def test_add_transition_dtype(self):
env = DummyDiscreteEnv()
obs = env.reset()
replay_buffer = SimpleReplayBuffer(env_spec=env,
size_in_transitions=3,
time_horizon=1)
replay_buffer.add_transition(observation=obs,
action=env.action_space.sample())
sample = replay_buffer.sample(1)
sample_obs = sample['observation']
sample_action = sample['action']
assert sample_obs.dtype == env.observation_space.dtype
assert sample_action.dtype == env.action_space.dtype
def test_pickleable(self):
env = DummyDiscreteEnv()
obs = env.reset()
replay_buffer = SimpleReplayBuffer(env_spec=env,
size_in_transitions=100,
time_horizon=1)
for _ in range(0, 100):
replay_buffer.add_transitions(observation=[obs], action=[1])
replay_buffer_pickled = pickle.loads(pickle.dumps(replay_buffer))
assert replay_buffer_pickled._buffer.keys(
) == replay_buffer._buffer.keys()
for k in replay_buffer_pickled._buffer:
assert replay_buffer_pickled._buffer[
k].shape == replay_buffer._buffer[k].shape
def test_ddpg_double_pendulum(self):
"""Test DDPG with Pendulum environment."""
with LocalTFRunner(snapshot_config, sess=self.sess) as runner:
env = TfEnv(gym.make('InvertedDoublePendulum-v2'))
action_noise = OUStrategy(env.spec, sigma=0.2)
policy = ContinuousMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e5),
time_horizon=100)
algo = DDPG(
env_spec=env.spec,
policy=policy,
policy_lr=1e-4,
qf_lr=1e-3,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=20,
target_update_tau=1e-2,
n_train_steps=50,
discount=0.9,
min_buffer_size=int(5e3),
exploration_strategy=action_noise,
)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 60
env.close()
def run_metarl(env, seed, log_dir):
'''
Create metarl model and training.
Replace the ddpg with the algorithm you want to run.
:param env: Environment of the task.
:param seed: Random seed for the trial.
:param log_dir: Log dir path.
:return:
'''
deterministic.set_seed(seed)
with LocalTFRunner(snapshot_config) as runner:
env = TfEnv(normalize(env))
# Set up params for ddpg
action_noise = OUStrategy(env.spec, sigma=params['sigma'])
policy = ContinuousMLPPolicy(
env_spec=env.spec,
hidden_sizes=params['policy_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=params['qf_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu)
replay_buffer = SimpleReplayBuffer(
env_spec=env.spec,
size_in_transitions=params['replay_buffer_size'],
time_horizon=params['n_rollout_steps'])
ddpg = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=params['steps_per_epoch'],
policy_lr=params['policy_lr'],
qf_lr=params['qf_lr'],
target_update_tau=params['tau'],
n_train_steps=params['n_train_steps'],
discount=params['discount'],
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
policy_optimizer=tf.train.AdamOptimizer,
qf_optimizer=tf.train.AdamOptimizer)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
tensorboard_log_dir = osp.join(log_dir)
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(tensorboard_log_dir))
runner.setup(ddpg, env)
runner.train(n_epochs=params['n_epochs'],
batch_size=params['n_rollout_steps'])
dowel_logger.remove_all()
return tabular_log_file
def run_task(snapshot_config, *_):
"""Wrap TD3 training task in the run_task function.
Args:
snapshot_config (metarl.experiment.SnapshotConfig): Configuration
values for snapshotting.
*_ (object): Hyperparameters (unused).
