def test_get_action(self, mock_rand, obs_dim, action_dim, filter_dims,
filter_sizes, strides, padding, hidden_sizes):
mock_rand.return_value = 0
env = TfEnv(DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('metarl.tf.policies.'
'categorical_cnn_policy.MLPModel'),
new=SimpleMLPModel):
with mock.patch(('metarl.tf.policies.'
'categorical_cnn_policy.CNNModel'),
new=SimpleCNNModel):
policy = CategoricalCNNPolicy(env_spec=env.spec,
conv_filters=filter_dims,
conv_filter_sizes=filter_sizes,
conv_strides=strides,
conv_pad=padding,
hidden_sizes=hidden_sizes)
env.reset()
obs, _, _, _ = env.step(1)
action, prob = policy.get_action(obs)
expected_prob = np.full(action_dim, 0.5)
assert env.action_space.contains(action)
assert action == 0
assert np.array_equal(prob['prob'], expected_prob)
actions, probs = policy.get_actions([obs, obs, obs])
for action, prob in zip(actions, probs['prob']):
assert env.action_space.contains(action)
assert action == 0
assert np.array_equal(prob, expected_prob)
def test_dist_info_sym(self, obs_dim, action_dim, filter_dims,
filter_sizes, strides, padding, hidden_sizes):
env = TfEnv(DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('metarl.tf.policies.'
'categorical_cnn_policy.MLPModel'),
new=SimpleMLPModel):
with mock.patch(('metarl.tf.policies.'
'categorical_cnn_policy.CNNModel'),
new=SimpleCNNModel):
policy = CategoricalCNNPolicy(env_spec=env.spec,
conv_filters=filter_dims,
conv_filter_sizes=filter_sizes,
conv_strides=strides,
conv_pad=padding,
hidden_sizes=hidden_sizes)
env.reset()
obs, _, _, _ = env.step(1)
expected_prob = np.full(action_dim, 0.5)
obs_dim = env.spec.observation_space.shape
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, ) + obs_dim)
dist1 = policy.dist_info_sym(state_input, name='policy2')
prob = self.sess.run(dist1['prob'], feed_dict={state_input: [obs]})
assert np.array_equal(prob[0], expected_prob)
def test_clone(self, filters, strides, padding, hidden_sizes):
env = MetaRLEnv(DummyDiscretePixelEnv())
policy = CategoricalCNNPolicy(env_spec=env.spec,
filters=filters,
strides=strides,
padding=padding,
hidden_sizes=hidden_sizes)
policy_clone = policy.clone('CategoricalCNNPolicyClone')
assert policy.env_spec == policy_clone.env_spec
def test_is_pickleable(self, mock_rand, obs_dim, action_dim):
mock_rand.return_value = 0
env = TfEnv(DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('metarl.tf.policies.'
'categorical_cnn_policy.MLPModel'),
new=SimpleMLPModel):
with mock.patch(('metarl.tf.policies.'
'categorical_cnn_policy.CNNModel'),
new=SimpleCNNModel):
policy = CategoricalCNNPolicy(env_spec=env.spec,
conv_filters=(32, ),
conv_filter_sizes=(3, ),
conv_strides=(1, ),
conv_pad='SAME',
hidden_sizes=(4, ))
env.reset()
obs, _, _, _ = env.step(1)
with tf.compat.v1.variable_scope(
'CategoricalCNNPolicy/Sequential/MLPModel', reuse=True):
return_var = tf.compat.v1.get_variable('return_var')
# assign it to all one
return_var.load(tf.ones_like(return_var).eval())
output1 = self.sess.run(policy.model.outputs,
feed_dict={policy.model.input: [obs]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
output2 = sess.run(policy_pickled.model.outputs,
feed_dict={policy_pickled.model.input: [obs]})
assert np.array_equal(output1, output2)
def test_trpo_cnn_cubecrash(self):
with LocalTFRunner(snapshot_config, sess=self.sess) as runner:
env = MetaRLEnv(normalize(gym.make('CubeCrash-v0')))
policy = CategoricalCNNPolicy(env_spec=env.spec,
filters=((32, (8, 8)), (64, (4, 4))),
strides=(4, 2),
padding='VALID',
hidden_sizes=(32, 32))
baseline = GaussianCNNBaseline(
env_spec=env.spec,
regressor_args=dict(filters=((32, (8, 8)), (64, (4, 4))),
strides=(4, 2),
padding='VALID',
hidden_sizes=(32, 32),
use_trust_region=True))
algo = TRPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
gae_lambda=0.98,
max_kl_step=0.01,
policy_ent_coeff=0.0,
flatten_input=False)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=2048)
assert last_avg_ret > -1.5
env.close()
def ppo_memorize_digits(ctxt=None, seed=1, batch_size=4000):
"""Train PPO on MemorizeDigits-v0 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.
batch_size (int): Number of timesteps to use in each training step.
