n_epoch_cycles,
target_network_update_freq=1,
buffer_batch_size=32)
runner.setup(algo, env)
last_avg_ret = runner.train(
n_epochs=n_epochs,
n_epoch_cycles=n_epoch_cycles,
batch_size=sampler_batch_size)
assert last_avg_ret > 20
env.close()
def run_task(snapshot_config, *_):
"""Run task.
Args:
snapshot_config (garage.experiment.SnapshotConfig): The snapshot
configuration used by LocalRunner to create the snapshotter.
*_ (object): Ignored by this function.
"""
with LocalTFRunner(snapshot_config=snapshot_config) as runner:
n_epochs = 10
steps_per_epoch = 10
sampler_batch_size = 500
num_timesteps = n_epochs * steps_per_epoch * sampler_batch_size
env = TfEnv(gym.make('CartPole-v0'))
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e4),
time_horizon=1)
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(64, 64))
policy = DiscreteQfDerivedPolicy(env_spec=env.spec, qf=qf)
epilson_greedy_strategy = EpsilonGreedyStrategy(
env_spec=env.spec,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
algo = DQN(env_spec=env.spec,
policy=policy,
qf=qf,
exploration_strategy=epilson_greedy_strategy,
replay_buffer=replay_buffer,
steps_per_epoch=steps_per_epoch,
qf_lr=1e-4,
discount=1.0,
min_buffer_size=int(1e3),
double_q=True,
n_train_steps=500,
target_network_update_freq=1,
buffer_batch_size=32)
runner.setup(algo, env)
runner.train(n_epochs=n_epochs, batch_size=sampler_batch_size)
def test_dqn_cartpole_double_q(self):
"""Test DQN with CartPole environment."""
deterministic.set_seed(100)
with LocalTFRunner(snapshot_config, sess=self.sess) as runner:
n_epochs = 10
steps_per_epoch = 10
sampler_batch_size = 500
num_timesteps = n_epochs * steps_per_epoch * sampler_batch_size
env = GarageEnv(gym.make('CartPole-v0'))
replay_buffer = PathBuffer(capacity_in_transitions=int(1e4))
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(64, 64))
policy = DiscreteQfDerivedPolicy(env_spec=env.spec, qf=qf)
epilson_greedy_policy = EpsilonGreedyPolicy(
env_spec=env.spec,
policy=policy,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
algo = DQN(env_spec=env.spec,
policy=policy,
qf=qf,
exploration_policy=epilson_greedy_policy,
replay_buffer=replay_buffer,
max_path_length=100,
qf_lr=1e-4,
discount=1.0,
min_buffer_size=int(1e3),
double_q=True,
n_train_steps=500,
steps_per_epoch=steps_per_epoch,
target_network_update_freq=1,
buffer_batch_size=32)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=n_epochs,
batch_size=sampler_batch_size)
assert last_avg_ret > 9
env.close()
def run_task(snapshot_config, *_):
"""Run task."""
with LocalTFRunner(snapshot_config=snapshot_config) as runner:
n_epochs = 500
n_epoch_cycles = 20
sampler_batch_size = 100
num_timesteps = n_epochs * n_epoch_cycles * sampler_batch_size
env_name = 'MountainCar-v0'
env = TfEnv(gym.make(env_name))
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e4),
time_horizon=1)
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(20,))
policy = DiscreteQfDerivedPolicy(env_spec=env.spec, qf=qf)
epilson_greedy_strategy = EpsilonGreedyStrategy(
env_spec=env.spec,
total_timesteps=num_timesteps,
max_epsilon=0.5,
min_epsilon=0.01,
decay_ratio=0.1)
algo = DQN(env_spec=env.spec,
policy=policy,
qf=qf,
exploration_strategy=epilson_greedy_strategy,
replay_buffer=replay_buffer,
qf_lr=1e-3,
discount=0.99,
min_buffer_size=int(1e3),
double_q=False,
n_train_steps=50,
n_epoch_cycles=n_epoch_cycles,
target_network_update_freq=5,
buffer_batch_size=64)
runner.setup(algo, env)
runner.train(n_epochs=n_epochs,
n_epoch_cycles=n_epoch_cycles,
batch_size=sampler_batch_size)
def run_task(*_):
"""Run task."""
