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
Paddle
提交
508548f8
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 2 年 前同步成功
通知
2325
Star
20933
Fork
5424
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
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
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
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
“89b49ec0784a6a1fb99e28600f9d541f7067fed4”上不存在“tools/python/validate.py”
提交
508548f8
编写于
8月 22, 2018
作者:
T
tensor-tang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
implement attention lstm cpu forward
上级
9affc36c
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
278 addition
and
194 deletion
+278
-194
paddle/fluid/operators/attention_lstm_op.cc
paddle/fluid/operators/attention_lstm_op.cc
+276
-190
paddle/fluid/operators/attention_lstm_op.h
paddle/fluid/operators/attention_lstm_op.h
+2
-3
paddle/fluid/operators/fusion_lstm_op.h
paddle/fluid/operators/fusion_lstm_op.h
+0
-1
未找到文件。
paddle/fluid/operators/attention_lstm_op.cc
浏览文件 @
508548f8
...
@@ -20,10 +20,12 @@ limitations under the License. */
...
@@ -20,10 +20,12 @@ limitations under the License. */
#include "paddle/fluid/operators/math/lstm_compute.h"
#include "paddle/fluid/operators/math/lstm_compute.h"
#include "paddle/fluid/operators/math/sequence2batch.h"
#include "paddle/fluid/operators/math/sequence2batch.h"
#include "paddle/fluid/operators/math/cpu_vec.h"
namespace
paddle
{
namespace
paddle
{
namespace
operators
{
namespace
operators
{
void
Fus
ionLSTMOp
::
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
{
void
Attent
ionLSTMOp
::
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
"Input(X) of LSTM should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
"Input(X) of LSTM should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"WeightX"
),
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"WeightX"
),
"Input(WeightX) of LSTM should not be null."
);
"Input(WeightX) of LSTM should not be null."
);
...
@@ -57,6 +59,9 @@ void FusionLSTMOp::InferShape(framework::InferShapeContext* ctx) const {
...
@@ -57,6 +59,9 @@ void FusionLSTMOp::InferShape(framework::InferShapeContext* ctx) const {
"should be the same."
);
"should be the same."
);
}
}
// fc_out , shape (maxseqlen,1)
int
max_seq_len
=
0
;
auto
wx_dims
=
ctx
->
GetInputDim
(
"WeightX"
);
auto
wx_dims
=
ctx
->
GetInputDim
(
"WeightX"
);
PADDLE_ENFORCE_EQ
(
wx_dims
.
size
(),
2
,
PADDLE_ENFORCE_EQ
(
wx_dims
.
size
(),
2
,
"The rank of Input(WeightX) should be 2."
);
"The rank of Input(WeightX) should be 2."
);
...
@@ -103,241 +108,321 @@ void FusionLSTMOp::InferShape(framework::InferShapeContext* ctx) const {
...
@@ -103,241 +108,321 @@ void FusionLSTMOp::InferShape(framework::InferShapeContext* ctx) const {
ctx
->
ShareLoD
(
"X"
,
"XX"
);
ctx
->
ShareLoD
(
"X"
,
"XX"
);
}
}
framework
::
OpKernelType
Fus
ionLSTMOp
::
GetExpectedKernelType
(
framework
::
OpKernelType
Attent
ionLSTMOp
::
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
const
framework
::
ExecutionContext
&
ctx
)
const
{
return
framework
::
OpKernelType
(
return
framework
::
OpKernelType
(
framework
::
ToDataType
(
ctx
.
Input
<
framework
::
LoDTensor
>
(
"X"
)
->
type
()),
framework
::
ToDataType
(
ctx
.
Input
<
framework
::
LoDTensor
>
(
"X"
)
->
type
()),
ctx
.
device_context
());
ctx
.
device_context
());
}
}
void
Fus
ionLSTMOpMaker
::
Make
()
{
void
Attent
ionLSTMOpMaker
::
Make
()
{
AddInput
(
"X"
,
AddInput
(
"X"
,
"(LoDTensor) the input is a LodTensor, which support "
"(LoDTensor) the input is a LodTensor, which support "
"variable-time length input sequence. The underlying tensor in "
"variable-time length input sequence. The underlying tensor in "
"this LoDTensor is a matrix with shape (T X M), where T is the "
"this LoDTensor is a matrix with shape (T X M), where T is the "
"total time steps in this mini-batch, M is the dim size of x."
);
"total time steps in this mini-batch, M is the dim size of x."
);
AddInput
(
"WeightX"
,
AddInput
(
"C0"
,
"(Tensor) the learnable weights of X."
