wangmeng28
7 years ago
commit
13ec6f99fe
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,132 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#include "paddle/operators/sequence_slice_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
class SequenceSliceOp : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
void InferShape(framework::InferShapeContext* ctx) const override {
|
||||
PADDLE_ENFORCE(ctx->HasInput("X"),
|
||||
"Input(X) of SequenceSliceOp should not be null.");
|
||||
PADDLE_ENFORCE(ctx->HasInput("Offset"),
|
||||
"Input(Offset) of SequenceSliceOp should not be null.");
|
||||
PADDLE_ENFORCE(ctx->HasInput("Length"),
|
||||
"Input(Length) of SequenceSliceOp should not be null.");
|
||||
PADDLE_ENFORCE(ctx->HasOutput("Out"),
|
||||
"Output(Out) of SequenceSliceOp should not be null.");
|
||||
auto input_dims = ctx->GetInputDim("X");
|
||||
|
||||
auto offset_dim = ctx->GetInputDim("Offset");
|
||||
auto length_dim = ctx->GetInputDim("Length");
|
||||
|
||||
PADDLE_ENFORCE_EQ(
|
||||
offset_dim.size(), 2UL,
|
||||
"Only support one level sequence now, The rank of offset must be 2.");
|
||||
PADDLE_ENFORCE_EQ(
|
||||
length_dim.size(), 2UL,
|
||||
"Only support one level sequence now, The rank of Length must be 2.");
|
||||
|
||||
// Initialize the output's dims to maximum,
|
||||
// and re-set to real dims by the value of Offset and Length at kernel
|
||||
ctx->SetOutputDim("Out", input_dims);
|
||||
}
|
||||
|
||||
protected:
|
||||
framework::OpKernelType GetKernelType(
|
||||
const framework::ExecutionContext& ctx) const override {
|
||||
return framework::OpKernelType(
|
||||
framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
|
||||
ctx.device_context());
|
||||
}
|
||||
};
|
||||
|
||||
class SequenceSliceGradOp : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
void InferShape(framework::InferShapeContext* ctx) const override {
|
||||
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
|
||||
"The gradient of Out should not be null.");
|
||||
PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName("X")),
|
||||
"The gradient of X should not be null.");
|
||||
ctx->SetOutputsDim(framework::GradVarName("X"), ctx->GetInputsDim("X"));
|
||||
}
|
||||
|
||||
protected:
|
||||
framework::OpKernelType GetKernelType(
|
||||
const framework::ExecutionContext& ctx) const override {
|
||||
return framework::OpKernelType(
|
||||
framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
|
||||
ctx.device_context());
|
||||
}
|
||||
};
|
||||
|
||||
class SequenceSliceOpMaker : public framework::OpProtoAndCheckerMaker {
|
||||
public:
|
||||
SequenceSliceOpMaker(framework::OpProto* proto,
|
||||
framework::OpAttrChecker* op_checker)
|
||||
: OpProtoAndCheckerMaker(proto, op_checker) {
|
||||
AddInput("X",
|
||||
"(LoDTensor), "
|
||||
"the input of SequenceSliceOp.");
|
||||
AddInput("Offset",
|
||||
"(Tensor), "
|
||||
"a vector<int> to describe the offset of every input sequence for "
|
||||
"sub sequence item.");
|
||||
AddInput("Length",
|
||||
"(Tensor), "
|
||||
"a vector<int> to describe the length of every input sequence for "
|
||||
"sub sequence item.");
|
||||
AddOutput("Out",
|
||||
"(LoDTensor), the output of SequenceSliceOp.");
|
||||
AddComment(R"DOC(
|
||||
Sequence slice operator
|
||||
|
||||
The operator crops a subsequence from given sequence with given start offset and subsequence length.
|
||||
It only supports sequence (LoD Tensor with level number is 1).
|
||||
- Case:
|
||||
X = [[a1, a2;
|
||||
b1, b2;
|
||||
c1, c2]
|
||||
[d1, d2;
|
||||
e1, e2]]
|
||||
LoD(X) = {{0, 3, 5}}; Dims(X) = (5, 2)
|
||||
Offset = [[0], [1]]; Length = [[2], [1]]
|
||||
|
||||
Out = [[a1, a2;
|
||||
b1, b2]
|
||||
[e1, e2]]
|
||||
LoD(Out) = {{0, 2, 3}}; Dims(Out) = (3, 2)
|
||||
NOTE: The first dimension size of input, the size of offset and Length, should be equal. The offset start from 0.
