def update_opt(self,
loss,
target,
inputs,
extra_inputs=None,
*args,
**kwargs):
"""
:param loss: Symbolic expression for the loss function.
:param target: A parameterized object to optimize over. It should implement methods of the
:class:`rllab.core.paramerized.Parameterized` class.
:param leq_constraint: A constraint provided as a tuple (f, epsilon), of the form f(*inputs) <= epsilon.
:param inputs: A list of symbolic variables as inputs
:return: No return value.
"""
if config.TF_NN_SETTRACE:
ipdb.set_trace()
self._target = target
def get_opt_output():
flat_grad = tensor_utils.flatten_tensor_variables(
tf.gradients(loss, target.get_params(trainable=True)))
return [tf.cast(loss, tf.float64), tf.cast(flat_grad, tf.float64)]
if extra_inputs is None:
extra_inputs = list()
self._opt_fun = ext.lazydict(
f_loss=lambda: tensor_utils.compile_function(
inputs + extra_inputs, loss),
f_opt=lambda: tensor_utils.compile_function(
inputs=inputs + extra_inputs,
outputs=get_opt_output(),
))
def update_opt(self,
loss,
target,
leq_constraint,
inputs,
constraint_name="constraint",
*args,
**kwargs):
"""
:param loss: Symbolic expression for the loss function.
:param target: A parameterized object to optimize over. It should implement methods of the
:class:`rllab.core.paramerized.Parameterized` class.
:param leq_constraint: A constraint provided as a tuple (f, epsilon), of the form f(*inputs) <= epsilon.
:param inputs: A list of symbolic variables as inputs
:return: No return value.
"""
if config.TF_NN_SETTRACE:
ipdb.set_trace()
constraint_term, constraint_value = leq_constraint
with tf.variable_scope(self._name):
penalty_var = tf.placeholder(tf.float32, tuple(), name="penalty")
penalized_loss = loss + penalty_var * constraint_term
self._target = target
self._max_constraint_val = constraint_value
self._constraint_name = constraint_name
def get_opt_output():
params = target.get_params(trainable=True)
grads = tf.gradients(penalized_loss, params)
for idx, (grad, param) in enumerate(zip(grads, params)):
if grad is None:
grads[idx] = tf.zeros_like(param)
flat_grad = tensor_utils.flatten_tensor_variables(grads)
return [
tf.cast(penalized_loss, tf.float64),
tf.cast(flat_grad, tf.float64),
]
self._opt_fun = ext.lazydict(
f_loss=lambda: tensor_utils.compile_function(
inputs, loss, log_name="f_loss"),
f_constraint=lambda: tensor_utils.compile_function(
inputs, constraint_term, log_name="f_constraint"),
f_penalized_loss=lambda: tensor_utils.compile_function(
inputs=inputs + [penalty_var],
outputs=[penalized_loss, loss, constraint_term],
log_name="f_penalized_loss",
),
f_opt=lambda: tensor_utils.compile_function(
inputs=inputs + [penalty_var],
outputs=get_opt_output(),
))
def update_opt(self, f, target, inputs, reg_coeff):
self.target = target
self.reg_coeff = reg_coeff
params = target.get_params(trainable=True)
constraint_grads = tf.gradients(f, xs=params)
for idx, (grad, param) in enumerate(zip(constraint_grads, params)):
if grad is None:
constraint_grads[idx] = tf.zeros_like(param)
xs = tuple([
tensor_utils.new_tensor_like(p.name.split(":")[0], p)
for p in params
])
def Hx_plain():
Hx_plain_splits = tf.gradients(
tf.reduce_sum(
tf.pack([
tf.reduce_sum(g * x)
for g, x in zip(constraint_grads, xs)
])), params)
for idx, (Hx, param) in enumerate(zip(Hx_plain_splits, params)):
if Hx is None:
Hx_plain_splits[idx] = tf.zeros_like(param)
return tensor_utils.flatten_tensor_variables(Hx_plain_splits)
self.opt_fun = ext.lazydict(
f_Hx_plain=lambda: tensor_utils.compile_function(
inputs=inputs + xs,
outputs=Hx_plain(),
log_name="f_Hx_plain",
), )
def __init__(self, dim):
self._dim = dim
weights_var = tf.placeholder(dtype=tf.float32,
shape=(None, dim),
name="weights")
self._f_sample = tensor_utils.compile_function(
inputs=[weights_var],
outputs=tf.multinomial(weights_var, num_samples=1)[:, 0],
)
def update_opt(self, loss, target, inputs, extra_inputs=None, **kwargs):
"""
:param loss: Symbolic expression for the loss function.