"""
with LocalTFRunner(snapshot_config) as runner:
env = TfEnv(gym.make('InvertedDoublePendulum-v2'))
action_noise = GaussianStrategy(env.spec, max_sigma=0.1, min_sigma=0.1)
policy = ContinuousMLPPolicy(env_spec=env.spec,
hidden_sizes=[400, 300],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
qf = ContinuousMLPQFunction(name='ContinuousMLPQFunction',
env_spec=env.spec,
hidden_sizes=[400, 300],
action_merge_layer=0,
hidden_nonlinearity=tf.nn.relu)
qf2 = ContinuousMLPQFunction(name='ContinuousMLPQFunction2',
env_spec=env.spec,
hidden_sizes=[400, 300],
action_merge_layer=0,
hidden_nonlinearity=tf.nn.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=250)
td3 = TD3(env_spec=env.spec,
policy=policy,
policy_lr=1e-4,
qf_lr=1e-3,
qf=qf,
qf2=qf2,
replay_buffer=replay_buffer,
target_update_tau=1e-2,
steps_per_epoch=20,
n_train_steps=1,
smooth_return=False,
discount=0.99,
buffer_batch_size=100,
min_buffer_size=1e4,
exploration_strategy=action_noise,
policy_optimizer=tf.train.AdamOptimizer,
qf_optimizer=tf.train.AdamOptimizer)
runner.setup(td3, env)
runner.train(n_epochs=500, batch_size=250)
def test_td3_pendulum(self):
"""Test TD3 with Pendulum environment."""
with LocalTFRunner(snapshot_config) as runner:
env = TfEnv(gym.make('InvertedDoublePendulum-v2'))
action_noise = GaussianStrategy(env.spec,
max_sigma=0.1,
min_sigma=0.1)
policy = ContinuousMLPPolicy(env_spec=env.spec,
hidden_sizes=[400, 300],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
qf = ContinuousMLPQFunction(name='ContinuousMLPQFunction',
env_spec=env.spec,
hidden_sizes=[400, 300],
action_merge_layer=0,
hidden_nonlinearity=tf.nn.relu)
qf2 = ContinuousMLPQFunction(name='ContinuousMLPQFunction2',
env_spec=env.spec,
hidden_sizes=[400, 300],
action_merge_layer=0,
hidden_nonlinearity=tf.nn.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=250)
algo = TD3(env_spec=env.spec,
policy=policy,
policy_lr=1e-3,
qf_lr=1e-3,
qf=qf,
qf2=qf2,
replay_buffer=replay_buffer,
steps_per_epoch=20,
target_update_tau=0.005,
n_train_steps=50,
discount=0.99,
smooth_return=False,
min_buffer_size=int(1e4),
buffer_batch_size=100,
policy_weight_decay=0.001,
qf_weight_decay=0.001,
exploration_strategy=action_noise,
policy_optimizer=tf.compat.v1.train.AdamOptimizer,
qf_optimizer=tf.compat.v1.train.AdamOptimizer)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=250)
assert last_avg_ret > 400
def test_dqn_cartpole_pickle(self):
"""Test DQN with CartPole environment."""
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 = 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,
steps_per_epoch=steps_per_epoch,
target_network_update_freq=1,
buffer_batch_size=32)
runner.setup(algo, env)
with tf.compat.v1.variable_scope(
'DiscreteMLPQFunction/MLPModel/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/MLPModel/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 run_task(snapshot_config, *_):
"""Set up environment and algorithm and run the task."""
runner = LocalRunner(snapshot_config)
env = MetaRLEnv(normalize(gym.make('HalfCheetah-v2')))
policy = TanhGaussianMLPPolicy2(
env_spec=env.spec,
hidden_sizes=[256, 256],
hidden_nonlinearity=nn.ReLU,
output_nonlinearity=None,
min_std=np.exp(-20.),
max_std=np.exp(2.),
)
qf1 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[256, 256],
hidden_nonlinearity=F.relu)
qf2 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[256, 256],
hidden_nonlinearity=F.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=1)
# replay_buffer = SACReplayBuffer(env_spec=env.spec,
# max_size=int(1e6))
sampler_args = {
'agent': policy,
'max_path_length': 1000,
}
sac = SAC(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
gradient_steps_per_itr=1000,
use_automatic_entropy_tuning=True,
replay_buffer=replay_buffer,
min_buffer_size=1e4,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=256,
reward_scale=1.)
