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
PaddleDetection
提交
ac3370a4
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
大约 1 年 前同步成功
通知
694
Star
11112
Fork
2696
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
184
列表
看板
标记
里程碑
合并请求
40
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
PaddleDetection
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
184
Issue
184
列表
看板
标记
里程碑
合并请求
40
合并请求
40
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 >>
提交
ac3370a4
编写于
10月 26, 2017
作者:
D
dangqingqing
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Add unit testing for gemv and fix the gradien check for bais.
上级
2e029874
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
165 addition
and
42 deletion
+165
-42
paddle/framework/lod_tensor_test.cu
paddle/framework/lod_tensor_test.cu
+4
-4
paddle/operators/lstm_op.h
paddle/operators/lstm_op.h
+5
-2
paddle/operators/math/math_function_test.cc
paddle/operators/math/math_function_test.cc
+50
-0
paddle/operators/math/math_function_test.cu
paddle/operators/math/math_function_test.cu
+62
-0
python/paddle/v2/framework/tests/test_lstm_op.py
python/paddle/v2/framework/tests/test_lstm_op.py
+44
-36
未找到文件。
paddle/framework/lod_tensor_test.cu
浏览文件 @
ac3370a4
...
@@ -36,8 +36,8 @@ TEST(LoDTensor, LoDInGPU) {
...
@@ -36,8 +36,8 @@ TEST(LoDTensor, LoDInGPU) {
lod_tensor
.
mutable_data
<
float
>
(
place
);
lod_tensor
.
mutable_data
<
float
>
(
place
);
lod_tensor
.
set_lod
(
src_lod
);
lod_tensor
.
set_lod
(
src_lod
);
CHECK
_EQ
(
lod_tensor
.
lod_element
(
0
,
2
).
first
,
4UL
);
EXPECT
_EQ
(
lod_tensor
.
lod_element
(
0
,
2
).
first
,
4UL
);
CHECK
_EQ
(
lod_tensor
.
lod_element
(
0
,
4
).
first
,
8UL
);
EXPECT
_EQ
(
lod_tensor
.
lod_element
(
0
,
4
).
first
,
8UL
);
auto
lod
=
lod_tensor
.
lod
();
auto
lod
=
lod_tensor
.
lod
();
...
@@ -45,6 +45,6 @@ TEST(LoDTensor, LoDInGPU) {
...
@@ -45,6 +45,6 @@ TEST(LoDTensor, LoDInGPU) {
cudaDeviceSynchronize
();
cudaDeviceSynchronize
();
for
(
size_t
i
=
0
;
i
<
src_lod
[
0
].
size
();
++
i
)
{
for
(
size_t
i
=
0
;
i
<
src_lod
[
0
].
size
();
++
i
)
{
CHECK
_EQ
(
lod
[
0
].
data
()[
i
],
src_lod
[
0
].
data
()[
i
]
*
2
);
EXPECT
_EQ
(
lod
[
0
].
data
()[
i
],
src_lod
[
0
].
data
()[
i
]
*
2
);
}
}
}
}
\ No newline at end of file
paddle/operators/lstm_op.h
浏览文件 @
ac3370a4
...
@@ -162,9 +162,9 @@ class LSTMGradKernel : public framework::OpKernel<T> {
...
@@ -162,9 +162,9 @@ class LSTMGradKernel : public framework::OpKernel<T> {
auto
*
bias_g
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Bias"
));
auto
*
bias_g
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Bias"
));
auto
&
device_ctx
=
ctx
.
device_context
();
auto
&
device_ctx
=
ctx
.
device_context
();
math
::
SetConstant
<
Place
,
T
>
zero
;
if
(
weight_g
)
{
if
(
weight_g
)
{
weight_g
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
weight_g
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
math
::
SetConstant
<
Place
,
T
>
zero
;
zero
(
device_ctx
,
weight_g
,
static_cast
<
T
>
(
0.0
));
zero
(
device_ctx
,
weight_g
,
static_cast
<
T
>
(
0.0
));
}
}
...
@@ -188,6 +188,7 @@ class LSTMGradKernel : public framework::OpKernel<T> {
...