"(Tensor) LSTM C0"
" - The shape is (M x 4D), where M is the dim size of x, D is the "
"This is a tensor with shape (N x D), where N is the batch size, D "
"hidden size. "
"is the gate size."
" - Weight = {W_cx, W_ix, W_fx, W_ox}"
);
"C0 is necessary because of attention."
);
AddInput
(
"WeightH"
,
"(Tensor) same as LSTMOp, the learnable hidden-hidden weights."
" - The shape is (D x 4D), where D is the hidden size. "
" - Weight = {W_ch, W_ih, W_fh, W_oh}"
);
AddInput
(
"Bias"
,
"(Tensor) the learnable weights. Almost same as LSTMOp"
"Note: we should add the fc bias into this (1x4D) in bias."
"input-hidden bias weight and peephole connections weight if "
"setting `use_peepholes` True. "
"1. `use_peepholes = False` "
" - The shape is (1 x 4D). "
" - Bias = {b_c, b_i, b_f, b_o}."
"2. `use_peepholes = True` "
" - The shape is (1 x 7D). "
" - Bias = {b_c, b_i, b_f, b_o, W_ic, W_fc, W_oc}."
);
AddInput
(
"H0"
,
AddInput
(
"H0"
,
"(Tensor, optional) (same as LSTMOp) the initial hidden state is an "
"(Tensor, optional) LSTM H0"
"optional "
"This is a tensor with shape (N x D), where N is the "
"input. This is a tensor with shape (N x D), where N is the "
"batch size and D is the gate size."
)
"batch size and D is the hidden size."
)
.
AsDispensable
();
.
AsDispensable
();
AddInput
(
"C0"
,
AddInput
(
"AttentionWeight"
,
"(Tensor, optional) (same as LSTMOp) (the initial cell state is an "
"(Tensor) the weights of attention fc. Always relu the fc result."
"optional "
"The shape is ((M+D) x 1), where M is the dim size of x, D is the "
"input. This is a tensor with shape (N x D), where N is the "
"gate size of LSTM."
);
"batch size. `H0` and `C0` can be NULL but only at the same time."
)
AddInput
(
"AttentionBias, optional"
,
"(Tensor) the bias of attention fc."
"The shape is (1 x 1)"
)
.
AsDispensable
();
AddInput
(
"AttentionScalar"
,
"(Tensor, optional) the scalar on the result of attentioned fc. "
"Always relu the Scalar."
"The shape is (1 x 1)"
)
.
AsDispensable
();
AddInput
(
"AttentionScalarBias"
,
"(Tensor, optional) the scalar bias of attention fc."
"The shape is (1 x 1)"
)
.
AsDispensable
();
.
AsDispensable
();
AddInput
(
"LSTMWeight"
,
"(Tensor) the combined weight of LSTM"
" - The shape is ((D+M) x 4D), where D is the hidden gate size, M "
"is the dim size of x"
" - Weight = {W_forget, W_input, W_output, W_cell}"
);
AddInput
(
"LSTMBias"
,
"(Tensor) the combined bias of LSTM, shape (1x4D)."
"Note: we should add the bias of hidden and context accorindg to "
"the same gate: "
"{B_forget, B_input, B_output, B_cell}"
);
AddOutput
(
"Hidden"
,
AddOutput
(
"Hidden"
,
"(LoDTensor) (same as LSTMOp) the hidden state of LSTM operator. "
"(LoDTensor) (same as LSTMOp) the hidden state of LSTM operator. "
"The shape is (T x D), and lod is the same with the `Input`."
);
"The shape is (T x D), and lod is the same with the `Input`."
);
AddOutput
(
"Cell"
,
AddOutput
(
"Cell"
,
"(LoDTensor) (same as LSTMOp) the cell state of LSTM operator. "
"(LoDTensor) (same as LSTMOp) the cell state of LSTM operator. "
"The shape is (T x D), and lod is the same with the `Input`."
);
"The shape is (T x D), and lod is the same with the `Input`."
);
AddOutput
(
"XX"
,
AddOutput
(
"(LoDTensor) the result after X * WeightX (size is T x 4D)"
"AttentionedX"
,
" or batched_X (size is T x M), this will be automatically chosen
,"
"(LodTensor) shape is (T x 1), the result after X * AttentionWeight
,"
" where T is the total time steps in this mini-batch,"
" where T is the total time steps in this mini-batch,"
" D is the hidden size, M is the dim size of x input
."
)
" D is the hidden size
."
)
.
AsIntermediate
();
.
AsIntermediate
();
AddOutput
(
"
BatchedGate"
,
"(LoDTensor) (same as LSTMOp)."
).
AsIntermediate
();
AddOutput
(
"
AttentionFCOut"
,
AddOutput
(
"BatchCellPreAct"
,
"(LoDTensor) (same as LSTMOp)
."
)
"(Tensor) (max_seq_len, 1), compute at each step
."
)
.
AsIntermediate
();
.
AsIntermediate
();
AddAttr
<
bool
>
(
"use_peepholes"
,
AddOutput
(
"LSTMX"
,
"(bool, defalut: True) "
"(Tensor) the input X of LSTM for each step."
"whether to enable diagonal/peephole connections."
)
"Shape is (1 x M), where M is the x frame size"
)
.
SetDefault
(
true
);
.
AsIntermediate
();
AddAttr
<
bool
>
(
"is_reverse"
,
AddOutput
(
"(bool, defalut: False) "
"LSTMOUT"
,
"whether to compute reversed LSTM."
)
"(Tensor) the output of LSTM X(1*(D+M))* weight((D+M)*4D) for each step."
.
SetDefault
(
false
);
"Shape is (1 x 4D), where M is the x frame size"
)
.
AsIntermediate
();
// TODO(TJ): InEnum({"sigmoid", "tanh", "relu", "identity"});
AddAttr
<
std
::
string
>
(
"gate_activation"
,
AddAttr
<
std
::
string
>
(
"gate_activation"
,
"(string, default: sigmoid)"
"(string, default: sigmoid)"
"The activation for input gate, forget gate and output "
"The activation for input gate, forget gate and output "
"gate, `sigmoid` by default."
)
"gate, `sigmoid` by default."
)
.
SetDefault
(
"sigmoid"
)
.
SetDefault
(
"sigmoid"
)
.
InEnum
({
"sigmoid"
,
"tanh"
,
"relu"
,
"identity"
});
.
InEnum
({
"sigmoid"
});
AddAttr
<
std
::
string
>
(
"cell_activation"
,
AddAttr
<
std
::
string
>
(
"cell_activation"
,
"(string, default: tanh)"
"(string, default: tanh)"
"The activation for cell output, `tanh` by defalut."
)
"The activation for cell output, `tanh` by defalut."
)
.
SetDefault
(
"tanh"
)
.
SetDefault
(
"tanh"
)
.
InEnum
({
"
sigmoid"
,
"tanh"
,
"relu"
,
"identity
"
});
.
InEnum
({
"
tanh
"
});
AddAttr
<
std
::
string
>
(
"candidate_activation"
,
AddAttr
<
std
::
string
>
(
"candidate_activation"
,
"(string, default: tanh)"
"(string, default: tanh)"
"The activation for candidate hidden state, "
"The activation for candidate hidden state, "
"`tanh` by default."
)
"`tanh` by default."
)
.
SetDefault
(
"tanh"
)
.
SetDefault
(
"tanh"
)
.
InEnum
({
"
sigmoid"
,
"tanh"
,
"relu"
,
"identity
"
});
.
InEnum
({
"
tanh
"
});
AddComment
(
R"DOC(
AddComment
(
R"DOC(
Fusion Long-Short Term Memory (LSTM) Operator.
Attention Long-Short Term Memory (LSTM) Operator.
This operator fuse the X into LSTM, more details can refer to LSTM op.
Attention part:
concat( x(seqlen * M), expand( cell_t-1(1,D) ) ) => tmp(seqlen*(M+D))
tmp(seqlen*(M+D)) * fc((M+D)*1) => fcout(seqlen*1) with bias, relu
fcout(seqlen*1) * scalar => fcout(seqlen*1) with bias, relu
dotmul and sum pool ( fcout(seqlen*1), x(seqlen * M) ) => lstm_x_t(1, M)
LSTM part:
use lstm_x_t as input and compute as standard LSTM.
)DOC"
);
)DOC"
);
}
}
// y[i] = (x[i] + bias[0]) > 0 ? (x[i] + bias[0]) : 0;
template
<
typename
T
>
inline
void
bias_relu
(
const
int
n
,
const
T
*
x
,
const
T
*
bias
,
T
*
y
)
{
if
(
bias
)
{
for
(
int
i
=
0
;
i
<
n
;
++
i
)
{
y
[
i
]
=
x
[
i
]
+
bias
[
0
];
}
vec_relu
(
n
,
y
,
y
);
}
else
{
vec_relu
(
n
,
x
,
y
);
}
}
template
<
typename
DeviceContext
,
typename
T
>
template
<
typename
DeviceContext
,
typename
T
>
inline
void
ReorderInitState
(
const
DeviceContext
&
ctx
,
inline
void
vec_softmax
(
const
BlasT
<
DeviceContext
,
T
>&
blas
,
const
int
n
,
const
framework
::
Tensor
&
src
,
const
T
*
x
,
T
*
y
)
{
framework
::
Vector
<
size_t
>
index_lod
,
T
scalar
=
x
[
0
];
framework
::
Tensor
*
dst
,
bool
indexed_src
)
{
// max
math
::
CopyMatrixRowsFunctor
<
DeviceContext
,
T
>
row_shuffle
;
for
(
int
i
=
1
;
i
<
n
;
++
i
)
{
dst
->
mutable_data
<
T
>
(
src
.
dims
(),
ctx
.
