def initialize(self, what: Tensor, config=None, configuration_key=None):
"""Initialize tensor with provided configuration.
The initializers are taken from options "initialize" and "initialize_args".
If set, config and configuration_key overwrite the default configuration used in
this class. When both are set, self can be None.
"""
if config is None:
config = self.config
if configuration_key is None:
configuration_key = self.configuration_key
configurable = Configurable(config, configuration_key)
initialize = configurable.get_option("initialize")
try:
initialize_args_key = "initialize_args." + initialize
initialize_args = configurable.get_option(initialize_args_key)
except KeyError:
initialize_args_key = "initialize_args"
initialize_args = configurable.get_option(initialize_args_key)
# Automatically set arg a (lower bound) for uniform_ if not given
if initialize == "uniform_" and "a" not in initialize_args:
initialize_args["a"] = initialize_args["b"] * -1
KgeBase._initialize(what, initialize, initialize_args)
def _init_configuration(self, config: Config, configuration_key: Optional[str]):
Configurable._init_configuration(self, config, configuration_key)
if not hasattr(self, "model") or not self.model:
if self.configuration_key:
self.model: str = config.get(self.configuration_key + ".type")
else:
self.model: str = config.get("model")
self.configuration_key = self.model
def __init__(self,
config: Config,
dataset: Dataset,
configuration_key=None):
Configurable.__init__(self, config, configuration_key)
torch.nn.Module.__init__(self)
self.dataset = dataset
self.meta: Dict[str,
Any] = dict() #: meta-data stored with this module
def __init__(self, config: Config, dataset: Dataset, configuration_key=None):
Configurable.__init__(self, config, configuration_key)
torch.nn.Module.__init__(self)
self.dataset = dataset
self.meta: Dict[str, Any] = dict() #: meta-data stored with this module
self.backward_compatible_keys = {
"_entity_embedder.embeddings.weight": "_entity_embedder._embeddings.weight",
"_relation_embedder.embeddings.weight": "_relation_embedder._embeddings.weight",
"_base_model._entity_embedder.embeddings.weight": "_base_model._entity_embedder._embeddings.weight",
"_base_model._relation_embedder.embeddings.weight": "_base_model._relation_embedder._embeddings.weight",
}