def __init__(self, elements, nodes, emb_size, compact_dst=True):
ElementEmbedderBase.__init__(self,
elements=elements,
nodes=nodes,
compact_dst=compact_dst)
nn.Module.__init__(self)
Scorer.__init__(self,
num_embs=len(self.elements["dst"].unique()),
emb_size=emb_size,
src2dst=self.element_lookup)
self.emb_size = emb_size
n_elems = self.elements['emb_id'].unique().size
self.embed = nn.Embedding(n_elems, emb_size)
self.norm = nn.LayerNorm(emb_size)
def sample_negative(self, size, ids=None, strategy="closest"):
if strategy == "w2v" or self.scorer_index is None:
negative = ElementEmbedderBase.sample_negative(self, size)
else:
negative = Scorer.sample_closest_negative(self,
ids,
k=size // len(ids))
assert len(negative) == size
return negative
def __init__(self,
elements,
nodes,
emb_size,
num_buckets=5000,
max_len=100,
gram_size=3):
ElementEmbedderBase.__init__(self,
elements=elements,
nodes=nodes,
compact_dst=False)
nn.Module.__init__(self)
Scorer.__init__(self,
num_embs=len(self.elements["dst"].unique()),
emb_size=emb_size,
src2dst=self.element_lookup)
self.gram_size = gram_size
self.emb_size = emb_size
self.init_subwords(elements, num_buckets=num_buckets, max_len=max_len)
def sample_negative(self,
size,
ids=None,
strategy="w2v"): # TODO switch to w2v?
if strategy == "w2v":
negative = ElementEmbedderBase.sample_negative(self, size)
else:
negative = Scorer.sample_closest_negative(self,
ids,
k=size // len(ids))
assert len(negative) == size
return negative
def sample_negative(self, size, ids=None, strategy="closest"):
# TODO
# Try other distributions
if strategy == "w2v":
negative = ElementEmbedderBase.sample_negative(self, size)
else:
### negative = random.choices(Scorer.sample_closest_negative(self, ids), k=size)
negative = Scorer.sample_closest_negative(self,
ids,
k=size // len(ids))
assert len(negative) == size
return torch.LongTensor(negative)
def __init__(self,
elements,
nodes,
compact_dst=True,
dst_to_global=True,
emb_size=None,
device="cpu",
method="inner_prod",
nn_index="brute",
ns_groups=None):
assert emb_size is not None
ElementEmbedderBase.__init__(self,
elements=elements,
nodes=nodes,
compact_dst=compact_dst,
dst_to_global=dst_to_global)
Scorer.__init__(self,
num_embs=len(self.elements["dst"].unique()),
emb_size=emb_size,
src2dst=self.element_lookup,
device=device,
method=method,
index_backend=nn_index,
ns_groups=ns_groups)
def prepare_index(self):
self.set_embed()
Scorer.prepare_index(self)