def test_is_pickleable(self, obs_dim, action_dim):
"""Test if ContinuousMLPPolicy is pickleable"""
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
policy = ContinuousMLPPolicy(env_spec=env.spec)
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, np.prod(obs_dim)))
outputs = policy.build(state_input, name='policy')
env.reset()
obs, _, _, _ = env.step(1)
with tf.compat.v1.variable_scope('ContinuousMLPPolicy', reuse=True):
bias = tf.compat.v1.get_variable('mlp/hidden_0/bias')
# assign it to all one
bias.load(tf.ones_like(bias).eval())
output1 = self.sess.run([outputs],
feed_dict={state_input: [obs.flatten()]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None,
np.prod(obs_dim)))
outputs = policy_pickled.build(state_input, name='policy')
output2 = sess.run([outputs],
feed_dict={state_input: [obs.flatten()]})
assert np.array_equal(output1, output2)
def test_is_pickleable(self):
env = GarageEnv(DummyDiscretePixelEnv())
policy = CategoricalCNNPolicy(env_spec=env.spec,
filters=((3, (32, 32)), ),
strides=(1, ),
padding='SAME',
hidden_sizes=(4, ))
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)
output2 = sess.run(policy_pickled.distribution.probs,
feed_dict={policy_pickled.model.input: [[obs]]})
assert np.array_equal(output1, output2)
def test_is_pickleable(self, obs_dim, embedding_dim):
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=embedding_dim))
embedding_spec = InOutSpec(input_space=env.spec.observation_space,
output_space=env.spec.action_space)
embedding = GaussianMLPEncoder(embedding_spec)
env.reset()
obs, _, _, _ = env.step(1)
obs_dim = env.spec.observation_space.flat_dim
with tf.compat.v1.variable_scope('GaussianMLPEncoder/GaussianMLPModel',
reuse=True):
bias = tf.compat.v1.get_variable(
'dist_params/mean_network/hidden_0/bias')
# assign it to all one
bias.load(tf.ones_like(bias).eval())
output1 = self.sess.run(
[embedding.distribution.loc,
embedding.distribution.stddev()],
feed_dict={embedding.model.input: [[obs.flatten()]]})
p = pickle.dumps(embedding)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
embedding_pickled = pickle.loads(p)
output2 = sess.run(
[
embedding_pickled.distribution.loc,
embedding_pickled.distribution.stddev()
],
feed_dict={embedding_pickled.model.input: [[obs.flatten()]]})
assert np.array_equal(output1, output2)
def test_get_qval_sym(self, obs_dim, action_dim):
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('garage.tf.q_functions.'
'continuous_mlp_q_function.MLPMergeModel'),
new=SimpleMLPMergeModel):
qf = ContinuousMLPQFunction(env_spec=env.spec)
env.reset()
obs, _, _, _ = env.step(1)
obs = obs.flatten()
act = np.full(action_dim, 0.5).flatten()
output1 = qf.get_qval([obs], [act])
input_var1 = tf.compat.v1.placeholder(tf.float32,
shape=(None, obs.shape[0]))
input_var2 = tf.compat.v1.placeholder(tf.float32,
shape=(None, act.shape[0]))
q_vals = qf.get_qval_sym(input_var1, input_var2, 'another')
output2 = self.sess.run(q_vals,
feed_dict={
input_var1: [obs],
input_var2: [act]
})
expected_output = np.full((1, ), 0.5)
assert np.array_equal(output1, output2)
assert np.array_equal(output2[0], expected_output)
def test_is_pickleable(self):
env = GarageEnv(DummyDiscretePixelEnv())
policy = CategoricalCNNPolicy(env_spec=env.spec,
filters=((3, (32, 32)), ),
strides=(1, ),
padding='SAME',
hidden_sizes=(4, ))
env.reset()
obs, _, _, _ = env.step(1)
with tf.compat.v1.variable_scope('CategoricalCNNPolicy', 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())
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, None) +
policy.input_dim)
dist_sym = policy.build(state_input, name='dist_sym').dist
output1 = self.sess.run(dist_sym.probs,
feed_dict={state_input: [[obs]]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, None) +
policy.input_dim)
dist_sym = policy_pickled.build(state_input, name='dist_sym').dist
output2 = sess.run(dist_sym.probs,
feed_dict={state_input: [[obs]]})
assert np.array_equal(output1, output2)
def test_is_pickleable(self, obs_dim, action_dim):
env = GarageEnv(
DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('garage.tf.q_functions.'
