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
_土豆不是豆
Paddle
提交
33cc0f7a
P
Paddle
项目概览
_土豆不是豆
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
列表
看板
标记
里程碑
合并请求
0
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle
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
0
Issue
0
列表
看板
标记
里程碑
合并请求
0
合并请求
0
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
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
33cc0f7a
编写于
7月 26, 2022
作者:
Z
Zhong Hui
提交者:
GitHub
7月 26, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[Eager] Add warpctc yaml (#44617)
上级
b6e84806
变更
11
显示空白变更内容
内联
并排
Showing
11 changed file
with
73 addition
and
26 deletion
+73
-26
paddle/phi/api/yaml/legacy_api.yaml
paddle/phi/api/yaml/legacy_api.yaml
+12
-0
paddle/phi/api/yaml/legacy_backward.yaml
paddle/phi/api/yaml/legacy_backward.yaml
+12
-0
paddle/phi/infermeta/multiary.cc
paddle/phi/infermeta/multiary.cc
+1
-1
paddle/phi/infermeta/multiary.h
paddle/phi/infermeta/multiary.h
+1
-1
paddle/phi/kernels/impl/warpctc_grad_kernel_impl.h
paddle/phi/kernels/impl/warpctc_grad_kernel_impl.h
+9
-9
paddle/phi/kernels/impl/warpctc_kernel_impl.h
paddle/phi/kernels/impl/warpctc_kernel_impl.h
+6
-6
paddle/phi/kernels/warpctc_grad_kernel.h
paddle/phi/kernels/warpctc_grad_kernel.h
+2
-2
paddle/phi/kernels/warpctc_kernel.h
paddle/phi/kernels/warpctc_kernel.h
+2
-2
paddle/phi/ops/compat/warpctc_sig.cc
paddle/phi/ops/compat/warpctc_sig.cc
+2
-2
python/paddle/fluid/layers/loss.py
python/paddle/fluid/layers/loss.py
+9
-0
python/paddle/fluid/tests/unittests/test_warpctc_op.py
python/paddle/fluid/tests/unittests/test_warpctc_op.py
+17
-3
未找到文件。
paddle/phi/api/yaml/legacy_api.yaml
浏览文件 @
33cc0f7a
...
...
@@ -2403,6 +2403,18 @@
func
:
viterbi_decode
data_type
:
input
-
api
:
warpctc
args
:
(Tensor logits, Tensor label, Tensor logits_length, Tensor labels_length, int blank, bool norm_by_times)
output
:
Tensor(loss), Tensor(warpctcgrad)
infer_meta
:
func
:
WarpctcInferMeta
kernel
:
func
:
warpctc
data_type
:
logits
optional
:
logits_length, labels_length
intermediate
:
warpctcgrad
backward
:
warpctc_grad
-
api
:
where
args
:
(Tensor condition, Tensor x, Tensor y)
output
:
Tensor
...
...
paddle/phi/api/yaml/legacy_backward.yaml
浏览文件 @
33cc0f7a
...
...
@@ -2376,6 +2376,18 @@
inplace
:
(out_grad -> x_grad)
backward
:
unsqueeze_double_grad
-
backward_api
:
warpctc_grad
forward
:
warpctc (Tensor logits, Tensor label, Tensor logits_length, Tensor labels_length, int blank, bool norm_by_times) -> Tensor(loss), Tensor(warpctcgrad)
args
:
(Tensor logits, Tensor logits_length, Tensor warpctcgrad, Tensor loss_grad, int blank, bool norm_by_times)
output
:
Tensor(logits_grad)
infer_meta
:
func
:
UnchangedInferMeta
param
:
[
logits
]
kernel
:
func
:
warpctc_grad
optional
:
logits_length
no_need_buffer
:
logits
-
backward_api
:
where_grad
forward
:
where (Tensor condition, Tensor x, Tensor y) -> Tensor(out)
args
:
(Tensor condition, Tensor x, Tensor y, Tensor out_grad)
...
...
paddle/phi/infermeta/multiary.cc
浏览文件 @
33cc0f7a
...
...
@@ -2049,7 +2049,7 @@ void WarpctcInferMeta(const MetaTensor& logits,
const
MetaTensor
&
labels_length
,
int
blank
,
bool
norm_by_times
,
MetaTensor
*
warpctc
_
grad
,
MetaTensor
*
warpctcgrad
,
MetaTensor
*
loss
)
{
auto
logits_dims
=
logits
.
dims
();
int
sequence_width
=
0
;
...
...
paddle/phi/infermeta/multiary.h
浏览文件 @
33cc0f7a
...
...
@@ -358,7 +358,7 @@ void WarpctcInferMeta(const MetaTensor& logits,
const
MetaTensor
&
labels_length
,
int
blank
,
bool
norm_by_times
,
MetaTensor
*
warpctc
_
grad
,
MetaTensor
*
warpctcgrad
,
MetaTensor
*
loss
);
void
WhereInferMeta
(
const
MetaTensor
&
condition
,
...
...
paddle/phi/kernels/impl/warpctc_grad_kernel_impl.h
浏览文件 @
33cc0f7a
...