runner.setup(algo=sac,
env=env,
sampler_cls=SimpleSampler,
sampler_args=sampler_args)
runner.train(n_epochs=1000, batch_size=1000)
def test_eviction_policy(self):
env = DummyDiscreteEnv()
obs = env.reset()
replay_buffer = SimpleReplayBuffer(env_spec=env,
size_in_transitions=3,
time_horizon=1)
replay_buffer.add_transitions(observation=[obs, obs], action=[1, 2])
assert not replay_buffer.full
replay_buffer.add_transitions(observation=[obs, obs], action=[3, 4])
assert replay_buffer.full
replay_buffer.add_transitions(observation=[obs, obs], action=[5, 6])
replay_buffer.add_transitions(observation=[obs, obs], action=[7, 8])
assert np.array_equal(replay_buffer._buffer['action'], [[7], [8], [6]])
assert replay_buffer.n_transitions_stored == 3
def tf_ddpg_pendulum(ctxt=None, seed=1):
"""Train DDPG with InvertedDoublePendulum-v2 environment.
Args:
ctxt (metarl.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(snapshot_config=ctxt) as runner:
env = TfEnv(gym.make('InvertedDoublePendulum-v2'))
action_noise = OUStrategy(env.spec, sigma=0.2)
policy = ContinuousMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=100)
ddpg = DDPG(env_spec=env.spec,
policy=policy,
policy_lr=1e-4,
qf_lr=1e-3,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=20,
target_update_tau=1e-2,
n_train_steps=50,
discount=0.9,
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
policy_optimizer=tf.train.AdamOptimizer,
qf_optimizer=tf.train.AdamOptimizer)
runner.setup(algo=ddpg, env=env)
runner.train(n_epochs=500, batch_size=100)
def run_task(snapshot_config, *_):
"""Set up environment and algorithm and run the task.
Args:
snapshot_config (metarl.experiment.SnapshotConfig): The snapshot
configuration used by LocalRunner to create the snapshotter.
If None, it will create one with default settings.
_ : Unused parameters
"""
runner = LocalRunner(snapshot_config)
env = MetaRLEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
action_noise = OUStrategy(env.spec, sigma=0.2)
policy = DeterministicMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu,
output_nonlinearity=torch.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=100)
policy_optimizer = (torch.optim.Adagrad, {'lr': 1e-4, 'lr_decay': 0.99})
ddpg = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=20,
n_train_steps=50,
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
target_update_tau=1e-2,
discount=0.9,
policy_optimizer=policy_optimizer,
qf_optimizer=torch.optim.Adam)
runner.setup(algo=ddpg, env=env)
runner.train(n_epochs=500, batch_size=100)
def test_algo_with_goal_without_es(self):
# This tests if sampler works properly when algorithm
# includes goal but is without exploration policy
env = DummyDictEnv()
policy = DummyPolicy(env)
replay_buffer = SimpleReplayBuffer(env_spec=env,
size_in_transitions=int(1e6),
time_horizon=100)
algo = DummyOffPolicyAlgo(env_spec=env,
qf=None,
replay_buffer=replay_buffer,
policy=policy,
exploration_strategy=None,
input_include_goal=True)
sampler = OffPolicyVectorizedSampler(algo, env, 1, no_reset=True)
sampler.start_worker()
sampler.obtain_samples(0, 30)
def run_task(snapshot_config, *_):
"""Run task.
Args:
snapshot_config (metarl.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_grad_clip(self):
"""Test DQN with CartPole environment."""
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 = 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,
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 > 15
env.close()
def test_ddpg_pendulum(self):
"""Test DDPG with Pendulum environment.
This environment has a [-3, 3] action_space bound.
"""
deterministic.set_seed(0)
runner = LocalRunner(snapshot_config)
env = MetaRLEnv(normalize(gym.make('InvertedPendulum-v2')))
action_noise = OUStrategy(env.spec, sigma=0.2)
policy = DeterministicMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu,
output_nonlinearity=torch.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=100)
algo = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=20,
n_train_steps=50,
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
target_update_tau=1e-2,
discount=0.9)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 10
env.close()
def run_metarl(env, seed, log_dir):
"""Create metarl model and training.