'discrete_mlp_q_function.MLPModel'),
new=SimpleMLPModel):
qf = DiscreteMLPQFunction(env_spec=env.spec)
env.reset()
obs, _, _, _ = env.step(1)
with tf.compat.v1.variable_scope('DiscreteMLPQFunction/SimpleMLPModel',
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(qf.q_vals, feed_dict={qf.input: [obs]})
h_data = pickle.dumps(qf)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
qf_pickled = pickle.loads(h_data)
output2 = sess.run(qf_pickled.q_vals,
feed_dict={qf_pickled.input: [obs]})
assert np.array_equal(output1, output2)
class TestQfDerivedPolicy(TfGraphTestCase):
def setup_method(self):
super().setup_method()
self.env = GarageEnv(DummyDiscreteEnv())
self.qf = SimpleQFunction(self.env.spec)
self.policy = DiscreteQfDerivedPolicy(env_spec=self.env.spec,
qf=self.qf)
self.sess.run(tf.compat.v1.global_variables_initializer())
self.env.reset()
def test_discrete_qf_derived_policy(self):
obs, _, _, _ = self.env.step(1)
action, _ = self.policy.get_action(obs)
assert self.env.action_space.contains(action)
actions, _ = self.policy.get_actions([obs])
for action in actions:
assert self.env.action_space.contains(action)
def test_is_pickleable(self):
with tf.compat.v1.variable_scope('SimpleQFunction/SimpleMLPModel',
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())
obs, _, _, _ = self.env.step(1)
action1, _ = self.policy.get_action(obs)
p = pickle.dumps(self.policy)
with tf.compat.v1.Session(graph=tf.Graph()):
policy_pickled = pickle.loads(p)
action2, _ = policy_pickled.get_action(obs)
assert action1 == action2
def test_is_pickleable(self, obs_dim, action_dim):
"""Test if ContinuousMLPPolicy is pickleable"""
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('garage.tf.policies.'
'continuous_mlp_policy.MLPModel'),
new=SimpleMLPModel):
policy = ContinuousMLPPolicy(env_spec=env.spec)
env.reset()
obs, _, _, _ = env.step(1)
with tf.compat.v1.variable_scope('ContinuousMLPPolicy/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.flatten()]})
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.flatten()]})
assert np.array_equal(output1, output2)
def osimArmResume(ctxt=None,
snapshot_dir='data/local/experiment/osimArm_153',
seed=1):
set_seed(seed)
with LocalTFRunner(snapshot_config=ctxt) as runner:
runner.restore(snapshot_dir)
ddpg = runner._algo
env = GarageEnv(Arm2DVecEnv(visualize=True))
env.reset()
policy = ddpg.policy
env.render()
obs = env.step(env.action_space.sample())
steps = 0
n_steps = 100
while True:
if steps == n_steps:
env.close()
break
temp = policy.get_action(obs[0])
obs = env.step(temp[0])
env.render()
steps += 1
def test_get_action(self, obs_dim, task_num, latent_dim, action_dim):
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
embedding_spec = InOutSpec(
input_space=akro.Box(low=np.zeros(task_num),
high=np.ones(task_num)),
output_space=akro.Box(low=np.zeros(latent_dim),
high=np.ones(latent_dim)))
encoder = GaussianMLPEncoder(embedding_spec)
policy = GaussianMLPTaskEmbeddingPolicy(env_spec=env.spec,
encoder=encoder)
env.reset()
obs, _, _, _ = env.step(1)
latent = np.random.random((latent_dim, ))
task = np.zeros(task_num)
task[0] = 1
action1, _ = policy.get_action_given_latent(obs, latent)
action2, _ = policy.get_action_given_task(obs, task)
action3, _ = policy.get_action(np.concatenate([obs.flatten(), task]))
assert env.action_space.contains(action1)
assert env.action_space.contains(action2)
assert env.action_space.contains(action3)
obses, latents, tasks = [obs] * 3, [latent] * 3, [task] * 3
aug_obses = [np.concatenate([obs.flatten(), task])] * 3
action1n, _ = policy.get_actions_given_latents(obses, latents)
action2n, _ = policy.get_actions_given_tasks(obses, tasks)
action3n, _ = policy.get_actions(aug_obses)
for action in chain(action1n, action2n, action3n):
assert env.action_space.contains(action)
def test_is_pickleable(self):
env = GarageEnv(DummyBoxEnv(obs_dim=(1, ), action_dim=(1, )))
policy = GaussianLSTMPolicy(env_spec=env.spec,
state_include_action=False)
env.reset()
obs = env.reset()
with tf.compat.v1.variable_scope(
'GaussianLSTMPolicy/GaussianLSTMModel', reuse=True):
param = tf.compat.v1.get_variable(
'dist_params/log_std_param/parameter')
# assign it to all one
param.load(tf.ones_like(param).eval())
output1 = self.sess.run(
[policy.distribution.loc,
policy.distribution.stddev()],
feed_dict={policy.model.input: [[obs.flatten()], [obs.flatten()]]})
p = pickle.dumps(policy)
# yapf: disable
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
output2 = sess.run(
[
policy_pickled.distribution.loc,
policy_pickled.distribution.stddev()
],
feed_dict={
policy_pickled.model.input: [[obs.flatten()],
[obs.flatten()]]
})
assert np.array_equal(output1, output2)
def test_is_pickleable(self, obs_dim, action_dim):
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('garage.tf.q_functions.'
'continuous_mlp_q_function.MLPMergeModel'),
new=SimpleMLPMergeModel):
qf = ContinuousMLPQFunction(env_spec=env.spec)
env.reset()
obs, _, _, _ = env.step(1)
obs = obs.flatten()
act = np.full(action_dim, 0.5).flatten()
with tf.compat.v1.variable_scope(
'ContinuousMLPQFunction/SimpleMLPMergeModel', 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 = qf.get_qval([obs], [act])
h_data = pickle.dumps(qf)
with tf.compat.v1.Session(graph=tf.Graph()):
qf_pickled = pickle.loads(h_data)
output2 = qf_pickled.get_qval([obs], [act])
assert np.array_equal(output1, output2)
def test_time_limit_env(self):
garage_env = GarageEnv(env_name='Pendulum-v0')
garage_env.reset()
for _ in range(200):
_, _, done, info = garage_env.step(
garage_env.spec.action_space.sample())
assert not done and info['TimeLimit.truncated']
assert info['GarageEnv.TimeLimitTerminated']
def test_output_shape(self, obs_dim, action_dim):
env = GarageEnv(
DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
qf = DiscreteMLPQFunction(env_spec=env.spec)
env.reset()
obs, _, _, _ = env.step(1)
outputs = self.sess.run(qf.q_vals, feed_dict={qf.input: [obs]})
assert outputs.shape == (1, action_dim)
def test_output_shape(self, obs_dim, action_dim):
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
qf = ContinuousMLPQFunction(env_spec=env.spec)
env.reset()
obs, _, _, _ = env.step(1)
obs = obs.flatten()
act = np.full(action_dim, 0.5).flatten()
outputs = qf.get_qval([obs], [act])
assert outputs.shape == (1, 1)
def test_get_embedding(self, obs_dim, embedding_dim):
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=embedding_dim))
embedding_spec = InOutSpec(input_space=env.spec.observation_space,
output_space=env.spec.action_space)
embedding = GaussianMLPEncoder(embedding_spec)
env.reset()
obs, _, _, _ = env.step(1)
latent, _ = embedding.forward(obs)
assert env.action_space.contains(latent)
def test_output_shape_dueling(self, obs_dim, action_dim):
env = GarageEnv(
DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('garage.tf.q_functions.'
'discrete_mlp_q_function.MLPDuelingModel'),
new=SimpleMLPModel):
qf = DiscreteMLPQFunction(env_spec=env.spec, dueling=True)
env.reset()
obs, _, _, _ = env.step(1)
outputs = self.sess.run(qf.q_vals, feed_dict={qf.input: [obs]})
assert outputs.shape == (1, action_dim)
def test_output_shape(self, obs_dim, action_dim):
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('garage.tf.q_functions.'