"""
set_seed(seed)
with LocalTFRunner(ctxt) as runner:
env = MetaRLEnv(normalize(gym.make('MemorizeDigits-v0')),
is_image=True)
policy = CategoricalCNNPolicy(env_spec=env.spec,
filters=(
(32, (5, 5)),
(64, (3, 3)),
(64, (2, 2)),
),
strides=(4, 2, 1),
padding='VALID',
hidden_sizes=(256, )) # yapf: disable
baseline = GaussianCNNBaseline(
env_spec=env.spec,
regressor_args=dict(filters=(
(32, (5, 5)),
(64, (3, 3)),
(64, (2, 2)),
),
strides=(4, 2, 1),
padding='VALID',
hidden_sizes=(256, ),
use_trust_region=True)) # yapf: disable
algo = PPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
gae_lambda=0.95,
lr_clip_range=0.2,
policy_ent_coeff=0.0,
optimizer_args=dict(
batch_size=32,
max_epochs=10,
learning_rate=1e-3,
),
flatten_input=False)
runner.setup(algo, env)
runner.train(n_epochs=1000, batch_size=batch_size)
def categorical_cnn_policy(ctxt, env_id, seed):
"""Create Categorical CNN Policy on TF-PPO.
Args:
ctxt (metarl.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the
snapshotter.
env_id (str): Environment id of the task.
seed (int): Random positive integer for the trial.
"""
deterministic.set_seed(seed)
with LocalTFRunner(ctxt, max_cpus=12) as runner:
env = MetaRLEnv(normalize(gym.make(env_id)))
policy = CategoricalCNNPolicy(
env_spec=env.spec,
conv_filters=hyper_params['conv_filters'],
conv_strides=hyper_params['conv_strides'],
conv_pad=hyper_params['conv_pad'],
hidden_sizes=hyper_params['hidden_sizes'])
baseline = GaussianCNNBaseline(
env_spec=env.spec,
regressor_args=dict(
filters=hyper_params['conv_filters'],
strides=hyper_params['conv_strides'],
padding=hyper_params['conv_pad'],
hidden_sizes=hyper_params['hidden_sizes'],
use_trust_region=hyper_params['use_trust_region']))
algo = PPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
gae_lambda=0.95,
lr_clip_range=0.2,
policy_ent_coeff=0.0,
optimizer_args=dict(
batch_size=32,
max_epochs=10,
learning_rate=1e-3,
),
flatten_input=False,
)
runner.setup(algo, env)
runner.train(n_epochs=hyper_params['n_epochs'],
batch_size=hyper_params['batch_size'])
def test_is_pickleable(self):
env = MetaRLEnv(DummyDiscretePixelEnv())
policy = CategoricalCNNPolicy(env_spec=env.spec,
filters=((3, (32, 32)), ),
strides=(1, ),
padding='SAME',
hidden_sizes=(4, ))
obs_var = tf.compat.v1.placeholder(tf.float32,
shape=(None, None) +
env.observation_space.shape,
name='obs')
policy.build(obs_var)
env.reset()
obs, _, _, _ = env.step(1)
with tf.compat.v1.variable_scope(
'CategoricalCNNPolicy/CategoricalCNNModel', reuse=True):
cnn_bias = tf.compat.v1.get_variable('CNNModel/cnn/h0/bias')
bias = tf.compat.v1.get_variable('MLPModel/mlp/hidden_0/bias')
cnn_bias.load(tf.ones_like(cnn_bias).eval())
bias.load(tf.ones_like(bias).eval())
output1 = self.sess.run(policy.distribution.probs,
feed_dict={policy.model.input: [[obs]]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
obs_var = tf.compat.v1.placeholder(tf.float32,
shape=(None, None) +
env.observation_space.shape,
name='obs')
policy_pickled.build(obs_var)
output2 = sess.run(policy_pickled.distribution.probs,
feed_dict={policy_pickled.model.input: [[obs]]})
assert np.array_equal(output1, output2)
def test_dist_info(self, obs_dim, action_dim, filter_dims, filter_sizes,
strides, padding, hidden_sizes):
env = TfEnv(DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('metarl.tf.policies.'