GetPlace
());
scalar
=
scalar
<
x
[
i
]
?
x
[
i
]
:
scalar
;
// TODO(TJ): check mem copy perf
}
row_shuffle
(
ctx
,
src
,
index_lod
,
dst
,
indexed_src
);
// sub
for
(
int
i
=
0
;
i
<
n
;
++
i
)
{
y
[
c
]
=
x
[
c
]
-
alpha
;
}
// exp
blas
.
VEXP
(
n
,
y
,
y
);
// sum
scalar
=
T
(
0
);
for
(
int
i
=
0
;
i
<
n
;
++
i
)
{
scalar
+=
y
[
i
];
}
// scale
blas
.
VSCAL
(
n
,
static_cast
<
T
>
(
1
)
/
scalar
,
y
);
}
__m256
exp
(
__m256
a
)
{
return
exp256_ps
(
a
);
}
__m256
log
(
__m256
a
)
{
return
log256_ps
(
a
);
}
__m256
sin
(
__m256
a
)
{
return
sin256_ps
(
a
);
}
__m256
cos
(
__m256
a
)
{
return
cos256_ps
(
a
);
}
__m256
relu
(
const
__m256
a
)
{
__m256
tmp
=
_mm256_set1_ps
(
0.0
f
);
return
_mm256_max_ps
(
a
,
tmp
);
}
__m256
sigmoid
(
const
__m256
a
)
{
__m256
max
=
_mm256_set1_ps
(
SIGMOID_THRESHOLD_MAX
);
__m256
min
=
_mm256_set1_ps
(
SIGMOID_THRESHOLD_MIN
);
__m256
tmp
=
_mm256_max_ps
(
a
,
min
);
tmp
=
_mm256_min_ps
(
tmp
,
max
);
tmp
=
_mm256_sub_ps
(
_mm256_set1_ps
(
0.0
f
),
tmp
);
tmp
=
exp
(
tmp
);
tmp
=
_mm256_add_ps
(
_mm256_set1_ps
(
1.0
f
),
tmp
);
tmp
=
_mm256_div_ps
(
_mm256_set1_ps
(
1.0
f
),
tmp
);
return
tmp
;
}
__m256
tanh
(
const
__m256
a
)
{
__m256
max
=
_mm256_set1_ps
(
EXP_MAX_INPUT
);
__m256
tmp
=
_mm256_mul_ps
(
_mm256_set1_ps
(
-
2.0
f
),
a
);
tmp
=
_mm256_min_ps
(
tmp
,
max
);
tmp
=
exp
(
tmp
);
return
_mm256_sub_ps
(
_mm256_div_ps
(
_mm256_set1_ps
(
2.0
f
),
_mm256_add_ps
(
_mm256_set1_ps
(
1.0
f
),
tmp
)),
_mm256_set1_ps
(
1.0
f
));
}
__m256
linear
(
const
__m256
a
)
{
return
a
;
}
inline
void
vec_sigmoid
(
const
T
*
x
,
T
*
y
)
{
const
real
min
=
SIGMOID_THRESHOLD_MIN
;
const
real
max
=
SIGMOID_THRESHOLD_MAX
;
real
tmp
=
(
a
<
min
)
?
min
:
((
a
>
max
)
?
max
:
a
);
return
1.0
/
(
1.0
+
exp
(
-
tmp
));
}
}
template
<
typename
DeviceContext
,
typename
T
>
template
<
typename
DeviceContext
,
typename
T
>
class
Fuis
onLSTMKernel
:
public
framework
::
OpKernel
<
T
>
{
class
Attenti
onLSTMKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
x
=
ctx
.
Input
<
LoDTensor
>
(
"X"
);
auto
*
x
=
ctx
.
Input
<
LoDTensor
>
(
"X"
);
// T x M
auto
*
wx
=
ctx
.
Input
<
Tensor
>
(
"WeightX"
);
auto
*
h0
=
ctx
.
Input
<
Tensor
>
(
"H0"
);
// N x D
auto
*
wh
=
ctx
.
Input
<
Tensor
>
(
"WeightH"
);
auto
*
c0
=
ctx
.
Input
<
Tensor
>
(
"C0"
);
// N x D
auto
*
bias
=
ctx
.
