def from_spec(spec, kwargs=None):
"""
Creates a network from a specification dict.
"""
if isinstance(spec, list) and len(spec) > 0:
# Default case is a list of dict() with each dict describing a layer.
if type(spec[0]) is dict:
network = util.get_object(
obj=spec,
default_object=LayeredNetwork,
kwargs=kwargs
)
# ComplexLayeredNetwork forced for testing
if type(spec[0]) is list:
# Spec contains List of List of Dict(), Complex network specification
# Load "ComplexLayeredNetwork" here to avoid a recurring loop which fails
from tensorforce.core.networks.complex_network import ComplexLayeredNetwork
network = util.get_object(
obj=spec,
default_object=ComplexLayeredNetwork,
kwargs=kwargs
)
else:
network = util.get_object(
obj=spec,
default_object=LayeredNetwork,
kwargs=kwargs
)
# If neither format, invalid spec and will fail on assert
assert isinstance(network, Network)
return network
def network_builder(inputs, summary_level=0):
input_length = len(inputs)
if input_length != 1:
raise TensorForceError('Layered network must have only one input,'
' input length {} given.'.format(input_length))
x = next(iter(inputs.values()))
internal_inputs = []
internal_outputs = []
internal_inits = []
layer_counter = Counter()
for layer_config in layers_config:
if callable(layer_config['type']):
scope = layer_config['type'].__name__ + str(layer_counter[layer_config['type']])
else:
scope = layer_config['type'] + str(layer_counter[layer_config['type']])
x = util.get_object(
obj=layer_config,
predefined=layers,
kwargs=dict(x=x, scope=scope, summary_level=summary_level)
)
layer_counter[layer_config['type']] += 1
if isinstance(x, list) or isinstance(x, tuple):
assert len(x) == 4
internal_inputs.extend(x[1])
internal_outputs.extend(x[2])
internal_inits.extend(x[3])
x = x[0]
if internal_inputs:
return x, internal_inputs, internal_outputs, internal_inits
else:
return x
def from_spec(spec, kwargs=None):
lr_schedule = util.get_object(obj=spec,
predefined_objects=lr_schedulers,
kwargs=kwargs)
assert isinstance(lr_schedule, DecaySchedule)
return lr_schedule
def network_builder(inputs):
input_length = len(inputs)
if input_length != 1:
raise TensorForceError(
'Layered network must have only one input,'
' input length {} given.'.format(input_length))
x = next(iter(inputs.values()))
internal_inputs = []
internal_outputs = []
internal_inits = []
for layer_config in layers_config:
x = util.get_object(obj=layer_config,
predefined=layers,
kwargs=dict(x=x))
if isinstance(x, list) or isinstance(x, tuple):
assert len(x) == 4
internal_inputs.extend(x[1])
internal_outputs.extend(x[2])
internal_inits.extend(x[3])
x = x[0]
if internal_inputs:
return x, internal_inputs, internal_outputs, internal_inits
else:
return x
def from_config(config, kwargs=None):
"""
Creates a solver from a specification dict.
"""
return util.get_object(
obj=config,
predefined=tensorforce.core.optimizers.solvers.solvers,
kwargs=kwargs)
def from_spec(spec):
"""
Creates an exploration object from a specification dict.
"""
exploration = util.get_object(
obj=spec,
predefined_objects=tensorforce.core.explorations.explorations)
assert isinstance(exploration, Exploration)
return exploration
def from_spec(spec, kwargs):
"""
Creates an agent from a specification dict.
"""
agent = util.get_object(obj=spec,
predefined_objects=tensorforce.agents.agents,
kwargs=kwargs)
assert isinstance(agent, Agent)
return agent
def from_spec(spec, kwargs=None):
"""
Creates an optimizer from a specification dict.
"""
optimizer = util.get_object(obj=spec,
predefined_objects=optimizers,
kwargs=kwargs)
assert isinstance(optimizer, Optimizer)
return optimizer
def from_spec(spec, kwargs=None):
"""
Creates a network from a specification dict.
"""
network = util.get_object(obj=spec,
default_object=LayeredNetwork,
kwargs=kwargs)
assert isinstance(network, Network)
return network
def from_spec(spec, kwargs=None):
"""
Creates a layer from a specification dict.
