def __init__(
self,
optimizer=None,
optimizer_args=None,
**kwargs):
if optimizer is None:
if optimizer_args is None:
optimizer_args = dict()
optimizer = ConjugateGradientOptimizer(**optimizer_args)
super(TRPO, self).__init__(optimizer=optimizer, **kwargs)
def __init__(
self,
name,
input_shape,
output_dim,
prob_network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
optimizer=None,
tr_optimizer=None,
use_trust_region=True,
step_size=0.01,
normalize_inputs=True,
no_initial_trust_region=True,
):
"""
:param input_shape: Shape of the input data.
:param output_dim: Dimension of output.
:param hidden_sizes: Number of hidden units of each layer of the mean network.
:param hidden_nonlinearity: Non-linearity used for each layer of the mean network.
:param optimizer: Optimizer for minimizing the negative log-likelihood.
:param use_trust_region: Whether to use trust region constraint.
:param step_size: KL divergence constraint for each iteration
"""
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
if optimizer is None:
optimizer = LbfgsOptimizer(name="optimizer")
if tr_optimizer is None:
tr_optimizer = ConjugateGradientOptimizer()
self.output_dim = output_dim
self.optimizer = optimizer
self.tr_optimizer = tr_optimizer
if prob_network is None:
prob_network = MLP(input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=tf.nn.softmax,
name="prob_network")
l_prob = prob_network.output_layer
LayersPowered.__init__(self, [l_prob])
xs_var = prob_network.input_layer.input_var
ys_var = tf.placeholder(dtype=tf.float32,
shape=[None, output_dim],
name="ys")
old_prob_var = tf.placeholder(dtype=tf.float32,
shape=[None, output_dim],
name="old_prob")
x_mean_var = tf.get_variable(name="x_mean",
shape=(1, ) + input_shape,
initializer=tf.constant_initializer(
0., dtype=tf.float32))
x_std_var = tf.get_variable(name="x_std",
shape=(1, ) + input_shape,
initializer=tf.constant_initializer(
1., dtype=tf.float32))
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
prob_var = L.get_output(
l_prob, {prob_network.input_layer: normalized_xs_var})
old_info_vars = dict(prob=old_prob_var)
info_vars = dict(prob=prob_var)
dist = self._dist = Categorical(output_dim)
mean_kl = tf.reduce_mean(dist.kl_sym(old_info_vars, info_vars))
loss = -tf.reduce_mean(dist.log_likelihood_sym(ys_var, info_vars))
predicted = tensor_utils.to_onehot_sym(
tf.argmax(prob_var, dimension=1), output_dim)
self.prob_network = prob_network
self.f_predict = tensor_utils.compile_function([xs_var], predicted)
self.f_prob = tensor_utils.compile_function([xs_var], prob_var)
self.l_prob = l_prob
self.optimizer.update_opt(loss=loss,
target=self,
network_outputs=[prob_var],
inputs=[xs_var, ys_var])
self.tr_optimizer.update_opt(loss=loss,
target=self,
network_outputs=[prob_var],
inputs=[xs_var, ys_var, old_prob_var],
leq_constraint=(mean_kl, step_size))
self.use_trust_region = use_trust_region
self.name = name
self.normalize_inputs = normalize_inputs
self.x_mean_var = x_mean_var
self.x_std_var = x_std_var
self.first_optimized = not no_initial_trust_region
stub(globals())
env = TfEnv(normalize(CartpoleEnv()))
policy = GaussianLSTMPolicy(
name="policy",
env_spec=env.spec,
lstm_layer_cls=L.TfBasicLSTMLayer,
# gru_layer_cls=L.GRULayer,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TRPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=4000,
max_path_length=100,
n_itr=10,
discount=0.99,
step_size=0.01,
optimizer=ConjugateGradientOptimizer(hvp_approach=FiniteDifferenceHvp(base_eps=1e-5))
)
run_experiment_lite(
algo.train(),
n_parallel=4,
seed=1,
)