Input
<
Tensor
>
(
"Bias"
);
auto
*
atten_w
=
ctx
.
Input
<
Tensor
>
(
"AttentionWeight"
);
// (M+D) x 1
auto
*
hidden_t0
=
ctx
.
Input
<
Tensor
>
(
"H0"
);
auto
*
atten_b
=
ctx
.
Input
<
Tensor
>
(
"AttentionBias"
);
// 1x1
auto
*
cell_t0
=
ctx
.
Input
<
Tensor
>
(
"C0"
);
auto
*
atten_scalar
=
ctx
.
Input
<
Tensor
>
(
"AttentionScalar"
);
// 1x1
auto
*
atten_scalar_bias
=
ctx
.
Input
<
Tensor
>
(
"AttentionScalar"
);
// 1x1
auto
*
xx
=
ctx
.
Output
<
LoDTensor
>
(
"XX"
);
auto
*
lstm_w
=
ctx
.
Input
<
Tensor
>
(
"LSTMWeight"
);
// (D+M) x D*4
auto
*
batched_gate
=
ctx
.
Output
<
LoDTensor
>
(
"BatchedGate"
);
auto
*
lstm_b
=
ctx
.
Input
<
Tensor
>
(
"LSTMBias"
);
// 1 x D*4
auto
*
hidden_out
=
ctx
.
Output
<
LoDTensor
>
(
"Hidden"
);
auto
*
cell_out
=
ctx
.
Output
<
LoDTensor
>
(
"Cell"
);
auto
*
hidden_out
=
ctx
.
Output
<
LoDTensor
>
(
"Hidden"
);
// TxD
bool
is_reverse
=
ctx
.
Attr
<
bool
>
(
"is_reverse"
);
auto
*
cell_out
=
ctx
.
Output
<
LoDTensor
>
(
"Cell"
);
// TxD
auto
*
atted_x
=
ctx
.
Output
<
LoDTensor
>
(
"AttentionedX"
);
// T x 1
T
*
xx_data
=
xx
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
*
fc_out
=
ctx
.
Output
<
Tensor
>
(
'
AttentionFCOut
'
);
// max_seq_len x 1
T
*
batched_gate_data
=
batched_gate
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
*
lstm_x
=
ctx
.
Output
<
Tensor
>
(
"LSTMX"
);
// 1 x M
hidden_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
*
lstm_out
=
ctx
.
Output
<
Tensor
>
(
"LSTMOUT"
);
// 1 x 4D
cell_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
const
T
*
x_data
=
x
->
data
<
T
>
();
const
T
*
x_data
=
x
->
data
<
T
>
();
const
T
*
wx_data
=
wx
->
data
<
T
>
();
const
T
*
h0_data
=
h0
->
data
<
T
>
();
auto
x_dims
=
x
->
dims
();
const
T
*
c0_data
=
c0
->
data
<
T
>
();
auto
wx_dims
=
wx
->
dims
();
const
T
*
lstm_w_data
=
lstm_w
->
data
<
T
>
();
const
T
*
lstm_b_data
=
lstm_b
->
data
<
T
>
();
math
::
LoDTensor2BatchFunctor
<
DeviceContext
,
T
>
to_batch
;
const
T
*
atten_w_data
=
atten_w
->
data
<
T
>
();
auto
&
dev_ctx
=
ctx
.
template
device_context
<
DeviceContext
>();
const
T
*
atten_b_data
=
atten_b
?
atten_b
->
data
<
T
>
()
:
NULL
;
auto
blas
=
math
::
GetBlas
<
DeviceContext
,
T
>
(
dev_ctx
);
const
T
*
atten_scalar_data
=
atten_scalar
?
atten_scalar
->
data
<
T
>
()
:
NULL
;
if
(
x_dims
[
1
]
>
wx_dims
[
1
])
{
const
T
*
atten_scalar_bias_data
=
math
::
FCCompute
<
DeviceContext
,
T
>
(
blas
,
x_dims
[
0
],
wx_dims
[
1
],
x_dims
[
1
],
atten_scalar_bias
?
