Open UTimeStrange opened 7 months ago
import torch.nn as nn class SegmentEmbedding(nn.Embedding): def __init__(self, embed_size=512): super().__init__(3, embed_size, padding_idx=0)
This is the source code. First idx is padding, thus only 2 segment is supported. Why does Bert support 2 segments only?
因为一次放入2个句子,需要区分哪些token属于第一个句子,哪些token输入第二个句子, 再加上padding整好三个.
This is the source code. First idx is padding, thus only 2 segment is supported. Why does Bert support 2 segments only?