:param target: A parameterized object to optimize over. It should implement methods of the
:class:`rllab.core.paramerized.Parameterized` class.
:param leq_constraint: A constraint provided as a tuple (f, epsilon), of the form f(*inputs) <= epsilon.
:param inputs: A list of symbolic variables as inputs
:return: No return value.
"""
self._target = target
self._train_op = self._tf_optimizer.minimize(
loss, var_list=target.get_params(trainable=True))
# define operations for updating prior.
update_mus = [(l.bayesreg.hyperparams['empirical'],
l.bayesreg.w_mu.assign(l.W_mu)) for l in target.layers
if hasattr(l, 'bayesreg')]
update_rhos = [(l.bayesreg.hyperparams['empirical'],
l.bayesreg.w_sig.assign(tf.log(1.0 + tf.exp(l.W_rho))))
for l in target.layers if hasattr(l, 'bayesreg')]
self._update_priors_ops = update_mus + update_rhos
# updates = OrderedDict([(k, v.astype(k.dtype)) for k, v in updates.iteritems()])
if extra_inputs is None:
extra_inputs = list()
self._input_vars = inputs + extra_inputs
self._opt_fun = ext.lazydict(
f_loss=lambda: tensor_utils.compile_function(
inputs + extra_inputs, loss), )
if kwargs.has_key('like_loss'):
l_loss = lambda: tensor_utils.compile_function(
inputs + extra_inputs, kwargs['like_loss'])
self._opt_fun.set('l_loss', l_loss)
if kwargs.has_key('cmpx_loss'):
c_loss = lambda: tensor_utils.compile_function(
inputs + extra_inputs, kwargs['cmpx_loss'])
self._opt_fun.set('c_loss', c_loss)
def update_opt(self, f, target, inputs, reg_coeff):
if config.TF_NN_SETTRACE:
ipdb.set_trace()
self.target = target
self.reg_coeff = reg_coeff
params = target.get_params(trainable=True)
constraint_grads = tf.gradients(f, xs=params)
for idx, (grad, param) in enumerate(zip(constraint_grads, params)):
if grad is None:
constraint_grads[idx] = tf.zeros_like(param)
flat_grad = tensor_utils.flatten_tensor_variables(constraint_grads)
def f_Hx_plain(*args):
inputs_ = args[:len(inputs)]
xs = args[len(inputs):]
flat_xs = np.concatenate([np.reshape(x, (-1, )) for x in xs])
param_val = self.target.get_param_values(trainable=True)
eps = np.cast['float32'](self.base_eps /
(np.linalg.norm(param_val) + 1e-8))
self.target.set_param_values(param_val + eps * flat_xs,
trainable=True)
flat_grad_dvplus = self.opt_fun["f_grad"](*inputs_)
self.target.set_param_values(param_val, trainable=True)
if self.symmetric:
self.target.set_param_values(param_val - eps * flat_xs,
trainable=True)
flat_grad_dvminus = self.opt_fun["f_grad"](*inputs_)
hx = (flat_grad_dvplus - flat_grad_dvminus) / (2 * eps)
self.target.set_param_values(param_val, trainable=True)
else:
flat_grad = self.opt_fun["f_grad"](*inputs_)
hx = (flat_grad_dvplus - flat_grad) / eps
return hx
self.opt_fun = ext.lazydict(
f_grad=lambda: tensor_utils.compile_function(
inputs=inputs,
outputs=flat_grad,
log_name="f_grad",
),
f_Hx_plain=lambda: f_Hx_plain,
)
def c_loss():
return tensor_utils.compile_function(inputs + extra_inputs,
kwargs['cmpx_loss'])
def l_loss():
return tensor_utils.compile_function(inputs + extra_inputs,
kwargs['like_loss'])
def __init__(
self,
name,
env_spec,
hidden_sizes=(32, 32),
learn_std=True,
init_std=1.0,
adaptive_std=False,
std_share_network=False,
std_hidden_sizes=(32, 32),
min_std=1e-6,
std_hidden_nonlinearity=tf.nn.tanh,
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
mean_network=None,
std_network=None,
std_parametrization='exp',
batch_normalization=False
):
"""
:param env_spec:
:param hidden_sizes: list of sizes for the fully-connected hidden layers
:param learn_std: Is std trainable
:param init_std: Initial std
:param adaptive_std:
:param std_share_network:
:param std_hidden_sizes: list of sizes for the fully-connected layers for std
:param min_std: whether to make sure that the std is at least some threshold value, to avoid numerical issues
:param std_hidden_nonlinearity:
:param hidden_nonlinearity: nonlinearity used for each hidden layer
:param output_nonlinearity: nonlinearity for the output layer
:param mean_network: custom network for the output mean
:param std_network: custom network for the output log std
:param std_parametrization: how the std should be parametrized. There are a few options:
- exp: the logarithm of the std will be stored, and applied a exponential transformation
- softplus: the std will be computed as log(1+exp(x))
:return:
"""
Serializable.quick_init(self, locals())
assert isinstance(env_spec.action_space, Box)
with tf.variable_scope(name):
obs_dim = env_spec.observation_space.flat_dim
action_dim = env_spec.action_space.flat_dim
# create network
if mean_network is None:
mean_network = MLP(
name="mean_network",
input_shape=(obs_dim,),
output_dim=action_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
batch_normalization=batch_normalization
)
self._mean_network = mean_network
l_mean = mean_network.output_layer
obs_var = mean_network.input_layer.input_var
if std_network is not None:
l_std_param = std_network.output_layer
else:
if adaptive_std:
std_network = MLP(
name="std_network",
input_shape=(obs_dim,),
input_layer=mean_network.input_layer,
output_dim=action_dim,
hidden_sizes=std_hidden_sizes,
hidden_nonlinearity=std_hidden_nonlinearity,
output_nonlinearity=None,
batch_normalization=batch_normalization
)
l_std_param = std_network.output_layer
else:
if std_parametrization == 'exp':
init_std_param = np.log(init_std)
elif std_parametrization == 'softplus':
init_std_param = np.log(np.exp(init_std) - 1)
else:
raise NotImplementedError
l_std_param = L.ParamLayer(
mean_network.input_layer,
num_units=action_dim,
param=tf.constant_initializer(init_std_param),
name="output_std_param",
trainable=learn_std,
)
self.std_parametrization = std_parametrization
if std_parametrization == 'exp':
min_std_param = np.log(min_std)
elif std_parametrization == 'softplus':
min_std_param = np.log(np.exp(min_std) - 1)
else:
raise NotImplementedError
self.min_std_param = min_std_param
# mean_var, log_std_var = L.get_output([l_mean, l_std_param])
#
# if self.min_std_param is not None:
# log_std_var = tf.maximum(log_std_var, np.log(min_std))
#
# self._mean_var, self._log_std_var = mean_var, log_std_var
self._l_mean = l_mean
self._l_std_param = l_std_param
self._dist = DiagonalGaussian(action_dim)
LayersPowered.__init__(self, [l_mean, l_std_param])
super(GaussianMLPPolicy, self).__init__(env_spec)
dist_info_sym = self.dist_info_sym(
mean_network.input_layer.input_var, dict())
mean_var = dist_info_sym["mean"]
log_std_var = dist_info_sym["log_std"]
self._f_dist = tensor_utils.compile_function(
inputs=[obs_var],
outputs=[mean_var, log_std_var],
)
def __init__(self,
name,
input_shape,
output_dim,
mean_network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
optimizer=None,
use_trust_region=True,
step_size=0.01,
learn_std=True,
init_std=1.0,
adaptive_std=False,
std_share_network=False,
std_hidden_sizes=(32, 32),
std_nonlinearity=None,
normalize_inputs=True,
normalize_outputs=True,
subsample_factor=1.0):
"""
: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
:param learn_std: Whether to learn the standard deviations. Only effective if adaptive_std is False. If
adaptive_std is True, this parameter is ignored, and the weights for the std network are always learned.