"""
layer = util.get_object(
obj=spec,
predefined_objects=tensorforce.core.networks.layers,
kwargs=kwargs)
assert isinstance(layer, Layer)
return layer
def from_spec(spec, kwargs=None):
"""
Creates a memory from a specification dict.
"""
memory = util.get_object(
obj=spec,
predefined_objects=tensorforce.core.memories.memories,
kwargs=kwargs)
assert isinstance(memory, Memory)
return memory
def from_spec(spec, kwargs=None):
"""
Creates a baseline from a specification dict.
"""
baseline = util.get_object(
obj=spec,
predefined_objects=tensorforce.core.baselines.baselines,
kwargs=kwargs)
assert isinstance(baseline, Baseline)
return baseline
def from_spec(spec, kwargs=None):
"""
Creates a distribution from a specification dict.
"""
distribution = util.get_object(
obj=spec,
predefined_objects=tensorforce.core.distributions.distributions,
kwargs=kwargs)
assert isinstance(distribution, Distribution)
return distribution
def from_config(config):
if not isinstance(config, list):
config = [config]
preprocessing = Preprocessing()
for config in config:
preprocessor = util.get_object(
obj=config,
predefined=tensorforce.core.preprocessing.preprocessors)
preprocessing.add(preprocessor=preprocessor)
return preprocessing
def from_config(config):
"""
Creates a baseline from a configuration dict.
Args:
config:
Returns:
"""
return util.get_object(obj=config,
predefined=tensorforce.core.baselines.baselines)
def __init__(self, layer, scope='tf-layer', summary_labels=(), **kwargs):
"""
Creates a new layer instance of a TensorFlow layer.
Args:
name: The name of the layer, one of 'dense'.
**kwargs: Additional arguments passed on to the TensorFlow layer constructor.
"""
self.layer_spec = layer
self.layer = util.get_object(obj=layer, predefined_objects=TFLayer.tf_layers, kwargs=kwargs)
self.first_scope = None
super(TFLayer, self).__init__(scope=scope, summary_labels=summary_labels)
def from_config(config, kwargs=None):
"""
Creates an optimizer from a configuration object.
Args:
config: Name of optimizer
kwargs: Dict of optimizer hyperparameters
Returns:
"""
return util.get_object(
obj=config,
predefined=tensorforce.core.optimizers.optimizers,
kwargs=kwargs)
def from_spec(spec):
"""
Creates a preprocessing stack from a specification dict.
"""
if not isinstance(spec, list):
spec = [spec]
preprocessing = Preprocessing()
for spec in spec:
preprocessor = util.get_object(obj=spec,
predefined_objects=tensorforce.core.
preprocessing.preprocessors)
assert isinstance(preprocessor, Preprocessor)
preprocessing.add(preprocessor=preprocessor)
return preprocessing
def from_spec(spec):
"""
Creates a preprocessing stack from a specification dict.
"""
if isinstance(spec, dict):
spec = [spec]
stack = PreprocessorStack()
for spec in spec:
preprocessor = util.get_object(obj=spec,
predefined_objects=tensorforce.core.
preprocessing.preprocessors)
assert isinstance(preprocessor, Preprocessor)
stack.preprocessors.append(preprocessor)
return stack
def from_spec(spec, kwargs=None):
"""
Creates a preprocessing stack from a specification dict.
"""
if isinstance(spec, dict):
spec = [spec]
stack = PreprocessorStack()
for preprocessor_spec in spec:
# need to deep copy, otherwise will add first processors spec_ to kwargs to second processor
preprocessor_kwargs = copy.deepcopy(kwargs)
preprocessor = util.get_object(obj=preprocessor_spec,
predefined_objects=tensorforce.core.
preprocessors.preprocessors,
kwargs=preprocessor_kwargs)
assert isinstance(preprocessor, Preprocessor)
stack.preprocessors.append(preprocessor)
return stack
def from_config(config):
return util.get_object(
obj=config,
predefined=tensorforce.core.explorations.explorations
)
def from_config(config, kwargs=None):
return util.get_object(obj=config,
predefined=tensorforce.core.memories.memories,
kwargs=kwargs)
def from_config(config, kwargs=None):
return util.get_object(
obj=config,
predefined=tensorforce.core.distributions.distributions,
kwargs=kwargs)
def from_config(config, kwargs=None):
return util.get_object(
obj=config,
predefined=tensorforce.core.optimizers.optimizers,
kwargs=kwargs
)
