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
50f10f37
D
DeepSpeech
项目概览
PaddlePaddle
/
DeepSpeech
大约 2 年 前同步成功
通知
210
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
提交
50f10f37
编写于
8月 20, 2021
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
support replace with mean by aug
上级
86d08f99
变更
20
隐藏空白更改
内联
并排
Showing
20 changed file
with
66 addition
and
72 deletion
+66
-72
deepspeech/__init__.py
deepspeech/__init__.py
+1
-40
deepspeech/frontend/augmentor/impulse_response.py
deepspeech/frontend/augmentor/impulse_response.py
+1
-1
deepspeech/frontend/augmentor/noise_perturb.py
deepspeech/frontend/augmentor/noise_perturb.py
+1
-1
deepspeech/frontend/augmentor/online_bayesian_normalization.py
...peech/frontend/augmentor/online_bayesian_normalization.py
+1
-1
deepspeech/frontend/augmentor/resample.py
deepspeech/frontend/augmentor/resample.py
+1
-1
deepspeech/frontend/augmentor/shift_perturb.py
deepspeech/frontend/augmentor/shift_perturb.py
+1
-1
deepspeech/frontend/augmentor/spec_augment.py
deepspeech/frontend/augmentor/spec_augment.py
+15
-6
deepspeech/frontend/augmentor/speed_perturb.py
deepspeech/frontend/augmentor/speed_perturb.py
+1
-1
deepspeech/frontend/augmentor/volume_perturb.py
deepspeech/frontend/augmentor/volume_perturb.py
+1
-1
examples/aishell/s0/conf/augmentation.json
examples/aishell/s0/conf/augmentation.json
+2
-1
examples/aishell/s1/conf/augmentation.json
examples/aishell/s1/conf/augmentation.json
+2
-1
examples/aug_conf/augmentation.json
examples/aug_conf/augmentation.json
+0
-10
examples/augmentation/augmentation.json
examples/augmentation/augmentation.json
+2
-1
examples/callcenter/s1/conf/augmentation.json
examples/callcenter/s1/conf/augmentation.json
+2
-1
examples/librispeech/s0/conf/augmentation.json
examples/librispeech/s0/conf/augmentation.json
+2
-1
examples/librispeech/s1/conf/augmentation.json
examples/librispeech/s1/conf/augmentation.json
+2
-1
examples/librispeech/s2/conf/augmentation.json
examples/librispeech/s2/conf/augmentation.json
+2
-1
examples/timit/s1/conf/augmentation.json
examples/timit/s1/conf/augmentation.json
+2
-1
examples/tiny/s0/conf/augmentation.json
examples/tiny/s0/conf/augmentation.json
+25
-0
examples/tiny/s1/conf/augmentation.json
examples/tiny/s1/conf/augmentation.json
+2
-1
未找到文件。
deepspeech/__init__.py
浏览文件 @
50f10f37
...
@@ -352,45 +352,6 @@ if not hasattr(paddle.Tensor, 'tolist'):
...
@@ -352,45 +352,6 @@ if not hasattr(paddle.Tensor, 'tolist'):
"register user tolist to paddle.Tensor, remove this when fixed!"
)
"register user tolist to paddle.Tensor, remove this when fixed!"
)
setattr
(
paddle
.
Tensor
,
'tolist'
,
tolist
)
setattr
(
paddle
.
Tensor
,
'tolist'
,
tolist
)
########### hcak paddle.nn.functional #############
def
glu
(
x
:
paddle
.
Tensor
,
axis
=-
1
)
->
paddle
.
Tensor
:
"""The gated linear unit (GLU) activation."""
a
,
b
=
x
.
split
(
2
,
axis
=
axis
)
act_b
=
F
.
sigmoid
(
b
)
return
a
*
act_b
if
not
hasattr
(
paddle
.
nn
.
functional
,
'glu'
):
logger
.
warn
(
"register user glu to paddle.nn.functional, remove this when fixed!"
)
setattr
(
paddle
.
nn
.
functional
,
'glu'
,
glu
)
# def softplus(x):
# """Softplus function."""
# if hasattr(paddle.nn.functional, 'softplus'):
# #return paddle.nn.functional.softplus(x.float()).type_as(x)
# return paddle.nn.functional.softplus(x)
# else:
# raise NotImplementedError
# def gelu_accurate(x):
# """Gaussian Error Linear Units (GELU) activation."""
# # [reference] https://github.com/pytorch/fairseq/blob/e75cff5f2c1d62f12dc911e0bf420025eb1a4e33/fairseq/modules/gelu.py
# if not hasattr(gelu_accurate, "_a"):
# gelu_accurate._a = math.sqrt(2 / math.pi)
# return 0.5 * x * (1 + paddle.tanh(gelu_accurate._a *
# (x + 0.044715 * paddle.pow(x, 3))))
# def gelu(x):
# """Gaussian Error Linear Units (GELU) activation."""
# if hasattr(nn.functional, 'gelu'):
# #return nn.functional.gelu(x.float()).type_as(x)
# return nn.functional.gelu(x)
# else:
# return x * 0.5 * (1.0 + paddle.erf(x / math.sqrt(2.0)))
########### hcak paddle.nn #############
########### hcak paddle.nn #############
class
GLU
(
nn
.
Layer
):
class
GLU
(
nn
.
Layer
):
...
@@ -401,7 +362,7 @@ class GLU(nn.Layer):
...
@@ -401,7 +362,7 @@ class GLU(nn.Layer):
self
.
dim
=
dim
self
.
dim
=
dim
def
forward
(
self
,
xs
):
def
forward
(
self
,
xs
):
return
glu
(
xs
,
dim
=
self
.
dim
)
return
F
.
glu
(
xs
,
dim
=
self
.
dim
)
if
not
hasattr
(
paddle
.
nn
,
'GLU'
):
if
not
hasattr
(
paddle
.
nn
,
'GLU'
):
...
...
deepspeech/frontend/augmentor/impulse_response.py
浏览文件 @
50f10f37
...
@@ -32,7 +32,7 @@ class ImpulseResponseAugmentor(AugmentorBase):
...
@@ -32,7 +32,7 @@ class ImpulseResponseAugmentor(AugmentorBase):
def
__call__
(
self
,
x
,
uttid
=
None
,
train
=
True
):
def
__call__
(
self
,
x
,
uttid
=
None
,
train
=
True
):
if
not
train
:
if
not
train
:
return
return
x
self
.
transform_audio
(
x
)
self
.
transform_audio
(
x
)
return
x
return
x
...
...
deepspeech/frontend/augmentor/noise_perturb.py
浏览文件 @
50f10f37
...
@@ -38,7 +38,7 @@ class NoisePerturbAugmentor(AugmentorBase):
...
@@ -38,7 +38,7 @@ class NoisePerturbAugmentor(AugmentorBase):
def
__call__
(
self
,
x
,
uttid
=
None
,
train
=
True
):
def
__call__
(
self
,
x
,
uttid
=
None
,
train
=
True
):
if
not
train
:
if
not
train
:
return
return
x
self
.
transform_audio
(
x
)
self
.
transform_audio
(
x
)
return
x
return
x
...
...
deepspeech/frontend/augmentor/online_bayesian_normalization.py
浏览文件 @
50f10f37
...
@@ -46,7 +46,7 @@ class OnlineBayesianNormalizationAugmentor(AugmentorBase):
...
@@ -46,7 +46,7 @@ class OnlineBayesianNormalizationAugmentor(AugmentorBase):
def
__call__
(
self
,
x
,
uttid
=
None
,
train
=
True
):
def
__call__
(
self
,
x
,
uttid
=
None
,
train
=
True
):
if
not
train
:
if
not
train
:
return
return
x
self
.
transform_audio
(
x
)
self
.
transform_audio
(
x
)
return
x
return
x
...
...
deepspeech/frontend/augmentor/resample.py
浏览文件 @
50f10f37
...
@@ -33,7 +33,7 @@ class ResampleAugmentor(AugmentorBase):
...
@@ -33,7 +33,7 @@ class ResampleAugmentor(AugmentorBase):
def
__call__
(
self
,
x
,
uttid
=
None
,
train
=
True
):
def
__call__
(
self
,
x
,
uttid
=
None
,
train
=
True
):
if
not
train
:
if
not
train
:
return
return
x
self
.
transform_audio
(
x
)
self
.
transform_audio
(
x
)
return
x
return
x
...
...
deepspeech/frontend/augmentor/shift_perturb.py
浏览文件 @
50f10f37
...
@@ -33,7 +33,7 @@ class ShiftPerturbAugmentor(AugmentorBase):
...
@@ -33,7 +33,7 @@ class ShiftPerturbAugmentor(AugmentorBase):
def
__call__
(
self
,
x
,
uttid
=
None
,
train
=
True
):
def
__call__
(
self
,
x
,
uttid
=
None
,
train
=
True
):
if
not
train
:
if
not
train
:
return
return
x
self
.
transform_audio
(
x
)
self
.
transform_audio
(
x
)
return
x
return
x
...
...
deepspeech/frontend/augmentor/spec_augment.py
浏览文件 @
50f10f37
...
@@ -41,7 +41,8 @@ class SpecAugmentor(AugmentorBase):
...
@@ -41,7 +41,8 @@ class SpecAugmentor(AugmentorBase):
W
=
40
,
W
=
40
,
adaptive_number_ratio
=
0
,
adaptive_number_ratio
=
0
,
adaptive_size_ratio
=
0
,
adaptive_size_ratio
=
0
,
max_n_time_masks
=
20
):
max_n_time_masks
=
20
,
replace_with_zero
=
True
):
"""SpecAugment class.
"""SpecAugment class.
Args:
Args:
rng (random.Random): random generator object.
rng (random.Random): random generator object.
...
@@ -54,9 +55,11 @@ class SpecAugmentor(AugmentorBase):
...
@@ -54,9 +55,11 @@ class SpecAugmentor(AugmentorBase):
adaptive_number_ratio (float): adaptive multiplicity ratio for time masking
adaptive_number_ratio (float): adaptive multiplicity ratio for time masking
adaptive_size_ratio (float): adaptive size ratio for time masking
adaptive_size_ratio (float): adaptive size ratio for time masking
max_n_time_masks (int): maximum number of time masking
max_n_time_masks (int): maximum number of time masking
replace_with_zero (bool): pad zero on mask if true else use mean
"""
"""
super
().
__init__
()
super
().
__init__
()
self
.
_rng
=
rng
self
.
_rng
=
rng
self
.
replace_with_zero
=
replace_with_zero
self
.
W
=
W
self
.
W
=
W
self
.
F
=
F
self
.
F
=
F
...
@@ -124,15 +127,18 @@ class SpecAugmentor(AugmentorBase):
...
@@ -124,15 +127,18 @@ class SpecAugmentor(AugmentorBase):
return
f
"specaug: F-
{
F
}
, T-
{
T
}
, F-n-
{
n_freq_masks
}
, T-n-
{
n_time_masks
}
"
return
f
"specaug: F-
{
F
}
, T-
{
T
}
, F-n-
{
n_freq_masks
}
, T-n-
{
n_time_masks
}
"
def
time_warp
(
xs
,
W
=
40
):
def
time_warp
(
xs
,
W
=
40
):
r
aise
NotImplementedError
r
eturn
xs
def
mask_freq
(
self
,
xs
,
replace_with_zero
=
False
):
def
mask_freq
(
self
,
xs
,
replace_with_zero
=
False
):
n_bins
=
xs
.
shape
[
0
]
n_bins
=
xs
.
shape
[
0
]
for
i
in
range
(
0
,
self
.
n_freq_masks
):
for
i
in
range
(
0
,
self
.
n_freq_masks
):
f
=
int
(
self
.
_rng
.
uniform
(
low
=
0
,
high
=
self
.
F
))
f
=
int
(
self
.
_rng
.
uniform
(
low
=
0
,
high
=
self
.
F
))
f_0
=
int
(
self
.
_rng
.
uniform
(
low
=
0
,
high
=
n_bins
-
f
))
f_0
=
int
(
self
.
_rng
.
uniform
(
low
=
0
,
high
=
n_bins
-
f
))
xs
[
f_0
:
f_0
+
f
,
:]
=
0
assert
f_0
<=
f_0
+
f
assert
f_0
<=
f_0
+
f
if
self
.
replace_with_zero
:
xs
[
f_0
:
f_0
+
f
,
:]
=
0
else
:
xs
[
f_0
:
f_0
+
f
,
:]
=
xs
.
mean
()
self
.
_freq_mask
=
(
f_0
,
f_0
+
f
)
self
.
_freq_mask
=
(
f_0
,
f_0
+
f
)
return
xs
return
xs
...
@@ -154,14 +160,17 @@ class SpecAugmentor(AugmentorBase):
...
@@ -154,14 +160,17 @@ class SpecAugmentor(AugmentorBase):
t
=
int
(
self
.
_rng
.
uniform
(
low
=
0
,
high
=
T
))
t
=
int
(
self
.
_rng
.
uniform
(
low
=
0
,
high
=
T
))
t
=
min
(
t
,
int
(
n_frames
*
self
.
p
))
t
=
min
(
t
,
int
(
n_frames
*
self
.
p
))
t_0
=
int
(
self
.
