def initialize(self, custom_getter):
super(PGModel, self).initialize(custom_getter)
# Baseline
if self.baseline_spec is None:
assert self.baseline_mode is None
elif all(name in self.states_spec for name in self.baseline_spec):
# Implies AggregatedBaseline.
assert self.baseline_mode == 'states'
self.baseline = AggregatedBaseline(baselines=self.baseline_spec)
else:
assert self.baseline_mode is not None
self.baseline = Baseline.from_spec(
spec=self.baseline_spec,
kwargs=dict(summary_labels=self.summary_labels))
# Baseline optimizer
if self.baseline_optimizer_spec is not None:
assert self.baseline_mode is not None
self.baseline_optimizer = Optimizer.from_spec(
spec=self.baseline_optimizer_spec)
# TODO: Baseline internal states !!! (see target_network q_model)
# Reward estimation
self.fn_reward_estimation = tf.make_template(
name_='reward-estimation',
func_=self.tf_reward_estimation,
custom_getter_=custom_getter)
# Baseline loss
self.fn_baseline_loss = tf.make_template(name_='baseline-loss',
func_=self.tf_baseline_loss,
custom_getter_=custom_getter)
def __init__(self, baselines, scope='aggregated-baseline', summary_labels=()):
"""
Aggregated baseline.
Args:
baselines: Dict of per-state baseline specification dicts
"""
self.baselines = dict()
for name in sorted(baselines):
self.baselines[name] = Baseline.from_spec(
spec=baselines[name],
kwargs=dict(summary_labels=summary_labels))
self.linear = Linear(size=1, bias=0.0, scope='prediction', summary_labels=summary_labels)
super(AggregatedBaseline, self).__init__(scope, summary_labels)
def __init__(self, baselines, scope='aggregated-baseline', summary_labels=()):
"""
Aggregated baseline.
Args:
baselines: Dict of per-state baseline specification dicts
"""
with tf.name_scope(name=scope):
self.baselines = dict()
for name, baseline_spec in baselines.items():
with tf.name_scope(name=(name + '-baseline')):
self.baselines[name] = Baseline.from_spec(
spec=baseline_spec,
kwargs=dict(summary_labels=summary_labels)
)
self.linear = Linear(size=1, bias=0.0, scope='prediction')
super(AggregatedBaseline, self).__init__(scope, summary_labels)
def __init__(self, states_spec, actions_spec, network_spec, config):
# Baseline mode
assert config.baseline_mode is None or config.baseline_mode in (
'states', 'network')
self.baseline_mode = config.baseline_mode
with tf.name_scope(name=config.scope):
# Baseline
if config.baseline is None:
assert self.baseline_mode is None
self.baseline = None
elif all(name in states_spec for name in config.baseline):
# Implies AggregatedBaseline
assert self.baseline_mode == 'states'
self.baseline = AggregatedBaseline(baselines=config.baseline)
else:
assert self.baseline_mode is not None
self.baseline = Baseline.from_spec(
spec=config.baseline,
kwargs=dict(summary_labels=config.summary_labels))
# Baseline optimizer
if config.baseline_optimizer is None:
self.baseline_optimizer = None
else:
assert self.baseline_mode is not None
self.baseline_optimizer = Optimizer.from_spec(
spec=config.baseline_optimizer)
# Generalized advantage function
assert config.gae_lambda is None or (
0.0 <= config.gae_lambda <= 1.0 and self.baseline_mode is not None)
self.gae_lambda = config.gae_lambda
super(PGModel, self).__init__(states_spec=states_spec,
actions_spec=actions_spec,
network_spec=network_spec,
config=config)
def initialize(self, custom_getter):
super(PGModel, self).initialize(custom_getter)
# Baseline
if self.baseline is None:
assert self.baseline_mode is None
self.baseline = None
elif all(name in self.states_spec for name in self.baseline):
# Implies AggregatedBaseline.
assert self.baseline_mode == 'states'
self.baseline = AggregatedBaseline(baselines=self.baseline)
else:
assert self.baseline_mode is not None
self.baseline = Baseline.from_spec(
spec=self.baseline,
kwargs=dict(summary_labels=self.summary_labels))
# Baseline optimizer
if self.baseline_optimizer is None:
self.baseline_optimizer = None
else:
assert self.baseline_mode is not None
self.baseline_optimizer = Optimizer.from_spec(
spec=self.baseline_optimizer)
# TODO: Baseline internal states !!! (see target_network q_model)
# Reward estimation
self.fn_reward_estimation = tf.make_template(
name_=(self.scope + '/reward-estimation'),
func_=self.tf_reward_estimation,
custom_getter_=custom_getter)
# PG loss per instance function
self.fn_pg_loss_per_instance = tf.make_template(
name_=(self.scope + '/pg-loss-per-instance'),
func_=self.tf_pg_loss_per_instance,
custom_getter_=custom_getter)