@@ -188,6 +188,7 @@ class LSTMGradKernel : public framework::OpKernel<T> {
math
::
LstmMetaGrad
<
T
>
lstm_grad
;
math
::
LstmMetaGrad
<
T
>
lstm_grad
;
if
(
bias
&&
bias_g
)
{
if
(
bias
&&
bias_g
)
{
T
*
bias_g_data
=
const_cast
<
T
*>
(
bias_g
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()));
T
*
bias_g_data
=
const_cast
<
T
*>
(
bias_g
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()));
zero
(
device_ctx
,
bias_g
,
static_cast
<
T
>
(
0.0
));
lstm_grad
.
checkIgGrad
=
bias_g_data
+
4
*
frame_size
;
lstm_grad
.
checkIgGrad
=
bias_g_data
+
4
*
frame_size
;
lstm_grad
.
checkFgGrad
=
lstm_grad
.
checkIgGrad
+
frame_size
;
lstm_grad
.
checkFgGrad
=
lstm_grad
.
checkIgGrad
+
frame_size
;
lstm_grad
.
checkOgGrad
=
lstm_grad
.
checkFgGrad
+
frame_size
;
lstm_grad
.
checkOgGrad
=
lstm_grad
.
checkFgGrad
+
frame_size
;
...
@@ -219,6 +220,8 @@ class LSTMGradKernel : public framework::OpKernel<T> {
...
@@ -219,6 +220,8 @@ class LSTMGradKernel : public framework::OpKernel<T> {
batch_cell_g
.
mutable_data
<
T
>
(
out_dims
,
ctx
.
GetPlace
());
batch_cell_g
.
mutable_data
<
T
>
(
out_dims
,
ctx
.
GetPlace
());
batch_cell_g
.
set_lod
(
batch_gate
->
lod
());
batch_cell_g
.
set_lod
(
batch_gate
->
lod
());
to_batch
(
device_ctx
,
*
cell_g
,
batch_cell_g
,
false
);
to_batch
(
device_ctx
,
*
cell_g
,
batch_cell_g
,
false
);
// TODO(qingqing) support the case output cell has gradient.
zero
(
device_ctx
,
&
batch_cell_g
,
static_cast
<
T
>
(
0.0
));
LoDTensor
batch_gate_g
;
LoDTensor
batch_gate_g
;
batch_gate_g
.
mutable_data
<
T
>
(
batch_gate
->
dims
(),
ctx
.
GetPlace
());
batch_gate_g
.
mutable_data
<
T
>
(
batch_gate
->
dims
(),
ctx
.
GetPlace
());
...
@@ -304,7 +307,7 @@ class LSTMGradKernel : public framework::OpKernel<T> {
...
@@ -304,7 +307,7 @@ class LSTMGradKernel : public framework::OpKernel<T> {
int
n
=
static_cast
<
int
>
(
batch_gate_g
.
dims
()[
1
]);
int
n
=
static_cast
<
int
>
(
batch_gate_g
.
dims
()[
1
]);
Tensor
ones
;
Tensor
ones
;
ones
.
mutable_data
<
T
>
({
1
,
m
},
ctx
.
GetPlace
());
ones
.
mutable_data
<
T
>
({
m
},
ctx
.
GetPlace
());
math
::
SetConstant
<
Place
,
T
>
set
;
math
::
SetConstant
<
Place
,
T
>
set
;
set
(
device_ctx
,
&
ones
,
static_cast
<
T
>
(
1.0
));
set
(
device_ctx
,
&
ones
,
static_cast
<
T
>
(
1.0
));
...
...
paddle/operators/math/math_function_test.cc
浏览文件 @
ac3370a4
...
@@ -89,3 +89,53 @@ TEST(math_function, zero) {
...