_rng
.
uniform
(
low
=
0
,
high
=
n_frames
-
t
))
t_0
=
int
(
self
.
_rng
.
uniform
(
low
=
0
,
high
=
n_frames
-
t
))
xs
[:,
t_0
:
t_0
+
t
]
=
0
assert
t_0
<=
t_0
+
t
assert
t_0
<=
t_0
+
t
if
self
.
replace_with_zero
:
xs
[:,
t_0
:
t_0
+
t
]
=
0
else
:
xs
[:,
t_0
:
t_0
+
t
]
=
xs
.
mean
()
self
.
_time_mask
=
(
t_0
,
t_0
+
t
)
self
.
_time_mask
=
(
t_0
,
t_0
+
t
)
return
xs
return
xs
def
__call__
(
self
,
x
,
train
=
True
):
def
__call__
(
self
,
x
,
train
=
True
):
if
not
train
:
if
not
train
:
return
return
x
return
self
.
transform_feature
(
x
)
return
self
.
transform_feature
(
x
)
def
transform_feature
(
self
,
xs
:
np
.
ndarray
):
def
transform_feature
(
self
,
xs
:
np
.
ndarray
):
...
@@ -171,7 +180,7 @@ class SpecAugmentor(AugmentorBase):
...
@@ -171,7 +180,7 @@ class SpecAugmentor(AugmentorBase):
Returns:
Returns:
xs (FloatTensor): `[F, T]`
xs (FloatTensor): `[F, T]`
"""
"""
#
xs = self.time_warp(xs)
xs
=
self
.
time_warp
(
xs
)
xs
=
self
.
mask_freq
(
xs
)
xs
=
self
.
mask_freq
(
xs
)
xs
=
self
.
mask_time
(
xs
)
xs
=
self
.
mask_time
(
xs
)
return
xs
return
xs
deepspeech/frontend/augmentor/speed_perturb.py
浏览文件 @
50f10f37
...
@@ -81,7 +81,7 @@ class SpeedPerturbAugmentor(AugmentorBase):
...
@@ -81,7 +81,7 @@ class SpeedPerturbAugmentor(AugmentorBase):
def
__call__
(
self
,
x
,
uttid
=
None
,
train
=
True
):
def
__call__
(
self
,
x
,
uttid
=
None
,
train
=
True
):
if
not
train
:
if
not
train
:
return
return
x
self
.
transform_audio
(
x
)
self
.
transform_audio
(
x
)
return
x
return
x
...
...
deepspeech/frontend/augmentor/volume_perturb.py
浏览文件 @
50f10f37
...
@@ -39,7 +39,7 @@ class VolumePerturbAugmentor(AugmentorBase):
...
@@ -39,7 +39,7 @@ class VolumePerturbAugmentor(AugmentorBase):
def
__call__
(
self
,
x
,
uttid
=
None
,
train
=
True
):
def
__call__
(
self
,
x
,
uttid
=
None
,
train
=
True
):
if
not
train
:
if
not
train
:
return
return
x
self
.
transform_audio
(
x
)
self
.
transform_audio
(
x
)
return
x
return
x
...
...
examples/aishell/s0/conf/augmentation.json
浏览文件 @
50f10f37
...
@@ -27,7 +27,8 @@
...
@@ -27,7 +27,8 @@
"W"
:
80
,
"W"
:
80
,
"adaptive_number_ratio"
:
0
,
"adaptive_number_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"max_n_time_masks"
:
20
"max_n_time_masks"
:
20
,
"replace_with_zero"
:
true
},
},
"prob"
:
1.0
"prob"
:
1.0
}
}
...
...
examples/aishell/s1/conf/augmentation.json
浏览文件 @
50f10f37
...
@@ -27,7 +27,8 @@
...
@@ -27,7 +27,8 @@
"W"
:
80
,
"W"
:
80
,
"adaptive_number_ratio"
:
0
,
"adaptive_number_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"max_n_time_masks"
:
20
"max_n_time_masks"
:
20
,
"replace_with_zero"
:
true
},
},
"prob"
:
1.0
"prob"
:
1.0
}
}
...
...
examples/aug_conf/augmentation.json
已删除
100644 → 0
浏览文件 @
86d08f99
[
{
"type"
:
"shift"
,
"params"
:
{
"min_shift_ms"
:
-5
,
"max_shift_ms"
:
5
},
"prob"
:
1.0
}
]
examples/aug
_conf/augmentation.example
.json
→
examples/aug
mentation/augmentation
.json
浏览文件 @
50f10f37
...