atten_scalar_bias
->
data
<
T
>
()
:
NULL
;
x_data
,
wx_data
,
xx_data
,
bias
->
data
<
T
>
());
T
*
hidden_out_data
=
hidden_out
->
mutable_data
<
T
>
();
to_batch
(
dev_ctx
,
*
xx
,
batched_gate
,
true
,
is_reverse
);
T
*
cell_out_data
=
cell_out
->
mutable_data
<
T
>
();
}
else
{
T
*
atted_x_data
=
atted_x
->
mutable_data
<
T
>
();
to_batch
(
dev_ctx
,
*
x
,
xx
,
true
,
is_reverse
);
T
*
fc_out_data
=
fc_out
->
mutable_data
<
T
>
();
batched_gate
->
set_lod
(
xx
->
lod
());
T
*
lstm_x_data
=
lstm_x
->
mutable_data
<
T
>
();
math
::
FCCompute
<
DeviceContext
,
T
>
(
blas
,
x_dims
[
0
],
wx_dims
[
1
],
x_dims
[
1
],
T
*
lstm_out_data
=
lstm_out
->
mutable_data
<
T
>
();
xx_data
,
wx_data
,
batched_gate_data
,
bias
->
data
<
T
>
());
auto
x_lod
=
x
->
lod
();
}
auto
x_dims
=
x
->
dims
();
// T x M
auto
w_dims
=
w
->
dims
();
// (D+M) x 4D
int
frame_size
=
static_cast
<
int
>
(
wx_dims
[
1
]
/
4
);
const
int
M
=
x_dims
[
1
];
// x frame size
framework
::
DDim
out_dims
({
x_dims
[
0
],
frame_size
});
const
int
D
=
w_dims
[
1
]
/
4
;
// gate frame size
math
::
LstmMetaValue
<
T
>
lstm_value
;
const
int
D2
=
D
*
2
;
// no peephole
const
int
D3
=
D
*
3
;
lstm_value
.
check_ig
=
nullptr
;
const
int
D4
=
w_dims
[
1
];
lstm_value
.
check_fg
=
nullptr
;
const
int
batch_size
=
x_lod
[
0
].
size
()
-
1
;
// assert lod.size() == 1
lstm_value
.
check_og
=
nullptr
;
lstm_value
.
prev_state_value
=
nullptr
;
// x(TxM) * fc (Mx1) part of atten_wgt(M+D)x1
Tensor
ordered_c0
;
auto
blas
=
math
::
GetBlas
<
DeviceContext
,
T
>
(
ctx
);
math
::
FCCompute
<
DeviceContext
,
T
>
(
blas
,
T
,
1
,
M
,
x_data
,
atten_w_data
,
framework
::
Vector
<
size_t
>
order
(
batched_gate
->
lod
()[
2
]);
atted_x_data
,
atten_b_data
);
if
(
cell_t0
)
{
const
T
*
cur_x_data
=
x_data
;
// Since the batch computing for LSTM reorders the input sequence
const
T
*
prev_cell_data
=
NULL
;
// according to their length. The initialized cell state also needs
const
T
*
prev_hidden_data
=
NULL
;
// to reorder.
T
*
cur_cell_out_data
=
cell_out_data
;
ReorderInitState
<
DeviceContext
,
T
>
(
dev_ctx
,
*
cell_t0
,
order
,
&
ordered_c0
,
T
*
cur_hidden_out_data
=
hidden_out_data
;
true
);
for
(
int
i
=
0
;
i
<
batch_size
;
++
i
)
{
lstm_value
.
prev_state_value
=
ordered_c0
.
data
<
T
>
();
int
seq_len
=
x_lod
[
0
][
i
+
1
];
}
prev_cell_data
=
c0_data
+
i
*
D
;
prev_hidden_data
=
h0
?
h0_data
+
i
*
D
:
NULL
;
// Use the local variable as here.
LoDTensor
batch_hidden
,
batch_cell
;
for
(
int
step
=
0
;
step
<
seq_len
;
++
step
)
{
auto
*
batch_cell_pre_act
=
ctx
.
Output
<
LoDTensor
>
(
"BatchCellPreAct"
);
/// compute attention vector
batch_hidden
.
mutable_data
<
T
>
(
out_dims
,
ctx
.
GetPlace
());
// prev_cell(1xD) * fc(D) rest part of atten_wgt
batch_cell
.
mutable_data
<
T
>
(
out_dims
,
ctx
.
GetPlace
());
// T = cblas_dot();
batch_cell_pre_act
->
mutable_data
<
T
>
(
out_dims
,
ctx
.
GetPlace
());
T
prev_cell_bias
=
blas
.
VDOT
(
D
,
prev_cell_data
,
atten_w_data
+
M
);
// add cell bias and relu
auto
batch_starts
=
batched_gate
->
lod
()[
0
];
bias_relu
<
T
>
(
seq_len
,
atted_x_data
,
&
prev_cell_bias
,
fc_out_data
);
size_t
max_seq_len
=
batch_starts
.
size
()
-
1
;
// fc2: scalar
auto
gate_act
=
math
::
detail
::
GetActivationType
(
if
(
atten_scalar_data
)
{
ctx
.