|
||||
)DOC");
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
REGISTER_OP(sequence_slice, ops::SequenceSliceOp, ops::SequenceSliceOpMaker,
|
||||
sequence_slice_grad, ops::SequenceSliceGradOp);
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
sequence_slice,
|
||||
ops::SequenceSliceOpKernel<paddle::platform::CPUPlace, float>);
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
sequence_slice_grad,
|
||||
ops::SequenceSliceGradOpKernel<paddle::platform::CPUPlace, float>);
|
||||
@ -0,0 +1,23 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#include "paddle/operators/sequence_slice_op.h"
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
REGISTER_OP_GPU_KERNEL(
|
||||
sequence_slice,
|
||||
ops::SequenceSliceOpKernel<paddle::platform::GPUPlace, float>);
|
||||
REGISTER_OP_GPU_KERNEL(
|
||||
sequence_slice_grad,
|
||||
ops::SequenceSliceGradOpKernel<paddle::platform::GPUPlace, float>);
|
||||
@ -0,0 +1,173 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#pragma once
|
||||
#include "paddle/framework/op_registry.h"
|
||||
#include "paddle/operators/math/math_function.h"
|
||||
#include "paddle/operators/strided_memcpy.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
using Tensor = framework::Tensor;
|
||||
using LoDTensor = framework::LoDTensor;
|
||||
using LoD = framework::LoD;
|
||||
|
||||
template <typename T>
|
||||
inline LoD SequenceSliceLoD(const T& in, const int64_t* offset_data,
|
||||
const int64_t* length_data) {
|
||||
auto out_lod = in.lod();
|
||||
size_t lod_offset = 0;
|
||||
|
||||
auto n = in.lod()[0].size() - 1;
|
||||
out_lod[0][0] = 0;
|
||||
for (size_t i = 0; i < n; ++i) {
|
||||
lod_offset += length_data[i];
|
||||
out_lod[0][i+1] = lod_offset;
|
||||
}
|
||||
return out_lod;
|
||||
}
|
||||
|
||||
template <typename Place, typename T>
|
||||
class SequenceSliceOpKernel : public framework::OpKernel<T> {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& ctx) const override {
|
||||
auto* in = ctx.Input<LoDTensor>("X");
|
||||
auto* offset = ctx.Input<Tensor>("Offset");
|
||||
auto* length = ctx.Input<Tensor>("Length");
|
||||
auto* out = ctx.Output<LoDTensor>("Out");
|
||||
|
||||
auto lod = in->lod();
|
||||
auto n = lod[0].size() - 1;
|
||||
|
||||
PADDLE_ENFORCE_EQ(lod.size(), 1UL,
|
||||
"Only support one level sequence now.");
|
||||
PADDLE_ENFORCE_EQ(
|
||||
n, static_cast<size_t>(length->dims()[0]),
|
||||
"The size of input-sequence and length-array should be the same")
|
||||
PADDLE_ENFORCE_EQ(
|
||||
n, static_cast<size_t>(offset->dims()[0]),
|
||||
"The size of input-sequence and offset-array should be the same")
|
||||
|
||||
const int64_t* offset_data = offset->data<int64_t>();
|
||||
const int64_t* length_data = length->data<int64_t>();
|
||||
framework::Tensor offset_cpu;
|
||||
framework::Tensor length_cpu;
|
||||
|
||||
if (platform::is_gpu_place(ctx.GetPlace())) {
|
||||
offset_cpu.mutable_data<T>(offset->dims(), platform::CPUPlace());
|
||||
offset_cpu.CopyFrom(*offset, platform::CPUPlace(), ctx.device_context());
|
||||
offset_data = offset_cpu.data<int64_t>();
|
||||
|
||||
length_cpu.mutable_data<T>(length->dims(), platform::CPUPlace());
|
||||
length_cpu.CopyFrom(*length, platform::CPUPlace(), ctx.device_context());
|
||||
length_data = length_cpu.data<int64_t>();
|
||||
}
|
||||
|
||||
for (size_t i = 0; i < n; ++i) {
|
||||
PADDLE_ENFORCE_LT(0, offset_data[i],
|
||||
"The offset[%d] must greater than zero.", i)
|
||||
PADDLE_ENFORCE_LT(0, length_data[i],
|
||||
"The length[%d] must greater than zero.", i)
|
||||
PADDLE_ENFORCE_LT(
|
||||
lod[0][i] + offset_data[i] + length_data[i],
|
||||
lod[0][i + 1],
|
||||
"The target tensor's length overflow.")
