Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
PaddlePaddle
DeepSpeech
提交
0a5cc555
D
DeepSpeech
项目概览
PaddlePaddle
/
DeepSpeech
9 个月 前同步成功
通知
200
Star
8425
Fork
1598
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
245
列表
看板
标记
里程碑
合并请求
3
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
D
DeepSpeech
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
245
Issue
245
列表
看板
标记
里程碑
合并请求
3
合并请求
3
Pages
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
提交
Issue看板
前往新版Gitcode,体验更适合开发者的 AI 搜索 >>
提交
0a5cc555
编写于
7月 14, 2023
作者:
H
Hui Zhang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
rope for streaming decoding
上级
b56fb85c
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
34 addition
and
14 deletion
+34
-14
examples/aishell/asr1/RESULTS.md
examples/aishell/asr1/RESULTS.md
+23
-10
paddlespeech/s2t/modules/attention.py
paddlespeech/s2t/modules/attention.py
+9
-3
paddlespeech/s2t/modules/embedding.py
paddlespeech/s2t/modules/embedding.py
+1
-0
paddlespeech/s2t/modules/encoder.py
paddlespeech/s2t/modules/encoder.py
+1
-1
未找到文件。
examples/aishell/asr1/RESULTS.md
浏览文件 @
0a5cc555
...
...
@@ -4,7 +4,10 @@
paddle version: 2.5.0
paddlespeech version: 1.5.0
Need set
`decoding.decoding_chunk_size=16`
when decoding.
Tesla V100-SXM2-32GB: 1 node, 4 card
Global BachSize: 32
*
4
Training Done: 1 day, 12:56:39.639646
### `decoding.decoding_chunk_size=16`
> chunk_size=16, ((16 - 1) * 4 + 7) * 10ms = (16 * 4 + 3) * 10ms = 670ms
...
...
@@ -15,15 +18,14 @@ Need set `decoding.decoding_chunk_size=16` when decoding.
| roformer | 44.80M | conf/chunk_roformer.yaml | spec_aug | test | ctc_prefix_beam_search | 16, -1 | - | |
| roformer | 44.80M | conf/chunk_roformer.yaml | spec_aug | test | attention_rescoring | 16, -1 | - | |
## Conformer
paddle version: 2.2.2
paddlespeech version: 1.0.1
| Model | Params | Config | Augmentation| Test set | Decode method | Loss | CER |
| --- | --- | --- | --- | --- | --- | --- | --- |
| conformer | 47.07M | conf/conformer.yaml | spec_aug | test | attention | - | 0.0522 |
| conformer | 47.07M | conf/conformer.yaml | spec_aug | test | ctc_greedy_search | - | 0.0481 |
| conformer | 47.07M | conf/conformer.yaml | spec_aug | test | ctc_prefix_beam_search | - | 0.0480 |
| conformer | 47.07M | conf/conformer.yaml | spec_aug | test | attention_rescoring | - | 0.0460 |
### `decoding.decoding_chunk_size=-1`
| Model | Params | Config | Augmentation| Test set | Decode method | Chunk Size & Left Chunks | Loss | CER |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| roformer | 44.80M | conf/chunk_roformer.yaml | spec_aug | test | attention | -1, -1 | - | 5.39 |
| roformer | 44.80M | conf/chunk_roformer.yaml | spec_aug | test | ctc_greedy_search | -1, -1 | - | 5.51 |
| roformer | 44.80M | conf/chunk_roformer.yaml | spec_aug | test | ctc_prefix_beam_search | -1, -1 | - | 5.51 |
| roformer | 44.80M | conf/chunk_roformer.yaml | spec_aug | test | attention_rescoring | -1, -1 | - | 4.99 |
## Conformer Streaming
...
...
@@ -39,6 +41,17 @@ Need set `decoding.decoding_chunk_size=16` when decoding.
| conformer | 47.06M | conf/chunk_conformer.yaml | spec_aug | test | attention_rescoring | 16, -1 | - | 0.051968 |
## Conformer
paddle version: 2.2.2
paddlespeech version: 1.0.1
| Model | Params | Config | Augmentation| Test set | Decode method | Loss | CER |
| --- | --- | --- | --- | --- | --- | --- | --- |
| conformer | 47.07M | conf/conformer.yaml | spec_aug | test | attention | - | 0.0522 |
| conformer | 47.07M | conf/conformer.yaml | spec_aug | test | ctc_greedy_search | - | 0.0481 |
| conformer | 47.07M | conf/conformer.yaml | spec_aug | test | ctc_prefix_beam_search | - | 0.0480 |
| conformer | 47.07M | conf/conformer.yaml | spec_aug | test | attention_rescoring | - | 0.0460 |
## Transformer
| Model | Params | Config | Augmentation| Test set | Decode method | Loss | CER |
...
...
paddlespeech/s2t/modules/attention.py
浏览文件 @
0a5cc555
...
...
@@ -15,8 +15,8 @@
# Modified from wenet(https://github.com/wenet-e2e/wenet)
"""Multi-Head Attention layer definition."""
import
math
from
typing
import
Tuple
from
typing
import
List
from
typing
import
Tuple
import
paddle
from
paddle
import
nn
...
...
@@ -428,7 +428,7 @@ class RoPERelPositionMultiHeadedAttention(MultiHeadedAttention):
# (B,H,T,D)
ndim
=
tensors
[
0
].
dim
()
_
,
H
,
T
,
D
=
tensors
[
0
].
shape
_
,
H
,
T
,
D
=
tensors
[
0
].
shape
# sinusoidal shape same with tensors[0]
# [B,T,D] -> [B,T,H,D/H] -> (B,H,T,D/H)
...
...
@@ -476,6 +476,7 @@ class RoPERelPositionMultiHeadedAttention(MultiHeadedAttention):
where `cache_t == chunk_size * num_decoding_left_chunks`
and `head * d_k == size`
"""
q
,
k
,
v
=
self
.
forward_qkv
(
query
,
key
,
value
)
# q = q.transpose([0, 2, 1, 3]) # (batch, time1, head, d_k)
...
...
@@ -504,7 +505,12 @@ class RoPERelPositionMultiHeadedAttention(MultiHeadedAttention):
new_cache
=
paddle
.
concat
((
k
,
v
),
axis
=-
1
)
# f{q,k}(x_m, m) = R^d_{\theta, m} W_{q,k} x_m, m is position index
q
,
k
=
self
.
apply_rotary_position_embeddings
(
pos_emb
,
q
,
k
)
# q_t always is chunk_size
q_t
=
q
.
shape
[
2
]
q
=
self
.
apply_rotary_position_embeddings
(
pos_emb
[:,
-
q_t
:,
:],
q
)
# k will increase when in streaming decoding.
k
=
self
.
apply_rotary_position_embeddings
(
pos_emb
,
k
)
# dot(q, k)
scores
=
paddle
.
matmul
(
q
,
k
,
transpose_y
=
True
)
/
math
.
sqrt
(
self
.
d_k
)
return
self
.
forward_attention
(
v
,
scores
,
mask
),
new_cache
paddlespeech/s2t/modules/embedding.py
浏览文件 @
0a5cc555
...
...
@@ -164,6 +164,7 @@ class RelPositionalEncoding(PositionalEncoding):
assert
offset
+
x
.
shape
[
1
]
<
self
.
max_len
,
"offset: {} + x.shape[1]: {} is larger than the max_len: {}"
.
format
(
offset
,
x
.
shape
[
1
],
self
.
max_len
)
x
=
x
*
self
.
xscale
pos_emb
=
self
.
pe
[:,
offset
:
offset
+
x
.
shape
[
1
]]
return
self
.
dropout
(
x
),
self
.
dropout
(
pos_emb
)
...
...
paddlespeech/s2t/modules/encoder.py
浏览文件 @
0a5cc555
...
...
@@ -233,7 +233,7 @@ class BaseEncoder(nn.Layer):
xs
=
self
.
global_cmvn
(
xs
)
# before embed, xs=(B, T, D1), pos_emb=(B=1, T, D)
xs
,
pos_emb
,
_
=
self
.
embed
(
xs
,
tmp_masks
,
offset
=
offset
)
xs
,
_
,
_
=
self
.
embed
(
xs
,
tmp_masks
,
offset
=
offset
)
# after embed, xs=(B=1, chunk_size, hidden-dim)
elayers
,
_
,
cache_t1
,
_
=
att_cache
.
shape
...
...
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录