:param adaptive_std: Whether to make the std a function of the states.
:param std_share_network: Whether to use the same network as the mean.
:param std_hidden_sizes: Number of hidden units of each layer of the std network. Only used if
`std_share_network` is False. It defaults to the same architecture as the mean.
:param std_nonlinearity: Non-linearity used for each layer of the std network. Only used if `std_share_network`
is False. It defaults to the same non-linearity as the mean.
"""
if config.TF_NN_SETTRACE:
ipdb.set_trace()
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
if optimizer is None:
if use_trust_region:
optimizer = PenaltyLbfgsOptimizer("optimizer")
else:
optimizer = LbfgsOptimizer("optimizer")
self._optimizer = optimizer
self._subsample_factor = subsample_factor
if mean_network is None:
mean_network = MLP(
name="mean_network",
input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=None,
)
l_mean = mean_network.output_layer
if adaptive_std:
l_log_std = MLP(
name="log_std_network",
input_shape=input_shape,
input_var=mean_network.input_layer.input_var,
output_dim=output_dim,
hidden_sizes=std_hidden_sizes,
hidden_nonlinearity=std_nonlinearity,
output_nonlinearity=None,
).output_layer
else:
l_log_std = L.ParamLayer(
mean_network.input_layer,
num_units=output_dim,
param=tf.constant_initializer(np.log(init_std)),
name="output_log_std",
trainable=learn_std,
)
LayersPowered.__init__(self, [l_mean, l_log_std])
xs_var = mean_network.input_layer.input_var
ys_var = tf.placeholder(dtype=tf.float32,
name="ys",
shape=(None, output_dim))
old_means_var = tf.placeholder(dtype=tf.float32,
name="ys",
shape=(None, output_dim))
old_log_stds_var = tf.placeholder(dtype=tf.float32,
name="old_log_stds",
shape=(None, output_dim))
x_mean_var = tf.Variable(
np.zeros((1, ) + input_shape, dtype=np.float32),
name="x_mean",
)
x_std_var = tf.Variable(
np.ones((1, ) + input_shape, dtype=np.float32),
name="x_std",
)
y_mean_var = tf.Variable(
np.zeros((1, output_dim), dtype=np.float32),
name="y_mean",
)
y_std_var = tf.Variable(
np.ones((1, output_dim), dtype=np.float32),
name="y_std",
)
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
normalized_ys_var = (ys_var - y_mean_var) / y_std_var
normalized_means_var = L.get_output(
l_mean, {mean_network.input_layer: normalized_xs_var})
normalized_log_stds_var = L.get_output(
l_log_std, {mean_network.input_layer: normalized_xs_var})
means_var = normalized_means_var * y_std_var + y_mean_var
log_stds_var = normalized_log_stds_var + tf.log(y_std_var)
normalized_old_means_var = (old_means_var - y_mean_var) / y_std_var
normalized_old_log_stds_var = old_log_stds_var - tf.log(y_std_var)
dist = self._dist = DiagonalGaussian(output_dim)
normalized_dist_info_vars = dict(mean=normalized_means_var,
log_std=normalized_log_stds_var)
mean_kl = tf.reduce_mean(
dist.kl_sym(
dict(mean=normalized_old_means_var,
log_std=normalized_old_log_stds_var),
normalized_dist_info_vars,
))
loss = - \
tf.reduce_mean(dist.log_likelihood_sym(
normalized_ys_var, normalized_dist_info_vars))
self._f_predict = tensor_utils.compile_function([xs_var],
means_var)