with LocalRunner() as runner:
n_epochs = 10
n_epoch_cycles = 10
sampler_batch_size = 500
num_timesteps = n_epochs * n_epoch_cycles * sampler_batch_size
env = TfEnv(gym.make('CartPole-v0'))
replay_buffer = SimpleReplayBuffer(
env_spec=env.spec, size_in_transitions=int(1e4), time_horizon=1)
qf = DiscreteMLPQFunction(env_spec=env.spec, hidden_sizes=(64, 64))
policy = DiscreteQfDerivedPolicy(env_spec=env.spec, qf=qf)
epilson_greedy_strategy = EpsilonGreedyStrategy(
env_spec=env.spec,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
algo = DQN(
env_spec=env.spec,
policy=policy,
qf=qf,
exploration_strategy=epilson_greedy_strategy,
replay_buffer=replay_buffer,
qf_lr=1e-4,
discount=1.0,
min_buffer_size=int(1e3),
double_q=True,
n_train_steps=500,
n_epoch_cycles=n_epoch_cycles,
target_network_update_freq=1,
buffer_batch_size=32)
runner.setup(algo, env)
runner.train(
n_epochs=n_epochs,
n_epoch_cycles=n_epoch_cycles,
batch_size=sampler_batch_size)
def dqn_pong(ctxt=None, seed=1, buffer_size=int(5e4), max_episode_length=500):
"""Train DQN on PongNoFrameskip-v4 environment.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by Trainer to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
buffer_size (int): Number of timesteps to store in replay buffer.
max_episode_length (int): Maximum length of an episode, after which an
episode is considered complete. This is used during testing to
minimize the memory required to store a single episode.
"""
set_seed(seed)
with TFTrainer(ctxt) as trainer:
n_epochs = 100
steps_per_epoch = 20
sampler_batch_size = 500
num_timesteps = n_epochs * steps_per_epoch * sampler_batch_size
env = gym.make('PongNoFrameskip-v4')
env = env.unwrapped
env = Noop(env, noop_max=30)
env = MaxAndSkip(env, skip=4)
env = EpisodicLife(env)
if 'FIRE' in env.unwrapped.get_action_meanings():
env = FireReset(env)
env = Grayscale(env)
env = Resize(env, 84, 84)
env = ClipReward(env)
env = StackFrames(env, 4)
env = GymEnv(env, is_image=True, max_episode_length=max_episode_length)
replay_buffer = PathBuffer(capacity_in_transitions=buffer_size)
qf = DiscreteCNNQFunction(env_spec=env.spec,
filters=(
(32, (8, 8)),
(64, (4, 4)),
(64, (3, 3)),
),
strides=(4, 2, 1),
dueling=False) # yapf: disable
policy = DiscreteQFArgmaxPolicy(env_spec=env.spec, qf=qf)
exploration_policy = EpsilonGreedyPolicy(env_spec=env.spec,
policy=policy,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
algo = DQN(env_spec=env.spec,
policy=policy,
qf=qf,
exploration_policy=exploration_policy,
replay_buffer=replay_buffer,
qf_lr=1e-4,
discount=0.99,
min_buffer_size=int(1e4),
double_q=False,
n_train_steps=500,
steps_per_epoch=steps_per_epoch,
target_network_update_freq=2,
buffer_batch_size=32)
trainer.setup(algo, env)
trainer.train(n_epochs=n_epochs, batch_size=sampler_batch_size)
def run_task(snapshot_config, variant_data, *_):
"""Run task.