@@ -60,7 +60,8 @@
...
@@ -60,7 +60,8 @@
"W"
:
80
,
"W"
:
80
,
"adaptive_number_ratio"
:
0
,
"adaptive_number_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"max_n_time_masks"
:
20
"max_n_time_masks"
:
20
,
"replace_with_zero"
:
true
},
},
"prob"
:
0.0
"prob"
:
0.0
}
}
...
...
examples/callcenter/s1/conf/augmentation.json
浏览文件 @
50f10f37
...
@@ -27,7 +27,8 @@
...
@@ -27,7 +27,8 @@
"W"
:
80
,
"W"
:
80
,
"adaptive_number_ratio"
:
0
,
"adaptive_number_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"max_n_time_masks"
:
20
"max_n_time_masks"
:
20
,
"replace_with_zero"
:
true
},
},
"prob"
:
1.0
"prob"
:
1.0
}
}
...
...
examples/librispeech/s0/conf/augmentation.json
浏览文件 @
50f10f37
...
@@ -27,7 +27,8 @@
...
@@ -27,7 +27,8 @@
"W"
:
80
,
"W"
:
80
,
"adaptive_number_ratio"
:
0
,
"adaptive_number_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"max_n_time_masks"
:
20
"max_n_time_masks"
:
20
,
"replace_with_zero"
:
true
},
},
"prob"
:
1.0
"prob"
:
1.0
}
}
...
...
examples/librispeech/s1/conf/augmentation.json
浏览文件 @
50f10f37
...
@@ -27,7 +27,8 @@
...
@@ -27,7 +27,8 @@
"W"
:
80
,
"W"
:
80
,
"adaptive_number_ratio"
:
0
,
"adaptive_number_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"max_n_time_masks"
:
20
"max_n_time_masks"
:
20
,
"replace_with_zero"
:
true
},
},
"prob"
:
1.0
"prob"
:
1.0
}
}
...
...
examples/librispeech/s2/conf/augmentation.json
浏览文件 @
50f10f37
...
@@ -10,7 +10,8 @@
...
@@ -10,7 +10,8 @@
"W"
:
80
,
"W"
:
80
,
"adaptive_number_ratio"
:
0
,
"adaptive_number_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"max_n_time_masks"
:
20
"max_n_time_masks"
:
20
,
"replace_with_zero"
:
true
},
},
"prob"
:
1.0
"prob"
:
1.0
}
}
...
...
examples/timit/s1/conf/augmentation.json
浏览文件 @
50f10f37
...
@@ -27,7 +27,8 @@
...
@@ -27,7 +27,8 @@
"W"
:
80
,
"W"
:
80
,
"adaptive_number_ratio"
:
0
,
"adaptive_number_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"max_n_time_masks"
:
20
"max_n_time_masks"
:
20
,
"replace_with_zero"
:
true
},
},
"prob"
:
1.0
"prob"
:
1.0
}
}
...
...
examples/tiny/s0/conf/augmentation.json
浏览文件 @
50f10f37
[
[
{
"type"
:
"speed"
,
"params"
:
{
"min_speed_rate"
:
0.9
,
"max_speed_rate"
:
1.1
,
"num_rates"
:
3
},
"prob"
:
1.0
},
{
{
"type"
:
"shift"
,
"type"
:
"shift"
,
"params"
:
{
"params"
:
{
...
@@ -6,5 +15,21 @@
...
@@ -6,5 +15,21 @@
"max_shift_ms"
:
5
"max_shift_ms"
:
5
},
},
"prob"
:
1.0
"prob"
:
1.0
},
{
"type"
:
"specaug"
,
"params"
:
{
"F"
:
10
,
"T"
:
50
,
"n_freq_masks"
:
2
,
"n_time_masks"
:
2
,
"p"
:
1.0
,
"W"
:
80
,
"adaptive_number_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"max_n_time_masks"
:
20
,
"replace_with_zero"
:
true
},
"prob"
:
1.0
}
}
]
]
examples/tiny/s1/conf/augmentation.json
浏览文件 @
50f10f37
...
@@ -27,7 +27,8 @@
...
@@ -27,7 +27,8 @@
"W"
:
80
,
"W"
:
80
,
"adaptive_number_ratio"
:
0
,
"adaptive_number_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"adaptive_size_ratio"
:
0
,
"max_n_time_masks"
:
20
"max_n_time_masks"
:
20
,
"replace_with_zero"
:
true
},
},
"prob"
:
1.0
"prob"
:
1.0
}
}
...
...
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