Attr
<
std
::
string
>
(
"gate_activation"
));
// x = a*x
auto
cell_act
=
math
::
detail
::
GetActivationType
(
blas
.
VSCAL
(
seq_len
,
atten_scalar_data
,
fc_out_data
);
ctx
.
Attr
<
std
::
string
>
(
"cell_activation"
));
bias_relu
<
T
>
(
seq_len
,
fc_out_data
,
atten_scalar_bias_data
,
auto
cand_act
=
math
::
detail
::
GetActivationType
(
fc_out_data
);
ctx
.
Attr
<
std
::
string
>
(
"candidate_activation"
));
}
vec_softmax
<
DeviceContext
,
T
>
(
blas
,
seq_len
,
fc_out_data
,
fc_out_data
);
for
(
size_t
n
=
0
;
n
<
max_seq_len
;
n
++
)
{
// mul x(seq_len*M) and sum pool
int
bstart
=
static_cast
<
int
>
(
batch_starts
[
n
]);
math
::
FCCompute
<
DeviceContext
,
T
>
(
blas
,
1
,
M
,
seq_len
,
fc_out_data
,
int
bend
=
static_cast
<
int
>
(
batch_starts
[
n
+
1
]);
cur_x_data
,
lstm_x_data
);
Tensor
gate_t
=
batched_gate
->
Slice
(
bstart
,
bend
);
/// compute LSTM step
Tensor
out_t
=
batch_hidden
.
Slice
(
bstart
,
bend
);
// lstm weight : concat[forget , input , output , tilde]
Tensor
cell_t
=
batch_cell
.
Slice
(
bstart
,
bend
);
// shape : (D + M) x (4 * D)
Tensor
cell_pre_act_t
=
batch_cell_pre_act
->
Slice
(
bstart
,
bend
);
// fc inputX(1xM) * weightX(M*(4D)) => 1 x 4D
blas
.
MatMul
(
1
,
D4
,
M
,
lstm_x_data
,
lstm_w_data
+
D
*
D4
,
lstm_out_data
);
int
cur_batch_size
=
bend
-
bstart
;
if
(
prev_hidden_data
)
{
blas
.
GEMM
(
CblasNoTrans
,
CblasNoTrans
,
1
,
D4
,
D
,
static_cast
<
T
>
(
1
),
if
(
n
>
0
)
{
prev_hidden_data
,
D
,
lstm_w_data
,
D4
,
static_cast
<
T
>
(
1
),
int
pre_h_start
=
static_cast
<
int
>
(
batch_starts
[
n
-
1
]);
lstm_out_data
,
D4
);
int
pre_h_end
=
pre_h_start
+
cur_batch_size
;
}
auto
pre_hidden_t
=
batch_hidden
.
Slice
(
pre_h_start
,
pre_h_end
);
// since input is 1xM, so can use add bias
// TODO(TJ): use gemm directly
blas
.
VADD
(
D4
,
lstm_b_data
,
lstm_out_data
,
lstm_out_data
);
blas
.
MatMul
(
pre_hidden_t
,
false
,
*
wh
,
false
,
static_cast
<
T
>
(
1.0
),
&
gate_t
,
static_cast
<
T
>
(
1.0
));
// gate act: sigmoid
}
else
if
(
hidden_t0
)
{
vec_sigmoid
(
D3
,
lstm_out_data
,
lstm_out_data
);
// TODO(TJ): move h0 outside for
// candicate act: tanh
// If n == 0 and there is no initialized hidden state, that is to say
vec_tanh
(
D
,
lstm_out_data
+
D3
,
lstm_out_data
+
D3
);
// the H0 is zeros, the calculation W_h * H0 will be skiped.
// If n == 0 and there is initialized hidden state, calculate W_h * H0.
// a = forget * prev_cell
blas
.
VMUL
(
D
,
lstm_out_data
,
prev_cell_data
,
lstm_out_data
);
// Since the batch computing for LSTM reorders the input sequence
// according to their length. The initialized hidden state also needs
// b = input * tilde
// to reorder.
blas
.
VMUL
(
D
,
lstm_out_data
+
D
,
lstm_out
+
D3
,
lstm_out_data
+
D
);
Tensor
ordered_h0
;
ReorderInitState
<
DeviceContext
,
T
>
(
dev_ctx
,
*
hidden_t0
,
order
,
// cell_out = a + b
&
ordered_h0
,
true
);
blas
.
VADD
(
D
,
lstm_out_data
,
lstm_out_data
+
D
,
cur_cell_out_data
);
// TODO(TJ): use gemm directly
blas
.
