def test_get_action(self, obs_dim, action_dim, filter_dims, filter_sizes,
strides, padding, hidden_sizes, mock_rand):
mock_rand.return_value = 0
env = TfEnv(DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('garage.tf.policies.'
'categorical_conv_policy_with_model.MLPModel'),
new=SimpleMLPModel):
with mock.patch(('garage.tf.policies.'
'categorical_conv_policy_with_model.CNNModel'),
new=SimpleCNNModel):
policy = CategoricalConvPolicyWithModel(
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(('garage.tf.policies.'
'categorical_conv_policy_with_model.MLPModel'),
new=SimpleMLPModel):
with mock.patch(('garage.tf.policies.'
'categorical_conv_policy_with_model.CNNModel'),
new=SimpleCNNModel):
policy = CategoricalConvPolicyWithModel(
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.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_is_pickleable(self, obs_dim, action_dim, mock_rand):
mock_rand.return_value = 0
env = TfEnv(DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('garage.tf.policies.'
'categorical_conv_policy_with_model.MLPModel'),
new=SimpleMLPModel):
with mock.patch(('garage.tf.policies.'
'categorical_conv_policy_with_model.CNNModel'),
new=SimpleCNNModel):
policy = CategoricalConvPolicyWithModel(
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.variable_scope('CategoricalConvPolicy/Sequential/MLPModel',
reuse=True):
return_var = tf.get_variable('return_var')
# assign it to all one
self.sess.run(tf.assign(return_var, tf.ones_like(return_var)))
output1 = self.sess.run(policy.model.outputs,
feed_dict={policy.model.input: [obs]})
p = pickle.dumps(policy)
with tf.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_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(('garage.tf.policies.'
'categorical_conv_policy_with_model.MLPModel'),
new=SimpleMLPModel):
with mock.patch(('garage.tf.policies.'
'categorical_conv_policy_with_model.CNNModel'),
new=SimpleCNNModel):
policy = CategoricalConvPolicyWithModel(
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)