@@ -89,3 +89,53 @@ TEST(math_function, zero) {
EXPECT_EQ
(
t
[
2
],
1
);
EXPECT_EQ
(
t
[
2
],
1
);
EXPECT_EQ
(
t
[
3
],
1
);
EXPECT_EQ
(
t
[
3
],
1
);
}
}
template
<
typename
T
>
void
GemvTest
(
int
m
,
int
n
,
bool
trans
)
{
paddle
::
framework
::
Tensor
mat_a
;
paddle
::
framework
::
Tensor
vec_b
;
paddle
::
framework
::
Tensor
vec_c
;
auto
*
cpu_place
=
new
paddle
::
platform
::
CPUPlace
();
int
b_num
=
trans
?
m
:
n
;
int
c_num
=
trans
?
n
:
m
;
T
*
data_a
=
mat_a
.
mutable_data
<
T
>
({
m
,
n
},
*
cpu_place
);
T
*
data_b
=
vec_b
.
mutable_data
<
T
>
({
b_num
},
*
cpu_place
);
T
*
data_c
=
vec_c
.
mutable_data
<
T
>
({
c_num
},
*
cpu_place
);
for
(
int
i
=
0
;
i
<
mat_a
.
numel
();
++
i
)
{
data_a
[
i
]
=
static_cast
<
T
>
(
i
);
}
for
(
int
i
=
0
;
i
<
vec_b
.
numel
();
++
i
)
{
data_b
[
i
]
=
static_cast
<
T
>
(
i
);
}
paddle
::
platform
::
CPUDeviceContext
context
(
*
cpu_place
);
paddle
::
operators
::
math
::
gemv
<
paddle
::
platform
::
CPUPlace
,
T
>
(
context
,
trans
,
static_cast
<
int
>
(
m
),
static_cast
<
int
>
(
n
),
1.
,
data_a
,
data_b
,
0.
,
data_c
);
if
(
!
trans
)
{
for
(
int
i
=
0
;
i
<
m
;
++
i
)
{
T
sum
=
0.0
;
for
(
int
j
=
0
;
j
<
n
;
++
j
)
{
sum
+=
data_a
[
i
*
n
+
j
]
*
data_b
[
j
];
}
ASSERT_FLOAT_EQ
(
data_c
[
i
],
sum
);
}
}
else
{
for
(
int
i
=
0
;
i
<
n
;
++
i
)
{
T
sum
=
0.0
;
for
(
int
j
=
0
;
j
<
m
;
++
j
)
{
sum
+=
data_a
[
j
*
n
+
i
]
*
data_b
[
j
];
}
ASSERT_FLOAT_EQ
(
data_c
[
i
],
sum
);
}
}
}
TEST
(
math_function
,
gemv
)
{
GemvTest
<
float
>
(
3
,
13
,
false
);
GemvTest
<
double
>
(
4
,
5
,
false
);
GemvTest
<
float
>
(
12
,
7
,
true
);
GemvTest
<
double
>
(
7
,
9
,
true
);
}
paddle/operators/math/math_function_test.cu
浏览文件 @
ac3370a4
...
@@ -177,3 +177,65 @@ TEST(math_function, gemm_trans_cublas) {
...
@@ -177,3 +177,65 @@ TEST(math_function, gemm_trans_cublas) {
EXPECT_EQ
(
input3_ptr
[
7
],
99
);
EXPECT_EQ
(
input3_ptr
[
7
],
99
);
delete
gpu_place
;
delete
gpu_place
;
}
}
template
<
typename
T
>
void
GemvTest
(
int
m
,
int
n
,
bool
trans
)
{
paddle
::
framework
::
Tensor
mat_a
;
paddle
::
framework
::
Tensor
vec_b
;
paddle
::
framework
::
Tensor
vec_c
;
auto
*
cpu_place
=
new
paddle
::
platform
::
CPUPlace
();
T
*
data_a
=
mat_a
.
mutable_data
<
T
>
({
m
,
n
},
*
cpu_place
);
T
*
data_b
=
vec_b
.
mutable_data
<
T
>
({
trans
?
m
:
n
},
*
cpu_place
);
T
*
data_c
=
vec_c
.
mutable_data
<
T
>
({
trans
?
n
:
m
},
*
cpu_place
);
auto
*
gpu_place
=
new
paddle
::
platform
::
GPUPlace
(
0
);
paddle
::
framework
::
Tensor
g_mat_a
;
paddle
::
framework
::
Tensor
g_vec_b
;
paddle
::
framework
::
Tensor
g_vec_c
;
T
*
g_data_a
=
g_mat_a
.
mutable_data
<
T
>
(
mat_a
.