'continuous_mlp_q_function.MLPMergeModel'),
new=SimpleMLPMergeModel):
qf = ContinuousMLPQFunction(env_spec=env.spec)
env.reset()
obs, _, _, _ = env.step(1)
obs = obs.flatten()
act = np.full(action_dim, 0.5).flatten()
outputs = qf.get_qval([obs], [act])
assert outputs.shape == (1, 1)
def test_get_action(self, obs_dim, action_dim):
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
policy = GaussianMLPPolicy(env_spec=env.spec)
env.reset()
obs, _, _, _ = env.step(1)
action, _ = policy.get_action(obs.flatten())
assert env.action_space.contains(action)
actions, _ = policy.get_actions(
[obs.flatten(), obs.flatten(),
obs.flatten()])
for action in actions:
assert env.action_space.contains(action)
def test_get_embedding(self, obs_dim, embedding_dim):
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=embedding_dim))
embedding_spec = InOutSpec(input_space=env.spec.observation_space,
output_space=env.spec.action_space)
embedding = GaussianMLPEncoder(embedding_spec)
task_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, None,
embedding.input_dim))
embedding.build(task_input)
env.reset()
obs, _, _, _ = env.step(1)
latent, _ = embedding.forward(obs)
assert env.action_space.contains(latent)
def test_get_action_img_obs(self, hidden_channels, kernel_sizes, strides,
hidden_sizes):
"""Test get_action function with akro.Image observation space."""
env = GarageEnv(DummyDiscretePixelEnv(), is_image=True)
env = self._initialize_obs_env(env)
policy = CategoricalCNNPolicy(env=env,
kernel_sizes=kernel_sizes,
hidden_channels=hidden_channels,
strides=strides,
hidden_sizes=hidden_sizes)
env.reset()
obs, _, _, _ = env.step(1)
action, _ = policy.get_action(obs)
assert env.action_space.contains(action)
def test_build(self, obs_dim, action_dim):
env = GarageEnv(
DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
qf = DiscreteMLPQFunction(env_spec=env.spec)
env.reset()
obs, _, _, _ = env.step(1)
output1 = self.sess.run(qf.q_vals, feed_dict={qf.input: [obs]})
input_var = tf.compat.v1.placeholder(tf.float32,
shape=(None, ) + obs_dim)
q_vals = qf.build(input_var, 'another')
output2 = self.sess.run(q_vals, feed_dict={input_var: [obs]})
assert np.array_equal(output1, output2)
def test_is_pickleable(self, obs_dim, action_dim):
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
policy = GaussianMLPPolicy(env_spec=env.spec)
obs = env.reset()
with tf.compat.v1.variable_scope('GaussianMLPPolicy', reuse=True):
bias = tf.compat.v1.get_variable(
'dist_params/mean_network/hidden_0/bias')
# assign it to all one
bias.load(tf.ones_like(bias).eval())
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, None,
policy.input_dim))
dist_sym = policy.build(state_input, name='dist_sym').dist
output1 = self.sess.run([dist_sym.loc, dist_sym.stddev()],
feed_dict={state_input: [[obs.flatten()]]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, None,
policy.input_dim))
dist_sym = policy_pickled.build(state_input, name='dist_sym').dist
output2 = sess.run([dist_sym.loc, dist_sym.stddev()],
feed_dict={state_input: [[obs.flatten()]]})
assert np.array_equal(output1, output2)
def test_is_pickleable(self):
env = GarageEnv(DummyDiscreteEnv(obs_dim=(1, ), action_dim=1))
policy = CategoricalLSTMPolicy(env_spec=env.spec,
state_include_action=False)
policy.reset()
obs = env.reset()
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, None,
policy.input_dim))
dist_sym = policy.build(state_input, name='dist_sym').dist
policy._lstm_cell.weights[0].load(
tf.ones_like(policy._lstm_cell.weights[0]).eval())
output1 = self.sess.run(
[dist_sym.probs],
feed_dict={state_input: [[obs.flatten()], [obs.flatten()]]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
state_input = tf.compat.v1.placeholder(
tf.float32, shape=(None, None, policy_pickled.input_dim))
dist_sym = policy_pickled.build(state_input, name='dist_sym').dist
output2 = sess.run(
[dist_sym.probs],
feed_dict={state_input: [[obs.flatten()],
[obs.flatten()]]}) # noqa: E126
assert np.array_equal(output1, output2)
def test_is_pickleable(self):
env = GarageEnv(DummyDiscreteEnv(obs_dim=(1, ), action_dim=1))
policy = CategoricalLSTMPolicy(env_spec=env.spec,
state_include_action=False)
policy.reset()
obs = env.reset()
policy.model._lstm_cell.weights[0].load(
tf.ones_like(policy.model._lstm_cell.weights[0]).eval())
output1 = self.sess.run(
[policy.distribution.probs],
feed_dict={policy.model.input: [[obs.flatten()], [obs.flatten()]]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
output2 = sess.run([policy_pickled.distribution.probs],
feed_dict={
policy_pickled.model.input:
[[obs.flatten()], [obs.flatten()]]
}) # noqa: E126
assert np.array_equal(output1, output2)
def test_get_qval_max_pooling(self, filters, strides, pool_strides,
pool_shapes):
env = GarageEnv(DummyDiscretePixelEnv())
obs = env.reset()
with mock.patch(('garage.tf.models.'