Replace the ddpg with the algorithm you want to run.
Args:
env (gym.Env): Environment of the task.
seed (int): Random seed for the trial.
log_dir (str): Log dir path.
Returns:
str: Log file path.
"""
deterministic.set_seed(seed)
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=12,
inter_op_parallelism_threads=12)
sess = tf.Session(config=config)
with LocalTFRunner(snapshot_config, sess=sess, max_cpus=12) as runner:
env = TfEnv(normalize(env))
# Set up params for ddpg
action_noise = OUStrategy(env.spec, sigma=params['sigma'])
policy = ContinuousMLPPolicy(
env_spec=env.spec,
name='ContinuousMLPPolicy',
hidden_sizes=params['policy_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=params['qf_hidden_sizes'],
hidden_nonlinearity=tf.nn.relu,
name='ContinuousMLPQFunction')
replay_buffer = SimpleReplayBuffer(
env_spec=env.spec,
size_in_transitions=params['replay_buffer_size'],
time_horizon=params['n_rollout_steps'])
ddpg = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=params['steps_per_epoch'],
policy_lr=params['policy_lr'],
qf_lr=params['qf_lr'],
target_update_tau=params['tau'],
n_train_steps=params['n_train_steps'],
discount=params['discount'],
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
policy_optimizer=tf.train.AdamOptimizer,
qf_optimizer=tf.train.AdamOptimizer)
# Set up logger since we are not using run_experiment
tabular_log_file = osp.join(log_dir, 'progress.csv')
dowel_logger.add_output(dowel.StdOutput())
dowel_logger.add_output(dowel.CsvOutput(tabular_log_file))
dowel_logger.add_output(dowel.TensorBoardOutput(log_dir))
runner.setup(ddpg, env, sampler_args=dict(n_envs=12))
runner.train(n_epochs=params['n_epochs'],
batch_size=params['n_rollout_steps'])
dowel_logger.remove_all()
return tabular_log_file
def test_no_reset(self):
with LocalTFRunner(snapshot_config, sess=self.sess) as runner:
# This tests if off-policy sampler respect batch_size
# when no_reset is set to True
env = TfEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
action_noise = OUStrategy(env.spec, sigma=0.2)
policy = ContinuousMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=100)
algo = DDPG(
env_spec=env.spec,
policy=policy,
policy_lr=1e-4,
qf_lr=1e-3,
qf=qf,
replay_buffer=replay_buffer,
target_update_tau=1e-2,
n_train_steps=50,
discount=0.9,
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
)
sampler = OffPolicyVectorizedSampler(algo, env, 1, no_reset=True)
sampler.start_worker()
runner.initialize_tf_vars()
paths1 = sampler.obtain_samples(0, 5)
paths2 = sampler.obtain_samples(0, 5)
len1 = sum([len(path['rewards']) for path in paths1])
len2 = sum([len(path['rewards']) for path in paths2])
assert len1 == 5 and len2 == 5, 'Sampler should respect batch_size'
# yapf: disable
# When done is False in 1st sampling, the next sampling should be
# stacked with the last batch in 1st sampling
case1 = (len(paths1[-1]['rewards']) + len(paths2[0]['rewards'])
== paths2[0]['running_length'])
# When done is True in 1st sampling, the next sampling should be
# separated
case2 = len(paths2[0]['rewards']) == paths2[0]['running_length']
done = paths1[-1]['dones'][-1]
assert (
(not done and case1) or (done and case2)
), 'Running length should be the length of full path'
# yapf: enable
case1 = np.isclose(
paths1[-1]['rewards'].sum() + paths2[0]['rewards'].sum(),
paths2[0]['undiscounted_return'])
case2 = np.isclose(paths2[0]['rewards'].sum(),
paths2[0]['undiscounted_return'])
assert (
(not done and case1) or (done and case2)
), 'Undiscounted_return should be the sum of rewards of full path'
def run_task(snapshot_config, variant_data, *_):
"""Run task.
Args:
snapshot_config (metarl.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)
