def test_obs_not_image(self):
env = self.env
with mock.patch(('garage.tf.models.'
'categorical_cnn_model.CNNModel._build'),
autospec=True,
side_effect=CNNModel._build) as build:
qf = DiscreteCNNQFunction(env_spec=env.spec,
filters=((5, (3, 3)), ),
strides=(2, ),
dueling=False)
normalized_obs = build.call_args_list[0][0][1]
input_ph = qf.input
assert input_ph == normalized_obs
fake_obs = [np.full(env.spec.observation_space.shape, 255)]
assert (self.sess.run(normalized_obs,
feed_dict={input_ph:
fake_obs}) == 255.).all()
obs_dim = env.spec.observation_space.shape
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, ) + obs_dim)
qf.get_qval_sym(state_input, name='another')
normalized_obs = build.call_args_list[1][0][1]
fake_obs = [np.full(env.spec.observation_space.shape, 255)]
assert (self.sess.run(normalized_obs,
feed_dict={state_input:
fake_obs}) == 255).all()
def test_obs_is_image(self):
image_env = TfEnv(DummyDiscretePixelEnv(), is_image=True)
with mock.patch(('garage.tf.models.'
'categorical_cnn_model.CNNModel._build'),
autospec=True,
side_effect=CNNModel._build) as build:
qf = DiscreteCNNQFunction(env_spec=image_env.spec,
filter_dims=(3, ),
num_filters=(5, ),
strides=(2, ),
dueling=False)
normalized_obs = build.call_args_list[0][0][1]
input_ph = qf.input
assert input_ph != normalized_obs
fake_obs = [np.full(image_env.spec.observation_space.shape, 255)]
assert (self.sess.run(normalized_obs,
feed_dict={input_ph: fake_obs}) == 1.).all()
obs_dim = image_env.spec.observation_space.shape
state_input = tf.compat.v1.placeholder(tf.uint8,
shape=(None, ) + obs_dim)
qf.get_qval_sym(state_input, name='another')
normalized_obs = build.call_args_list[1][0][1]
fake_obs = [np.full(image_env.spec.observation_space.shape, 255)]
assert (self.sess.run(normalized_obs,
feed_dict={state_input:
fake_obs}) == 1.).all()
def test_get_qval_sym(self, filters, strides):
with mock.patch(('garage.tf.q_functions.'
'discrete_cnn_q_function.CNNModel'),
new=SimpleCNNModel):
with mock.patch(('garage.tf.q_functions.'
'discrete_cnn_q_function.MLPModel'),
new=SimpleMLPModel):
qf = DiscreteCNNQFunction(env_spec=self.env.spec,
filters=filters,
strides=strides,
dueling=False)
output1 = self.sess.run(qf.q_vals, feed_dict={qf.input: [self.obs]})
obs_dim = self.env.observation_space.shape
action_dim = self.env.action_space.n
input_var = tf.compat.v1.placeholder(tf.float32,
shape=(None, ) + obs_dim)
q_vals = qf.get_qval_sym(input_var, 'another')
output2 = self.sess.run(q_vals, feed_dict={input_var: [self.obs]})
expected_output = np.full(action_dim, 0.5)
assert np.array_equal(output1, output2)
assert np.array_equal(output2[0], expected_output)
def test_obs_not_image(self):
env = self.env
with mock.patch(('garage.tf.policies.'
'categorical_cnn_policy.CNNModel._build'),
autospec=True,
side_effect=CNNModel._build) as build:
qf = DiscreteCNNQFunction(env_spec=env.spec,
filter_dims=(3, ),
num_filters=(5, ),
strides=(2, ),
dueling=False)
normalized_obs = build.call_args_list[0][0][1]
input_ph = tf.compat.v1.get_default_graph().get_tensor_by_name(
'obs:0')
fake_obs = [np.full(env.spec.observation_space.shape, 255)]
assert (self.sess.run(normalized_obs,
feed_dict={input_ph:
fake_obs}) == 255.).all()
obs_dim = env.spec.observation_space.shape
state_input = tf.compat.v1.placeholder(tf.float32,
shape=(None, ) + obs_dim)
qf.get_qval_sym(state_input, name='another')
normalized_obs = build.call_args_list[1][0][1]
input_ph = tf.compat.v1.get_default_graph().get_tensor_by_name(
'Placeholder:0')
fake_obs = [np.full(env.spec.observation_space.shape, 255)]
assert (self.sess.run(normalized_obs,
feed_dict={input_ph:
fake_obs}) == 255).all()
