def __init__(
self,
config: Config,
dataset: Dataset,
configuration_key=None,
init_for_load_only=False,
):
self._init_configuration(config, configuration_key)
ent_emb_dim = self.get_option("entity_embedder.dim")
ent_emb_conf_key = self.configuration_key + ".entity_embedder"
round_ent_emb_dim_to = self.get_option("entity_embedder.round_dim_to")
if len(round_ent_emb_dim_to) > 0:
ent_emb_dim = round_to_points(round_ent_emb_dim_to, ent_emb_dim)
config.set(ent_emb_conf_key + ".dim", ent_emb_dim, log=True)
rescal_set_relation_embedder_dim(
config, dataset, self.configuration_key + ".relation_embedder")
super().__init__(
config=config,
dataset=dataset,
scorer=RescalScorer,
configuration_key=self.configuration_key,
init_for_load_only=init_for_load_only,
)
def __init__(self, config: Config, dataset: Dataset,
configuration_key: str, vocab_size: int):
super().__init__(config, dataset, configuration_key)
# read config
self.normalize_p = self.get_option("normalize.p")
self.normalize_with_grad = self.get_option("normalize.with_grad")
self.regularize = self.check_option("regularize", ["", "lp"])
self.sparse = self.get_option("sparse")
self.config.check("train.trace_level", ["batch", "epoch"])
self.vocab_size = vocab_size
round_embedder_dim_to = self.get_option("round_dim_to")
if len(round_embedder_dim_to) > 0:
self.dim = round_to_points(round_embedder_dim_to, self.dim)
# setup embedder
self.embeddings = torch.nn.Embedding(self.vocab_size,
self.dim,
sparse=self.sparse)
# initialize weights
init_ = self.get_option("initialize")
try:
init_args = self.get_option("initialize_args." + init_)
except KeyError:
init_args = self.get_option("initialize_args")
# Automatically set arg a (lower bound) for uniform_ if not given
if init_ == "uniform_" and "a" not in init_args:
init_args["a"] = init_args["b"] * -1
self.set_option("initialize_args.a", init_args["a"], log=True)
self.initialize(self.embeddings.weight.data, init_, init_args)
# TODO handling negative dropout because using it with ax searches for now
dropout = self.get_option("dropout")
if dropout < 0:
if config.get("train.auto_correct"):
config.log("Setting {}.dropout to 0, "
"was set to {}.".format(configuration_key, dropout))
dropout = 0
self.dropout = torch.nn.Dropout(dropout)
def __init__(
self,
config: Config,
dataset: Dataset,
configuration_key: str,
vocab_size: int,
init_for_load_only=False,
):
super().__init__(config,
dataset,
configuration_key,
init_for_load_only=init_for_load_only)
# read config
self.normalize_p = self.get_option("normalize.p")
self.regularize = self.check_option("regularize", ["", "lp"])
self.sparse = self.get_option("sparse")
self.config.check("train.trace_level", ["batch", "epoch"])
self.vocab_size = vocab_size
round_embedder_dim_to = self.get_option("round_dim_to")
if len(round_embedder_dim_to) > 0:
self.dim = round_to_points(round_embedder_dim_to, self.dim)
self._embeddings = torch.nn.Embedding(
self.vocab_size,
self.dim,
sparse=self.sparse,
)
if not init_for_load_only:
# initialize weights
self._init_embeddings(self._embeddings.weight.data)
self._normalize_embeddings()
# TODO handling negative dropout because using it with ax searches for now
dropout = self.get_option("dropout")
if dropout < 0:
if config.get("train.auto_correct"):
config.log("Setting {}.dropout to 0, "
"was set to {}.".format(configuration_key, dropout))
dropout = 0
self.dropout = torch.nn.Dropout(dropout)