'categorical_cnn_policy.MLPModel'),
new=SimpleMLPModel):
with mock.patch(('metarl.tf.policies.'
'categorical_cnn_policy.CNNModel'),
new=SimpleCNNModel):
policy = CategoricalCNNPolicy(env_spec=env.spec,
conv_filters=filter_dims,
conv_filter_sizes=filter_sizes,
conv_strides=strides,
conv_pad=padding,
hidden_sizes=hidden_sizes)
env.reset()
obs, _, _, _ = env.step(1)
expected_prob = np.full(action_dim, 0.5)
policy_probs = policy.dist_info([obs])
assert np.array_equal(policy_probs['prob'][0], expected_prob)
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:
env = TfEnv(normalize(gym.make('CubeCrash-v0')))
policy = CategoricalCNNPolicy(env_spec=env.spec,
conv_filters=(32, 64),
conv_filter_sizes=(8, 4),
conv_strides=(4, 2),
conv_pad='VALID',
hidden_sizes=(32, 32))
baseline = GaussianCNNBaseline(env_spec=env.spec,
regressor_args=dict(
num_filters=(32, 64),
filter_dims=(8, 4),
strides=(4, 2),
padding='VALID',
hidden_sizes=(32, 32),
use_trust_region=True))
algo = TRPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
max_kl_step=0.01,
flatten_input=False)
runner.setup(algo, env)
runner.train(n_epochs=100, batch_size=variant_data['batch_size'])
def trpo_cubecrash(ctxt=None, seed=1, batch_size=4000):
"""Train TRPO with CubeCrash-v0 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.
batch_size (int): Number of timesteps to use in each training step.
"""
set_seed(seed)
with LocalTFRunner(ctxt) as runner:
env = MetaRLEnv(normalize(gym.make('CubeCrash-v0')))
policy = CategoricalCNNPolicy(env_spec=env.spec,
filters=((32, (8, 8)), (64, (4, 4))),
strides=(4, 2),
padding='VALID',
hidden_sizes=(32, 32))
baseline = GaussianCNNBaseline(
env_spec=env.spec,
regressor_args=dict(filters=((32, (8, 8)), (64, (4, 4))),
strides=(4, 2),
padding='VALID',
hidden_sizes=(32, 32),
use_trust_region=True))
algo = TRPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
gae_lambda=0.95,
lr_clip_range=0.2,
policy_ent_coeff=0.0,
flatten_input=False)
runner.setup(algo, env)
runner.train(n_epochs=100, batch_size=batch_size)
def test_get_action(self, filters, strides, padding, hidden_sizes):
env = MetaRLEnv(DummyDiscretePixelEnv())
policy = CategoricalCNNPolicy(env_spec=env.spec,
filters=filters,
strides=strides,
padding=padding,
hidden_sizes=hidden_sizes)
obs_var = tf.compat.v1.placeholder(tf.float32,
shape=(None, None) +
env.observation_space.shape,
name='obs')
policy.build(obs_var)
env.reset()
obs, _, _, _ = env.step(1)
action, _ = policy.get_action(obs)
assert env.action_space.contains(action)
actions, _ = policy.get_actions([obs, obs, obs])
for action in actions:
assert env.action_space.contains(action)
def run_metarl(env, seed, log_dir):
'''
Create metarl model and training.
Replace the ppo 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)
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))
policy = CategoricalCNNPolicy(
env_spec=env.spec,
conv_filters=params['conv_filters'],
conv_filter_sizes=params['conv_filter_sizes'],
conv_strides=params['conv_strides'],
conv_pad=params['conv_pad'],
hidden_sizes=params['hidden_sizes'])
baseline = GaussianCNNBaseline(
env_spec=env.spec,
regressor_args=dict(num_filters=params['conv_filters'],
filter_dims=params['conv_filter_sizes'],
strides=params['conv_strides'],
padding=params['conv_pad'],
hidden_sizes=params['hidden_sizes'],
use_trust_region=params['use_trust_region']))
algo = PPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
gae_lambda=0.95,
lr_clip_range=0.2,
policy_ent_coeff=0.0,
optimizer_args=dict(
batch_size=32,
max_epochs=10,
tf_optimizer_args=dict(learning_rate=1e-3),
),
flatten_input=False,
)
# 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(algo, env)
runner.train(n_epochs=params['n_epochs'],
batch_size=params['batch_size'])
dowel_logger.remove_all()
return tabular_log_file