Skip to content
项目
群组
代码片段
帮助
当前项目
正在载入...
登录 / 注册
切换导航面板
F
Fairseq-S2T
概览
Overview
Details
Activity
Cycle Analytics
版本库
Repository
Files
Commits
Branches
Tags
Contributors
Graph
Compare
Charts
问题
0
Issues
0
列表
Board
标记
里程碑
合并请求
0
Merge Requests
0
CI / CD
CI / CD
流水线
作业
日程表
图表
维基
Wiki
代码片段
Snippets
成员
Collapse sidebar
Close sidebar
活动
图像
聊天
创建新问题
作业
提交
Issue Boards
Open sidebar
xuchen
Fairseq-S2T
Commits
8645e75b
Commit
8645e75b
authored
Apr 05, 2021
by
xuchen
Browse files
Options
Browse Files
Download
Email Patches
Plain Diff
fix the bug of the RPE (TODO)
parent
8b47f344
隐藏空白字符变更
内嵌
并排
正在显示
1 个修改的文件
包含
3 行增加
和
4 行删除
+3
-4
fairseq/modules/rel_position_multihead_attention.py
+3
-4
没有找到文件。
fairseq/modules/rel_position_multihead_attention.py
查看文件 @
8645e75b
...
@@ -261,7 +261,7 @@ class RelPositionMultiheadAttention(MultiheadAttention):
...
@@ -261,7 +261,7 @@ class RelPositionMultiheadAttention(MultiheadAttention):
)
)
# (batch * head, time1, d_k)
# (batch * head, time1, d_k)
q_with_bias_v
=
(
q_with_bias_v
=
(
(
q
+
self
.
pos_bias_v
)
.
transpose
(
0
,
1
)
(
q
+
self
.
pos_bias_v
)
.
transpose
(
1
,
2
)
.
contiguous
()
.
contiguous
()
.
view
(
bsz
*
self
.
num_heads
,
tgt_len
,
self
.
head_dim
)
.
view
(
bsz
*
self
.
num_heads
,
tgt_len
,
self
.
head_dim
)
)
)
...
@@ -280,8 +280,7 @@ class RelPositionMultiheadAttention(MultiheadAttention):
...
@@ -280,8 +280,7 @@ class RelPositionMultiheadAttention(MultiheadAttention):
"""Compute relative positional encoding.
"""Compute relative positional encoding.
Args:
Args:
x (torch.Tensor): Input tensor (batch, head, time1, 2*time1-1).
x (torch.Tensor): Input tensor (batch * head, time1, time2).
time1 means the length of query vector.
zero_triu (bool): If true, return the lower triangular part of
zero_triu (bool): If true, return the lower triangular part of
the matrix.
the matrix.
Returns:
Returns:
...
@@ -302,7 +301,7 @@ class RelPositionMultiheadAttention(MultiheadAttention):
...
@@ -302,7 +301,7 @@ class RelPositionMultiheadAttention(MultiheadAttention):
x
=
x
*
torch
.
tril
(
ones
,
x
.
size
(
2
)
-
x
.
size
(
1
))[
None
,
:,
:]
x
=
x
*
torch
.
tril
(
ones
,
x
.
size
(
2
)
-
x
.
size
(
1
))[
None
,
:,
:]
return
x
return
x
matrix_bd
=
rel_shift
(
matrix_bd
)
#
matrix_bd = rel_shift(matrix_bd)
attn_weights
=
(
matrix_ac
+
matrix_bd
)
*
self
.
scaling
attn_weights
=
(
matrix_ac
+
matrix_bd
)
*
self
.
scaling
...
...
编写
预览
Markdown
格式
0%
重试
或
添加新文件
添加附件
取消
您添加了
0
人
到此讨论。请谨慎行事。
请先完成此评论的编辑!
取消
请
注册
或者
登录
后发表评论