dims
(),
*
gpu_place
);
T
*
g_data_b
=
g_vec_b
.
mutable_data
<
T
>
(
vec_b
.
dims
(),
*
gpu_place
);
T
*
g_data_c
=
g_vec_c
.
mutable_data
<
T
>
(
vec_c
.
dims
(),
*
gpu_place
);
for
(
int
i
=
0
;
i
<
mat_a
.
numel
();
++
i
)
{
data_a
[
i
]
=
static_cast
<
T
>
(
i
);
}
for
(
int
i
=
0
;
i
<
vec_b
.
numel
();
++
i
)
{
data_b
[
i
]
=
static_cast
<
T
>
(
i
);
}
paddle
::
platform
::
CUDADeviceContext
context
(
*
gpu_place
);
g_mat_a
.
CopyFrom
(
mat_a
,
*
gpu_place
,
context
);
g_vec_b
.
CopyFrom
(
vec_b
,
*
gpu_place
,
context
);
paddle
::
operators
::
math
::
gemv
<
paddle
::
platform
::
GPUPlace
,
T
>
(
context
,
trans
,
static_cast
<
int
>
(
m
),
static_cast
<
int
>
(
n
),
1.
,
g_data_a
,
g_data_b
,
0.
,
g_data_c
);
vec_c
.
CopyFrom
(
g_vec_c
,
paddle
::
platform
::
CPUPlace
(),
context
);
if
(
!
trans
)
{
for
(
int
i
=
0
;
i
<
m
;
++
i
)
{
T
sum
=
0.0
;
for
(
int
j
=
0
;
j
<
n
;
++
j
)
{
sum
+=
data_a
[
i
*
n
+
j
]
*
data_b
[
j
];
}
ASSERT_FLOAT_EQ
(
data_c
[
i
],
sum
);
}
}
else
{
for
(
int
i
=
0
;
i
<
n
;
++
i
)
{
T
sum
=
0.0
;
for
(
int
j
=
0
;
j
<
m
;
++
j
)
{
sum
+=
data_a
[
j
*
n
+
i
]
*
data_b
[
j
];
}
ASSERT_FLOAT_EQ
(
data_c
[
i
],
sum
);
}
}
}
TEST
(
math_function
,
gemv
)
{
GemvTest
<
float
>
(
3
,
13
,
false
);
GemvTest
<
double
>
(
3
,
13
,
false
);
GemvTest
<
float
>
(
3
,
13
,
true
);
GemvTest
<
double
>
(
3
,
13
,
true
);
}
python/paddle/v2/framework/tests/test_lstm_op.py
浏览文件 @
ac3370a4
...
@@ -114,26 +114,20 @@ def lstm(
...
@@ -114,26 +114,20 @@ def lstm(
class
TestLstmOp
(
OpTest
):
class
TestLstmOp
(
OpTest
):
def
set_data
(
self
):
def
set_argument
(
self
):
# self.lod = [[0, 2, 6, 9]]
self
.
lod
=
[[
0
,
2
,
6
,
9
]]
# self.D = 64
self
.
D
=
16
# self.sort_idx = [2, 6, 0, 3, 7, 1, 4, 8, 5]
self
.
sort_idx
=
[
2
,
6
,
0
,
3
,
7
,
1
,
4
,
8
,
5
]
self
.
lod
=
[[
0
,
1
]]
self
.
D
=
4
self
.
sort_idx
=
[
0
]
# self.act_gate = 'identity'
# self.act_cell = 'identity'
# self.act_cand = 'identity'
self
.
act_gate
=
'sigmoid'
self
.
act_gate
=
'sigmoid'
self
.
act_cell
=
'tanh'
self
.
act_cell
=
'tanh'
self
.
act_cand
=
'tanh'
self
.
act_cand
=
'tanh'
self
.
has_initial_state
=
True
self
.
is_reverse
=
False
self
.
is_reverse
=
False
def
setUp
(
self
):
def
setUp
(
self
):
self
.
set_
data
()
self
.
set_
argument
()
self
.
op_type
=
'lstm'
self
.
op_type
=
'lstm'
T
=
self
.
lod
[
0
][
-
1
]
T
=
self
.
lod
[
0
][
-
1
]
...
