def test_dist(self):
model = CategoricalGRUModel(output_dim=1, hidden_dim=1)
step_hidden_var = tf.compat.v1.placeholder(shape=(self.batch_size, 1),
name='step_hidden',
dtype=tf.float32)
dist = model.build(self.input_var, self.step_input_var,
step_hidden_var).dist
assert isinstance(dist, tfp.distributions.OneHotCategorical)
def test_is_pickleable(self):
model = CategoricalGRUModel(output_dim=1, hidden_dim=1)
step_hidden_var = tf.compat.v1.placeholder(shape=(self.batch_size, 1),
name='step_hidden',
dtype=tf.float32)
network = model.build(self.input_var, self.step_input_var,
step_hidden_var)
dist = network.dist
# assign bias to all one
with tf.compat.v1.variable_scope('CategoricalGRUModel/gru',
reuse=True):
init_hidden = tf.compat.v1.get_variable('initial_hidden')
init_hidden.load(tf.ones_like(init_hidden).eval())
hidden = np.zeros((self.batch_size, 1))
outputs1 = self.sess.run(dist.probs,
feed_dict={self.input_var: self.obs_inputs})
output1 = self.sess.run(
[network.step_output, network.step_hidden],
# yapf: disable
feed_dict={
self.step_input_var: self.obs_input,
step_hidden_var: hidden
})
# yapf: enable
h = pickle.dumps(model)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
input_var = tf.compat.v1.placeholder(tf.float32, shape=(None, 5))
model_pickled = pickle.loads(h)
input_var = tf.compat.v1.placeholder(tf.float32,
shape=(None, None,
self.feature_shape),
name='input')
step_input_var = tf.compat.v1.placeholder(
tf.float32, shape=(None, self.feature_shape), name='input')
step_hidden_var = tf.compat.v1.placeholder(shape=(self.batch_size,
1),
name='initial_hidden',
dtype=tf.float32)
network2 = model_pickled.build(input_var, step_input_var,
step_hidden_var)
dist2 = network2.dist
outputs2 = sess.run(dist2.probs,
feed_dict={input_var: self.obs_inputs})
output2 = sess.run(
[network2.step_output, network2.step_hidden],
# yapf: disable
feed_dict={
step_input_var: self.obs_input,
step_hidden_var: hidden
})
# yapf: enable
assert np.array_equal(outputs1, outputs2)
assert np.array_equal(output1, output2)
def test_output_nonlinearity(self):
model = CategoricalGRUModel(output_dim=1,
hidden_dim=4,
output_nonlinearity=lambda x: x / 2)
obs_ph = tf.compat.v1.placeholder(tf.float32, shape=(None, None, 1))
step_obs_ph = tf.compat.v1.placeholder(tf.float32, shape=(None, 1))
step_hidden_ph = tf.compat.v1.placeholder(tf.float32, shape=(None, 4))
obs = np.ones((1, 1, 1))
dist, _, _, _ = model.build(obs_ph, step_obs_ph, step_hidden_ph)
probs = tf.compat.v1.get_default_session().run(dist.probs,
feed_dict={obs_ph: obs})
assert probs == [0.5]
def test_output_normalized(self, output_dim):
model = CategoricalGRUModel(output_dim=output_dim, hidden_dim=4)
obs_ph = tf.compat.v1.placeholder(tf.float32,
shape=(None, None, output_dim))
step_obs_ph = tf.compat.v1.placeholder(tf.float32,
shape=(None, output_dim))
step_hidden_ph = tf.compat.v1.placeholder(tf.float32, shape=(None, 4))
obs = np.ones((1, 1, output_dim))
dist = model.build(obs_ph, step_obs_ph, step_hidden_ph).dist
probs = tf.compat.v1.get_default_session().run(tf.reduce_sum(
dist.probs),
feed_dict={obs_ph: obs})
assert np.isclose(probs, 1.0)
def __init__(self,
env_spec,
name='CategoricalGRUPolicy',
hidden_dim=32,
hidden_nonlinearity=tf.nn.tanh,
hidden_w_init=tf.initializers.glorot_uniform(),
hidden_b_init=tf.zeros_initializer(),
recurrent_nonlinearity=tf.nn.sigmoid,
recurrent_w_init=tf.initializers.glorot_uniform(),
output_nonlinearity=tf.nn.softmax,
output_w_init=tf.initializers.glorot_uniform(),
output_b_init=tf.zeros_initializer(),
hidden_state_init=tf.zeros_initializer(),
hidden_state_init_trainable=False,
state_include_action=True,
layer_normalization=False):
if not isinstance(env_spec.action_space, akro.Discrete):
raise ValueError('CategoricalGRUPolicy only works'
'with akro.Discrete action space.')
super().__init__(name, env_spec)
self._obs_dim = env_spec.observation_space.flat_dim
self._action_dim = env_spec.action_space.n
self._hidden_dim = hidden_dim
self._hidden_nonlinearity = hidden_nonlinearity
self._hidden_w_init = hidden_w_init
self._hidden_b_init = hidden_b_init
self._recurrent_nonlinearity = recurrent_nonlinearity
self._recurrent_w_init = recurrent_w_init
self._output_nonlinearity = output_nonlinearity
self._output_w_init = output_w_init
self._output_b_init = output_b_init
self._hidden_state_init = hidden_state_init
self._hidden_state_init_trainable = hidden_state_init_trainable
self._layer_normalization = layer_normalization
self._state_include_action = state_include_action
if state_include_action:
self._input_dim = self._obs_dim + self._action_dim
else:
self._input_dim = self._obs_dim
self._f_step_prob = None
self.model = CategoricalGRUModel(
output_dim=self._action_dim,
hidden_dim=self._hidden_dim,
name='prob_network',
hidden_nonlinearity=hidden_nonlinearity,
hidden_w_init=hidden_w_init,
hidden_b_init=hidden_b_init,
recurrent_nonlinearity=recurrent_nonlinearity,
recurrent_w_init=recurrent_w_init,
hidden_state_init=hidden_state_init,
hidden_state_init_trainable=hidden_state_init_trainable,
output_nonlinearity=output_nonlinearity,
output_w_init=output_w_init,
output_b_init=output_b_init,
layer_normalization=layer_normalization)
self._prev_actions = None
self._prev_hiddens = None