self._f_pdists = tensor_utils.compile_function(
[xs_var], [means_var, log_stds_var])
self._l_mean = l_mean
self._l_log_std = l_log_std
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[
normalized_means_var, normalized_log_stds_var
],
)
if use_trust_region:
optimizer_args["leq_constraint"] = (mean_kl, step_size)
optimizer_args["inputs"] = [
xs_var, ys_var, old_means_var, old_log_stds_var
]
else:
optimizer_args["inputs"] = [xs_var, ys_var]
self._optimizer.update_opt(**optimizer_args)
self._use_trust_region = use_trust_region
self._name = name
self._normalize_inputs = normalize_inputs
self._normalize_outputs = normalize_outputs
self._mean_network = mean_network
self._x_mean_var = x_mean_var
self._x_std_var = x_std_var
self._y_mean_var = y_mean_var
self._y_std_var = y_std_var
def init_opt(self):
is_recurrent = int(self.policy.recurrent)
obs_var = self.env.observation_space.new_tensor_variable(
'obs',
extra_dims=1 + is_recurrent,
)
action_var = self.env.action_space.new_tensor_variable(
'action',
extra_dims=1 + is_recurrent,
)
advantage_var = tensor_utils.new_tensor(
name='advantage',
ndim=1 + is_recurrent,
dtype=tf.float32,
)
dist = self.policy.distribution
old_dist_info_vars = {
k: tf.placeholder(tf.float32, shape=[None] * (1 + is_recurrent) + list(shape), name='old_%s' % k)
for k, shape in dist.dist_info_specs
}
old_dist_info_vars_list = [old_dist_info_vars[k] for k in dist.dist_info_keys]
state_info_vars = {
k: tf.placeholder(tf.float32, shape=[None] * (1 + is_recurrent) + list(shape), name=k)
for k, shape in self.policy.state_info_specs
}
state_info_vars_list = [state_info_vars[k] for k in self.policy.state_info_keys]
if is_recurrent:
valid_var = tf.placeholder(tf.float32, shape=[None, None], name="valid")
else:
valid_var = None
dist_info_vars = self.policy.dist_info_sym(obs_var, state_info_vars)
logli = dist.log_likelihood_sym(action_var, dist_info_vars)
kl = dist.kl_sym(old_dist_info_vars, dist_info_vars)
# formulate as a minimization problem
# The gradient of the surrogate objective is the policy gradient
if is_recurrent:
surr_obj = - tf.reduce_sum(logli * advantage_var * valid_var) / tf.reduce_sum(valid_var)
mean_kl = tf.reduce_sum(kl * valid_var) / tf.reduce_sum(valid_var)
max_kl = tf.reduce_max(kl * valid_var)
else:
surr_obj = - tf.reduce_mean(logli * advantage_var)
mean_kl = tf.reduce_mean(kl)
max_kl = tf.reduce_max(kl)
input_list = [obs_var, action_var, advantage_var] + state_info_vars_list
if is_recurrent:
input_list.append(valid_var)
self.optimizer.update_opt(loss=surr_obj, target=self.policy, inputs=input_list)
f_kl = tensor_utils.compile_function(
inputs=input_list + old_dist_info_vars_list,
outputs=[mean_kl, max_kl],
)
self.opt_info = dict(
f_kl=f_kl,
)
def __init__(
self,
input_shape,
output_dim,
name,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.relu,
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
"""
if config.TF_NN_SETTRACE:
ipdb.set_trace()
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
p_network = MLP(
input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=tf.nn.sigmoid,
name="p_network"
)
l_p = p_network.output_layer
LayersPowered.__init__(self, [l_p])