Args:
snapshot_config (garage.experiment.SnapshotConfig): The snapshot
configuration used by LocalRunner to create the snapshotter.
variant_data (dict): Custom arguments for the task.
*_ (object): Ignored by this function.
"""
with LocalTFRunner(snapshot_config=snapshot_config) as runner:
n_epochs = 100
steps_per_epoch = 20
sampler_batch_size = 500
num_timesteps = n_epochs * steps_per_epoch * sampler_batch_size
env = gym.make('PongNoFrameskip-v4')
env = Noop(env, noop_max=30)
env = MaxAndSkip(env, skip=4)
env = EpisodicLife(env)
if 'FIRE' in env.unwrapped.get_action_meanings():
env = FireReset(env)
env = Grayscale(env)
env = Resize(env, 84, 84)
env = ClipReward(env)
env = StackFrames(env, 4)
env = TfEnv(env)
replay_buffer = SimpleReplayBuffer(
env_spec=env.spec,
size_in_transitions=variant_data['buffer_size'],
time_horizon=1)
qf = DiscreteCNNQFunction(env_spec=env.spec,
filter_dims=(8, 4, 3),
num_filters=(32, 64, 64),
strides=(4, 2, 1),
dueling=False)
policy = DiscreteQfDerivedPolicy(env_spec=env.spec, qf=qf)
epilson_greedy_strategy = EpsilonGreedyStrategy(
env_spec=env.spec,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
algo = DQN(env_spec=env.spec,
policy=policy,
qf=qf,
exploration_strategy=epilson_greedy_strategy,
replay_buffer=replay_buffer,
qf_lr=1e-4,
discount=0.99,
min_buffer_size=int(1e4),
double_q=False,
n_train_steps=500,
steps_per_epoch=steps_per_epoch,
target_network_update_freq=2,
buffer_batch_size=32)
runner.setup(algo, env)
runner.train(n_epochs=n_epochs, batch_size=sampler_batch_size)
def dqn_pong(ctxt=None, seed=1, buffer_size=int(5e4)):
"""Train DQN on PongNoFrameskip-v4 environment.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
buffer_size (int): Number of timesteps to store in replay buffer.
"""
set_seed(seed)
with LocalTFRunner(ctxt) as runner:
n_epochs = 100
steps_per_epoch = 20
sampler_batch_size = 500
num_timesteps = n_epochs * steps_per_epoch * sampler_batch_size
env = gym.make('PongNoFrameskip-v4')
env = Noop(env, noop_max=30)
env = MaxAndSkip(env, skip=4)
env = EpisodicLife(env)
if 'FIRE' in env.unwrapped.get_action_meanings():
env = FireReset(env)
env = Grayscale(env)
env = Resize(env, 84, 84)
env = ClipReward(env)
env = StackFrames(env, 4)
env = TfEnv(env, is_image=True)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=buffer_size,
time_horizon=1)
qf = DiscreteCNNQFunction(env_spec=env.spec,
filter_dims=(8, 4, 3),
num_filters=(32, 64, 64),
strides=(4, 2, 1),
dueling=False)
policy = DiscreteQfDerivedPolicy(env_spec=env.spec, qf=qf)
exploration_policy = EpsilonGreedyPolicy(env_spec=env.spec,
policy=policy,
total_timesteps=num_timesteps,
max_epsilon=1.0,
min_epsilon=0.02,
decay_ratio=0.1)
algo = DQN(env_spec=env.spec,
policy=policy,
qf=qf,
exploration_policy=exploration_policy,
replay_buffer=replay_buffer,
qf_lr=1e-4,
discount=0.99,
min_buffer_size=int(1e4),
double_q=False,
n_train_steps=500,
steps_per_epoch=steps_per_epoch,
target_network_update_freq=2,
buffer_batch_size=32)
runner.setup(algo, env)
runner.train(n_epochs=n_epochs, batch_size=sampler_batch_size)
