test/fastNLP
1
0
Fork 0
mirror of https://gitee.com/fastnlp/fastNLP.git synced 2024-12-02 12:17:35 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

修复bert embedding的bug

Browse Source
This commit is contained in:
yh 2019-09-05 00:21:46 +08:00
parent 60a535db08
commit 14d048f340
1 changed files with 4 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
fastNLP/embeddings/bert_embedding.py
View File

@ -420,11 +420,11 @@ class _WordBertModel(nn.Module):
if self.pool_method == 'first':
batch_word_pieces_cum_length = batch_word_pieces_cum_length[:, :seq_len.max()]
batch_word_pieces_cum_length.masked_fill_(batch_word_pieces_cum_length.ge(word_piece_length), 0)
batch_indexes = batch_indexes[:, None].expand((batch_size, batch_word_pieces_cum_length.size(1)))
_batch_indexes = batch_indexes[:, None].expand((batch_size, batch_word_pieces_cum_length.size(1)))
elif self.pool_method == 'last':
batch_word_pieces_cum_length = batch_word_pieces_cum_length[:, 1:seq_len.max()+1] - 1
batch_word_pieces_cum_length.masked_fill_(batch_word_pieces_cum_length.ge(word_piece_length), 0)
batch_indexes = batch_indexes[:, None].expand((batch_size, batch_word_pieces_cum_length.size(1)))
_batch_indexes = batch_indexes[:, None].expand((batch_size, batch_word_pieces_cum_length.size(1)))
for l_index, l in enumerate(self.layers):
output_layer = bert_outputs[l]
@ -437,12 +437,12 @@ class _WordBertModel(nn.Module):
# 从word_piece collapse到word的表示
truncate_output_layer = output_layer[:, 1:-1] # 删除[CLS]与[SEP] batch_size x len x hidden_size
if self.pool_method == 'first':
tmp = truncate_output_layer[batch_indexes, batch_word_pieces_cum_length]
tmp = truncate_output_layer[_batch_indexes, batch_word_pieces_cum_length]
tmp = tmp.masked_fill(word_mask[:, :batch_word_pieces_cum_length.size(1), None].eq(0), 0)
outputs[l_index, :, s_shift:batch_word_pieces_cum_length.size(1)+s_shift] = tmp
elif self.pool_method == 'last':
tmp = truncate_output_layer[batch_indexes, batch_word_pieces_cum_length]
tmp = truncate_output_layer[_batch_indexes, batch_word_pieces_cum_length]
tmp = tmp.masked_fill(word_mask[:, :batch_word_pieces_cum_length.size(1), None].eq(0), 0)
outputs[l_index, :, s_shift:batch_word_pieces_cum_length.size(1)+s_shift] = tmp
elif self.pool_method == 'max':