xs_var = p_network.input_layer.input_var
ys_var = tf.placeholder(
dtype=tf.float32, shape=(None, output_dim), name="ys")
old_p_var = tf.placeholder(
dtype=tf.float32, shape=(None, output_dim), name="old_p")
x_mean_var = tf.get_variable(
name="x_mean", initializer=tf.zeros_initializer, shape=(1,) + input_shape)
x_std_var = tf.get_variable(
name="x_std", initializer=tf.ones_initializer, shape=(1,) + input_shape)
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
p_var = L.get_output(
l_p, {p_network.input_layer: normalized_xs_var})
old_info_vars = dict(p=old_p_var)
info_vars = dict(p=p_var)
dist = self._dist = Bernoulli(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 = p_var >= 0.5
self.f_predict = tensor_utils.compile_function([xs_var], predicted)
self.f_p = tensor_utils.compile_function([xs_var], p_var)
self.l_p = l_p
self.optimizer.update_opt(loss=loss, target=self, network_outputs=[
p_var], inputs=[xs_var, ys_var])
self.tr_optimizer.update_opt(loss=loss, target=self, network_outputs=[p_var],
inputs=[xs_var, ys_var, old_p_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
def __init__(
self,
name,
input_shape,
output_dim,
network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
optimizer=None,
normalize_inputs=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.
"""
if config.TF_NN_SETTRACE:
ipdb.set_trace()
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
if optimizer is None:
optimizer = LbfgsOptimizer(name="optimizer")
self.output_dim = output_dim
self.optimizer = optimizer
if network is None:
network = MLP(
input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
name="network"
)
l_out = network.output_layer
LayersPowered.__init__(self, [l_out])
xs_var = network.input_layer.input_var
ys_var = tf.placeholder(dtype=tf.float32, shape=[
None, output_dim], name="ys")
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
fit_ys_var = L.get_output(
l_out, {network.input_layer: normalized_xs_var})
loss = - tf.reduce_mean(tf.square(fit_ys_var - ys_var))
self.f_predict = tensor_utils.compile_function(
[xs_var], fit_ys_var)
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[fit_ys_var],
)
optimizer_args["inputs"] = [xs_var, ys_var]
self.optimizer.update_opt(**optimizer_args)
self.name = name
self.l_out = l_out
self.normalize_inputs = normalize_inputs
self.x_mean_var = x_mean_var
self.x_std_var = x_std_var
def update_opt(self,
loss,
target,
leq_constraint,
inputs,
extra_inputs=None,
constraint_name="constraint",
*args,
**kwargs):
"""
:param loss: Symbolic expression for the loss function.
:param target: A parameterized object to optimize over. It should implement methods of the
:class:`rllab.core.paramerized.Parameterized` class.
:param leq_constraint: A constraint provided as a tuple (f, epsilon), of the form f(*inputs) <= epsilon.
:param inputs: A list of symbolic variables as inputs, which could be subsampled if needed. It is assumed
that the first dimension of these inputs should correspond to the number of data points
:param extra_inputs: A list of symbolic variables as extra inputs which should not be subsampled
:return: No return value.