MatMul
(
ordered_h0
,
false
,
*
wh
,
false
,
static_cast
<
T
>
(
1.0
),
&
gate_t
,
// state act tanh(cell_out) * output_gate
static_cast
<
T
>
(
1.0
));
vec_tanh
(
D
,
cur_cell_out_data
,
lstm_out_data
);
blas
.
VMUL
(
D
,
lstm_out_data
,
lstm_out
+
D2
,
cur_hidden_out_data
);
prev_hidden_data
=
hidden_out
+
i
*
gate_size
;
prev_cell_data
=
cur_cell_out_data
;
cur_cell_out_data
=
cur_cell_out_data
+
D
;
cur_hidden_out_data
=
cur_hidden_out_data
+
D
;
}
}
cur_x_data
=
cur_x_data
+
seq_len
*
M
;
lstm_value
.
gate_value
=
gate_t
.
data
<
T
>
();
lstm_value
.
output_value
=
out_t
.
data
<
T
>
();
lstm_value
.
state_value
=
cell_t
.
data
<
T
>
();
lstm_value
.
state_active_value
=
cell_pre_act_t
.
data
<
T
>
();
math
::
LstmUnitFunctor
<
DeviceContext
,
T
>::
compute
(
dev_ctx
,
lstm_value
,
frame_size
,
cur_batch_size
,
gate_act
,
cell_act
,
cand_act
);
lstm_value
.
prev_state_value
=
lstm_value
.
state_value
;
}
}
math
::
Batch2LoDTensorFunctor
<
DeviceContext
,
T
>
to_seq
;
batch_hidden
.
set_lod
(
batched_gate
->
lod
());
// restore the output hidden in LoDTensor from the batch hidden
to_seq
(
dev_ctx
,
batch_hidden
,
hidden_out
);
batch_cell
.
set_lod
(
batched_gate
->
lod
());
// restore the output cell state in LoDTensor from the batch cell
to_seq
(
dev_ctx
,
batch_cell
,
cell_out
);
}
}
};
};
...
@@ -345,10 +430,11 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
...
@@ -345,10 +430,11 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
}
// namespace paddle
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
namespace
ops
=
paddle
::
operators
;
REGISTER_OPERATOR
(
fusion_lstm
,
ops
::
FusionLSTMOp
,
ops
::
FusionLSTMOpMaker
,
REGISTER_OPERATOR
(
attention_lstm
,
ops
::
AttentionLSTMOp
,
ops
::
AttentionLSTMOpMaker
,
paddle
::
framework
::
DefaultGradOpDescMaker
<
true
>
);
paddle
::
framework
::
DefaultGradOpDescMaker
<
true
>
);
REGISTER_OP_CPU_KERNEL
(
REGISTER_OP_CPU_KERNEL
(
fus
ion_lstm
,
attent
ion_lstm
,
ops
::
Fuis
onLSTMKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
,
ops
::
Attenti
onLSTMKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
,
ops
::
Fuis
onLSTMKernel
<
paddle
::
platform
::
CPUDeviceContext
,
double
>
);
ops
::
Attenti
onLSTMKernel
<
paddle
::
platform
::
CPUDeviceContext
,
double
>
);
paddle/fluid/operators/attention_lstm_op.h
浏览文件 @
508548f8
...
@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
...
@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
limitations under the License. */
limitations under the License. */
#pragma once
#pragma once
// #include <string>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/op_registry.h"
namespace
paddle
{
namespace
paddle
{
...
@@ -22,7 +21,7 @@ namespace operators {
...
@@ -22,7 +21,7 @@ namespace operators {
using
LoDTensor
=
framework
::
LoDTensor
;
using
LoDTensor
=
framework
::
LoDTensor
;
using
Tensor
=
framework
::
Tensor
;
using
Tensor
=
framework
::
Tensor
;
class
Fus
ionLSTMOp
:
public
framework
::
OperatorWithKernel
{
class
Attent
ionLSTMOp
:
public
framework
::
OperatorWithKernel
{
public:
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
...
@@ -33,7 +32,7 @@ class FusionLSTMOp : public framework::OperatorWithKernel {
...
@@ -33,7 +32,7 @@ class FusionLSTMOp : public framework::OperatorWithKernel {
const
framework
::
ExecutionContext
&
ctx
)
const
override
;
const
framework
::
ExecutionContext
&
ctx
)
const
override
;
};
};
class
Fus
ionLSTMOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
class
Attent
ionLSTMOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
public:
void
Make
()
override
;
void
Make
()
override
;
};
};
...
...
paddle/fluid/operators/fusion_lstm_op.h
浏览文件 @
508548f8
...
@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
...
@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
limitations under the License. */
limitations under the License. */
#pragma once
#pragma once
// #include <string>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/op_registry.h"
namespace
paddle
{
namespace
paddle
{
...
...
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