...
@@ -29,33 +29,33 @@ namespace phi {
template
<
typename
T
,
typename
Context
>
void
WarpctcGradKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
warpctc_grad
,
const
DenseTensor
&
logits
,
const
DenseTensor
&
loss_grad
,
const
paddle
::
optional
<
DenseTensor
>&
logits_length
,
const
DenseTensor
&
warpctcgrad
,
const
DenseTensor
&
loss_grad
,
int
blank
,
bool
norm_by_times
,
DenseTensor
*
logits_grad
)
{
dev_ctx
.
template
Alloc
<
T
>(
logits_grad
);
if
(
logits_length
.
is_initialized
())
{
int
max_seq_length
=
warpctc
_
grad
.
dims
()[
0
];
// Tmax
int
num_sequences
=
warpctc
_
grad
.
dims
()[
1
];
// B
int
seq_width
=
warpctc
_
grad
.
dims
()[
2
];
// D
int
max_seq_length
=
warpctcgrad
.
dims
()[
0
];
// Tmax
int
num_sequences
=
warpctcgrad
.
dims
()[
1
];
// B
int
seq_width
=
warpctcgrad
.
dims
()[
2
];
// D
// B
auto
logits_len_e
=
EigenTensor
<
int64_t
,
1
>::
From
(
*
logits_length
);
// (B, 1)
auto
loss_grad_e
=
EigenTensor
<
T
,
2
>::
From
(
loss_grad
);
// (T, B, D)
auto
warpctc
_grad_e
=
EigenTensor
<
T
,
3
>::
From
(
warpctc_
grad
);
auto
warpctc
grad_e
=
EigenTensor
<
T
,
3
>::
From
(
warpctc
grad
);
auto
logits_grad_e
=
EigenTensor
<
T
,
3
>::
From
(
*
logits_grad
);
Eigen
::
DSizes
<
int
,
3
>
grad_shape
(
1
,
num_sequences
,
1
);
Eigen
::
DSizes
<
int
,
3
>
bcast
(
max_seq_length
,
1
,
seq_width
);
auto
logits_g
=
warpctc_grad_e
*
loss_grad_e
.
reshape
(
grad_shape
).
broadcast
(
bcast
).
eval
();
auto
logits_g
=
warpctcgrad_e
*
loss_grad_e
.
reshape
(
grad_shape
).
broadcast
(
bcast
).
eval
();
auto
*
place
=
dev_ctx
.
eigen_device
();
if
(
norm_by_times
)
{
...
...
@@ -71,7 +71,7 @@ void WarpctcGradKernel(const Context& dev_ctx,
}
else
{
paddle
::
operators
::
math
::
UnpaddingLoDTensorFunctor
<
Context
,
T
>
()(
dev_ctx
,
warpctc
_
grad
,
warpctcgrad
,
logits_grad
,
-
1
,
0
,
...
...
paddle/phi/kernels/impl/warpctc_kernel_impl.h
浏览文件 @
33cc0f7a
...
...
@@ -233,8 +233,8 @@ void WarpctcKernel(const Context& dev_ctx,
const
paddle
::
optional
<
DenseTensor
>&
labels_length
,
int
blank
,
bool
norm_by_times
,
DenseTensor
*
warpctc_grad
,
DenseTensor
*
loss
)
{
DenseTensor
*
loss
,
DenseTensor
*
warpctcgrad
)
{
size_t
num_sequences
,
sequence_width
,
max_sequence_length
;
paddle
::
framework
::
Vector
<
size_t
>
logits_lod
;
paddle
::
framework
::
Vector
<
size_t
>
label_lod
;
...
...
@@ -383,11 +383,11 @@ void WarpctcKernel(const Context& dev_ctx,
// warpctc computes loss and gradient in one call, gradient data also stored
// in batch format
warpctc
_
grad
->
Resize
(
warpctc_logits
.
dims
());
T
*
warpctc
_grad_data
=
dev_ctx
.
template
Alloc
<
T
>(
warpctc_
grad
);
warpctcgrad
->
Resize
(
warpctc_logits
.
dims
());
T
*
warpctc
grad_data
=
dev_ctx
.
template
Alloc
<
T
>(
warpctc
grad
);
phi
::
funcs
::
SetConstant
<
Context
,
T
>
()(
dev_ctx
,
warpctc
_
grad
,
static_cast
<
T
>
(
0
));
dev_ctx
,
warpctcgrad
,
static_cast
<
T
>
(
0
));
// warpctc accesses labels in CPU memory
DenseTensor
warpctc_label
;
...
...
@@ -439,7 +439,7 @@ void WarpctcKernel(const Context& dev_ctx,
T
*
warpctc_loss_data
=
dev_ctx
.
template
HostAlloc
<
T
>(
&
warpctc_loss
);
WarpCTCFunctor
<
Context
,
T
>
()(
dev_ctx
,
warpctc_logits_data
,
warpctc
_
grad_data
,
warpctcgrad_data
,
warpctc_label_data
,
warpctc_label_lengths
.
data
(),
warpctc_logits_lengths
.
data
(),
...
...
paddle/phi/kernels/warpctc_grad_kernel.h
浏览文件 @
33cc0f7a
...
...
@@ -21,10 +21,10 @@ namespace phi {
template
<
typename
T
,
typename
Context
>
void
WarpctcGradKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
warpctc_grad
,
const
DenseTensor
&
logits
,
const
DenseTensor
&
loss_grad
,
const
paddle
::
optional
<
DenseTensor
>&
logits_length
,
const
DenseTensor
&
warpctcgrad
,
const
DenseTensor
&
loss_grad
,
int
blank
,
bool
norm_by_times
,
DenseTensor
*
logits_grad
);
...
...
paddle/phi/kernels/warpctc_kernel.h
浏览文件 @
33cc0f7a
...
...
@@ -27,7 +27,7 @@ void WarpctcKernel(const Context& dev_ctx,
const
paddle
::
optional
<
DenseTensor
>&
labels_length
,
int
blank
,
bool
norm_by_times
,
DenseTensor
*
warpctc_grad
,
DenseTensor
*
loss
);
DenseTensor
*
loss
,
DenseTensor
*
warpctcgrad
);
}
// namespace phi
paddle/phi/ops/compat/warpctc_sig.cc
浏览文件 @
33cc0f7a
...
...
@@ -20,13 +20,13 @@ KernelSignature WarpctcOpArgumentMapping(const ArgumentMappingContext& ctx) {
return
KernelSignature
(
"warpctc"
,
{
"Logits"
,
"Label"
,
"LogitsLength"
,
"LabelLength"
},
{
"blank"
,
"norm_by_times"
},
{
"
WarpCTCGrad"
,
"Loss
"
});
{
"
Loss"
,
"WarpCTCGrad
"
});
}
KernelSignature
WarpctcGradOpArgumentMapping
(
const
ArgumentMappingContext
&
ctx
)
{
return
KernelSignature
(
"warpctc_grad"
,
{
"
WarpCTCGrad"
,
"Logits"
,
"Loss@GRAD"
,
"LogitsLength
"
},
{
"
Logits"
,
"LogitsLength"
,
"WarpCTCGrad"
,
"Loss@GRAD
"
},
{
"blank"
,
"norm_by_times"
},
{
"Logits@GRAD"
});
}
...
...
python/paddle/fluid/layers/loss.py
浏览文件 @
33cc0f7a
...
...
@@ -546,6 +546,15 @@ def warpctc(input,
fetch_list=[cost.name])
print(output)
"""
if
in_dygraph_mode
():
if
input_length
is
None
or
label_length
is
None
:
raise
ValueError
(
"input_length and label_length must not be None in dygraph mode!"
)
loss_out
=
_C_ops
.
final_state_warpctc
(
input
,
label
,
input_length
,
label_length
,
blank
,
norm_by_times
)
return
loss_out
if
_non_static_mode
():
if
input_length
is
None
or
label_length
is
None
:
raise
ValueError
(
...
...
python/paddle/fluid/tests/unittests/test_warpctc_op.py
浏览文件 @
33cc0f7a
...
...
@@ -191,6 +191,16 @@ class CTCForward(object):
return
self
.
loss
def
python_api
(
logits
,
label
,
logits_length
=
None
,
labels_length
=
None
,
blank
=
0
,
norm_by_times
=
False
):
return
paddle
.
fluid
.
layers
.
warpctc
(
logits
,
label
,
blank
,
norm_by_times
,
logits_length
,
labels_length
)
class
TestWarpCTCOp
(
OpTest
):
def
config
(
self
):
...
...
@@ -280,6 +290,8 @@ class TestWarpCTCOpWithPadding(OpTest):
def
setUp
(
self
):
self
.
op_type
=
"warpctc"
self
.
python_api
=
python_api
self
.
python_out_sig
=
[
"Loss"
]
self
.
config
()
logits
=
np
.
random
.
uniform
(
...
...
@@ -344,7 +356,7 @@ class TestWarpCTCOpWithPadding(OpTest):
}
def
test_check_output
(
self
):
self
.
check_output
()
self
.
check_output
(
check_eager
=
True
)
def
test_check_grad
(
self
):
self
.
outputs
[
'WarpCTCGrad'
]
=
self
.
gradient
...
...
@@ -387,6 +399,8 @@ class TestWarpCTCOpFp64(OpTest):
def
setUp
(
self
):
self
.
op_type
=
"warpctc"
self
.
python_api
=
python_api
self
.
python_out_sig
=
[
"Loss"
]
self
.
config
()
logits
=
np
.
random
.
uniform
(
...
...
@@ -451,11 +465,11 @@ class TestWarpCTCOpFp64(OpTest):
}
def
test_check_output
(
self
):
self
.
check_output
()
self
.
check_output
(
check_eager
=
True
)
def
test_check_grad
(
self
):
self
.
outputs
[
'WarpCTCGrad'
]
=
self
.
gradient
self
.
check_grad
([
"Logits"
],
"Loss"
)
self
.
check_grad
([
"Logits"
],
"Loss"
,
check_eager
=
True
)
class
TestWarpCTCOpError
(
unittest
.
TestCase
):
...
...
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