@@ -155,17 +149,14 @@ class TestLstmOp(OpTest):
...
@@ -155,17 +149,14 @@ class TestLstmOp(OpTest):
for
i
,
j
in
enumerate
(
self
.
sort_idx
):
for
i
,
j
in
enumerate
(
self
.
sort_idx
):
g_sort
[
i
,
:]
=
g
[
j
,
:]
g_sort
[
i
,
:]
=
g
[
j
,
:]
self
.
inputs
=
{
self
.
inputs
=
{
'Input'
:
(
x
,
self
.
lod
),
'Weight'
:
w
,
'Bias'
:
b
}
'Input'
:
(
x
,
self
.
lod
),
self
.
inputs
[
'H0'
]
=
h0
'H0'
:
h0
,
self
.
inputs
[
'C0'
]
=
c0
'C0'
:
c0
,
'Weight'
:
w
,
'Bias'
:
b
}
self
.
outputs
=
{
self
.
outputs
=
{
'Hidden'
:
(
h
,
self
.
lod
),
'Hidden'
:
(
h
,
self
.
lod
),
'Cell'
:
(
c
,
self
.
lod
),
'Cell'
:
(
c
,
self
.
lod
),
#
'BatchGate': g_sort,
'BatchGate'
:
g_sort
,
}
}
self
.
attrs
=
{
self
.
attrs
=
{
'usePeepholes'
:
True
,
'usePeepholes'
:
True
,
...
@@ -175,26 +166,43 @@ class TestLstmOp(OpTest):
...
@@ -175,26 +166,43 @@ class TestLstmOp(OpTest):
'candidateActivation'
:
self
.
act_cand
'candidateActivation'
:
self
.
act_cand
}
}
def
not_
test_check_output
(
self
):
def
test_check_output
(
self
):
self
.
check_output
()
self
.
check_output
()
#TODO(qingqing) add more unit testing case
def
test_check_grad
(
self
):
def
test_check_grad
(
self
):
# TODO(qingqing) remove folowing two lines after the check_grad is refined.
self
.
outputs
[
'BatchGate'
]
=
None
self
.
outputs
[
'BatchGate'
]
=
None
self
.
outputs
[
'BatchCellPreAct'
]
=
None
self
.
outputs
[
'BatchCellPreAct'
]
=
None
self
.
check_grad
([
'Input'
,
'Weight'
],
[
'Hidden'
,
'Cell'
])
self
.
check_grad
([
'Input'
,
'Weight'
,
'Bias'
],
[
'Hidden'
])
#['Input', 'Weight', 'Bias'], ['Hidden', 'Cell'])
#class TestLstmOpRerverse(TestLstmOp):
class
TestLstmOpHasNoInitial
(
TestLstmOp
):
# def set_data(self):
def
set_argument
(
self
):
# self.lod = [[0, 2, 6, 9]]
self
.
lod
=
[[
0
,
2
,
6
,
9
]]
# self.D = 64
self
.
D
=
64
# self.sort_idx = [2, 6, 0, 3, 7, 1, 4, 8, 5]
self
.
sort_idx
=
[
2
,
6
,
0
,
3
,
7
,
1
,
4
,
8
,
5
]
#
# self.act_gate = 'sigmoid'
self
.
act_gate
=
'sigmoid'
# self.act_cell = 'tanh'
self
.
act_cell
=
'tanh'
# self.act_cand = 'tanh'
self
.
act_cand
=
'tanh'
#
# self.is_reverse = True
self
.
has_initial_state
=
False
self
.
is_reverse
=
True
class
TestLstmOpRerverse
(
TestLstmOp
):
def
set_argument
(
self
):
self
.
lod
=
[[
0
,
2
,
6
,
9
]]
self
.
D
=
64
self
.
sort_idx
=
[
2
,
6
,
0
,
3
,
7
,
1
,
4
,
8
,
5
]
self
.
act_gate
=
'sigmoid'
self
.
act_cell
=
'tanh'
self
.
act_cand
=
'tanh'
self
.
has_initial_state
=
True
self
.
is_reverse
=
True
if
__name__
==
'__main__'
:
if
__name__
==
'__main__'
:
...
...
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