"""
inputs = tuple(inputs)
if extra_inputs is None:
extra_inputs = tuple()
else:
extra_inputs = tuple(extra_inputs)
constraint_term, constraint_value = leq_constraint
params = target.get_params(trainable=True)
grads = tf.gradients(loss, xs=params)
for idx, (grad, param) in enumerate(zip(grads, params)):
if grad is None:
grads[idx] = tf.zeros_like(param)
flat_grad = tensor_utils.flatten_tensor_variables(grads)
self._hvp_approach.update_opt(f=constraint_term,
target=target,
inputs=inputs + extra_inputs,
reg_coeff=self._reg_coeff)
self._target = target
self._max_constraint_val = constraint_value
self._constraint_name = constraint_name
self._opt_fun = ext.lazydict(
f_loss=lambda: tensor_utils.compile_function(
inputs=inputs + extra_inputs,
outputs=loss,
log_name="f_loss",
),
f_grad=lambda: tensor_utils.compile_function(
inputs=inputs + extra_inputs,
outputs=flat_grad,
log_name="f_grad",
),
f_constraint=lambda: tensor_utils.compile_function(
inputs=inputs + extra_inputs,
outputs=constraint_term,
log_name="constraint",
),
f_loss_constraint=lambda: tensor_utils.compile_function(
inputs=inputs + extra_inputs,
outputs=[loss, constraint_term],
log_name="f_loss_constraint",
),
)
def __init__(
self,
name,
env_spec,
hidden_dim=32,
feature_network=None,
state_include_action=True,
hidden_nonlinearity=tf.tanh,
gru_layer_cls=L.GRULayer,
learn_std=True,
init_std=1.0,
output_nonlinearity=None,
):
"""
:param env_spec: A spec for the env.
:param hidden_dim: dimension of hidden layer
:param hidden_nonlinearity: nonlinearity used for each hidden layer
:return:
"""
with tf.variable_scope(name):
Serializable.quick_init(self, locals())
super(GaussianGRUPolicy, self).__init__(env_spec)
obs_dim = env_spec.observation_space.flat_dim
action_dim = env_spec.action_space.flat_dim
if state_include_action:
input_dim = obs_dim + action_dim
else:
input_dim = obs_dim
l_input = L.InputLayer(shape=(None, None, input_dim), name="input")
if feature_network is None:
feature_dim = input_dim
l_flat_feature = None
l_feature = l_input
else:
feature_dim = feature_network.output_layer.output_shape[-1]
l_flat_feature = feature_network.output_layer
l_feature = L.OpLayer(
l_flat_feature,
extras=[l_input],
name="reshape_feature",
op=lambda flat_feature, input: tf.reshape(
flat_feature,
tf.pack([
tf.shape(input)[0],
tf.shape(input)[1], feature_dim
])),
shape_op=lambda _, input_shape:
(input_shape[0], input_shape[1], feature_dim))
mean_network = GRUNetwork(input_shape=(feature_dim, ),
input_layer=l_feature,
output_dim=action_dim,
hidden_dim=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
gru_layer_cls=gru_layer_cls,
name="mean_network")
l_log_std = L.ParamLayer(
mean_network.input_layer,
num_units=action_dim,
param=tf.constant_initializer(np.log(init_std)),
name="output_log_std",
trainable=learn_std,
)
l_step_log_std = L.ParamLayer(
mean_network.step_input_layer,
num_units=action_dim,
param=l_log_std.param,
name="step_output_log_std",
trainable=learn_std,
)
self.mean_network = mean_network
self.feature_network = feature_network
self.l_input = l_input
self.state_include_action = state_include_action
flat_input_var = tf.placeholder(dtype=tf.float32,
shape=(None, input_dim),
name="flat_input")
if feature_network is None:
feature_var = flat_input_var
else:
feature_var = L.get_output(
l_flat_feature,
{feature_network.input_layer: flat_input_var})
self.f_step_mean_std = tensor_utils.compile_function(
[
flat_input_var,
#mean_network.step_prev_hidden_layer.input_var,
mean_network.step_prev_state_layer.input_var
],
L.get_output([
mean_network.step_output_layer,
l_step_log_std,
mean_network.step_hidden_layer,
], {mean_network.step_input_layer: feature_var}))
self.l_log_std = l_log_std
self.input_dim = input_dim
self.action_dim = action_dim
self.hidden_dim = hidden_dim
self.prev_actions = None
self.prev_hiddens = None
self.dist = RecurrentDiagonalGaussian(action_dim)
out_layers = [mean_network.output_layer, l_log_std, l_step_log_std]
if feature_network is not None:
out_layers.append(feature_network.output_layer)
LayersPowered.__init__(self, out_layers)
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