|
||||
}
|
||||
|
||||
out->mutable_data<T>(ctx.GetPlace());
|
||||
auto out_lod = SequenceSliceLoD(*in, offset_data, length_data);
|
||||
auto out_dims = in->dims();
|
||||
out_dims[0] = out_lod[0][out_lod[0].size() - 1];
|
||||
out->Resize(out_dims);
|
||||
out->set_lod(out_lod);
|
||||
|
||||
auto in_stride = framework::stride(in->dims());
|
||||
auto out_stride = framework::stride(out->dims());
|
||||
|
||||
size_t out_offset = 0;
|
||||
for (size_t i = 0; i < n; ++i) {
|
||||
Tensor in_t =
|
||||
in->Slice(static_cast<int>(lod[0][i] + offset_data[i]),
|
||||
static_cast<int>(lod[0][i] + offset_data[i] +
|
||||
length_data[i]));
|
||||
|
||||
StridedMemcpy<T>(ctx.device_context(), in_t.data<T>(),
|
||||
in_stride, in_t.dims(), out_stride,
|
||||
out->data<T>() + out_offset);
|
||||
out_offset += length_data[i] * in_stride[0];
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
template <typename Place, typename T>
|
||||
class SequenceSliceGradOpKernel : public framework::OpKernel<T> {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& ctx) const override {
|
||||
auto* in = ctx.Input<LoDTensor>("X");
|
||||
auto* offset = ctx.Input<Tensor>("Offset");
|
||||
auto* length = ctx.Input<Tensor>("Length");
|
||||
auto* out_grad =
|
||||
ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"));
|
||||
auto* x_grad =
|
||||
ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
|
||||
|
||||
const int64_t* offset_data = offset->data<int64_t>();
|
||||
const int64_t* length_data = length->data<int64_t>();
|
||||
framework::Tensor offset_cpu;
|
||||
framework::Tensor length_cpu;
|
||||
|
||||
if (platform::is_gpu_place(ctx.GetPlace())) {
|
||||
offset_cpu.mutable_data<T>(offset->dims(), platform::CPUPlace());
|
||||
offset_cpu.CopyFrom(*offset, platform::CPUPlace(), ctx.device_context());
|
||||
offset_data = offset_cpu.data<int64_t>();
|
||||
|
||||
length_cpu.mutable_data<T>(length->dims(), platform::CPUPlace());
|
||||
length_cpu.CopyFrom(*length, platform::CPUPlace(), ctx.device_context());
|
||||
length_data = length_cpu.data<int64_t>();
|
||||
}
|
||||
|
||||
auto lod = in->lod();
|
||||
auto out_lod = out_grad->lod();
|
||||
|
||||
if (x_grad) {
|
||||
x_grad->mutable_data<T>(ctx.GetPlace());
|
||||
x_grad->set_lod(in->lod());
|
||||
math::SetConstant<Place, T> set_zero;
|
||||
set_zero(ctx.device_context(), x_grad, static_cast<T>(0));
|
||||
|
||||
auto out_grad_stride = framework::stride(out_grad->dims());
|
||||
|
||||
for (size_t i = 0; i < out_lod[0].size() - 1; ++i) {
|
||||
Tensor out_grad_t =
|
||||
out_grad->Slice(static_cast<int>(out_lod[0][i]),
|
||||
static_cast<int>(out_lod[0][i + 1]));
|
||||
auto out_grad_stride = framework::stride(out_grad_t.dims());
|
||||
|
||||
auto x_grad_stride = framework::stride(x_grad->dims());
|
||||
|
||||
Tensor x_grad_t = x_grad->Slice(
|
||||
static_cast<int>(lod[0][i] + offset_data[i]),
|
||||
static_cast<int>(lod[0][i] + offset_data[i] + length_data[i]));
|
||||
|
||||
StridedMemcpy<T>(ctx.device_context(), out_grad_t.data<T>(),
|
||||
out_grad_stride, out_grad_t.dims(), x_grad_stride,
|
||||
x_grad_t.data<T>());
|
||||
}
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue