def _pick_embed(embedding: nn.Embedding, indices: LongTensor,
mask: torch.ByteTensor):
indices = indices.clone()
indices[~mask] = 0
embed = embedding.forward(indices)
embed *= mask.to(dtype=torch.float).unsqueeze(-1).expand_as(embed)
return embed
class PackedEmbedding(Module):
def __init__(self, dimension: int, vocabulary: Vocabulary):
super().__init__()
self.embedding = Embedding(
num_embeddings=len(vocabulary), embedding_dim=dimension
)
def forward(self, inputs: PackedSequence) -> PackedSequence: # type: ignore
"""Embed the packed sequence given as input."""
return PackedSequence(self.embedding.forward(inputs.data), inputs.batch_sizes)