'cnn_mlp_merge_model.CNNModelWithMaxPooling'),
new=SimpleCNNModelWithMaxPooling):
with mock.patch(('garage.tf.models.'
'cnn_mlp_merge_model.MLPMergeModel'),
new=SimpleMLPMergeModel):
qf = ContinuousCNNQFunction(env_spec=env.spec,
filters=filters,
strides=strides,
max_pooling=True,
pool_strides=pool_strides,
pool_shapes=pool_shapes)
action_dim = env.action_space.shape
obs, _, _, _ = env.step(1)
act = np.full(action_dim, 0.5)
expected_output = np.full((1, ), 0.5)
outputs = qf.get_qval([obs], [act])
assert np.array_equal(outputs[0], expected_output)
outputs = qf.get_qval([obs, obs, obs], [act, act, act])
for output in outputs:
assert np.array_equal(output, expected_output)
def test_is_pickleable(self, obs_dim, action_dim):
env = GarageEnv(
DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
obs_var = tf.compat.v1.placeholder(
tf.float32,
shape=[None, None, env.observation_space.flat_dim],
name='obs')
policy = CategoricalMLPPolicy(env_spec=env.spec)
policy.build(obs_var)
obs = env.reset()
with tf.compat.v1.variable_scope(
'CategoricalMLPPolicy/CategoricalMLPModel', reuse=True):
bias = tf.compat.v1.get_variable('mlp/hidden_0/bias')
# assign it to all one
bias.load(tf.ones_like(bias).eval())
output1 = self.sess.run(
[policy.distribution.probs],
feed_dict={policy.model.input: [[obs.flatten()]]})
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.flat_dim],
name='obs')
policy_pickled.build(obs_var)
output2 = sess.run(
[policy_pickled.distribution.probs],
feed_dict={policy_pickled.model.input: [[obs.flatten()]]})
assert np.array_equal(output1, output2)
def test_get_action_state_include_action(self, obs_dim, action_dim,
hidden_dim):
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
obs_var = tf.compat.v1.placeholder(
tf.float32,
shape=[
None, None,
env.observation_space.flat_dim + np.prod(action_dim)
],
name='obs')
policy = GaussianGRUPolicy(env_spec=env.spec,
hidden_dim=hidden_dim,
state_include_action=True)
policy.build(obs_var)
policy.reset()
obs = env.reset()
action, _ = policy.get_action(obs.flatten())
assert env.action_space.contains(action)
policy.reset()
actions, _ = policy.get_actions([obs.flatten()])
for action in actions:
assert env.action_space.contains(action)
def test_is_pickleable(self, obs_dim, action_dim):
env = GarageEnv(DummyBoxEnv(obs_dim=obs_dim, action_dim=action_dim))
policy = GaussianMLPPolicy(env_spec=env.spec)
obs = env.reset()
with tf.compat.v1.variable_scope('GaussianMLPPolicy/GaussianMLPModel',
reuse=True):
bias = tf.compat.v1.get_variable(
'dist_params/mean_network/hidden_0/bias')
# assign it to all one
bias.load(tf.ones_like(bias).eval())
output1 = self.sess.run(
[policy.distribution.loc,
policy.distribution.stddev()],
feed_dict={policy.model.input: [[obs.flatten()]]})
p = pickle.dumps(policy)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
policy_pickled = pickle.loads(p)
output2 = sess.run(
[
policy_pickled.distribution.loc,
policy_pickled.distribution.stddev()
],
feed_dict={policy_pickled.model.input: [[obs.flatten()]]})
assert np.array_equal(output1, output2)
def test_get_action(self, filters, strides, padding, hidden_sizes):
env = GarageEnv(DummyDiscretePixelEnv())
policy = CategoricalCNNPolicy(env_spec=env.spec,
filters=filters,
strides=strides,
padding=padding,
hidden_sizes=hidden_sizes)
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)
