wangmeng28
7 years ago
commit
28c98103de
@ -0,0 +1,124 @@
|
||||
/* 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/margin_rank_loss_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
class MarginRankLossOp : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
protected:
|
||||
void InferShape(framework::InferShapeContext *ctx) const override {
|
||||
// input check
|
||||
PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) shouldn't be null.");
|
||||
PADDLE_ENFORCE(ctx->HasInput("X1"), "Input(X1) shouldn't be null.");
|
||||
PADDLE_ENFORCE(ctx->HasInput("X2"), "Input(X2) shouldn't be null.");
|
||||
PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) shouldn't be null.");
|
||||
auto label_dims = ctx->GetInputDim("Label");
|
||||
auto x1_dims = ctx->GetInputDim("X1");
|
||||
auto x2_dims = ctx->GetInputDim("X2");
|
||||
PADDLE_ENFORCE(
|
||||
(label_dims == x1_dims) && (x1_dims == x2_dims) &&
|
||||
(label_dims.size() == 2) && (label_dims[1] == 1),
|
||||
"All inputs must be 2-D tensor with shape [batch_size x 1].");
|
||||
ctx->SetOutputDim("Activated", label_dims);
|
||||
ctx->SetOutputDim("Out", label_dims);
|
||||
}
|
||||
};
|
||||
|
||||
template <typename T>
|
||||
class MarginRankLossOpMaker : public framework::OpProtoAndCheckerMaker {
|
||||
public:
|
||||
MarginRankLossOpMaker(framework::OpProto *proto,
|
||||
framework::OpAttrChecker *op_checker)
|
||||
: OpProtoAndCheckerMaker(proto, op_checker) {
|
||||
AddInput("X1",
|
||||
"(2-D tensor with shape [batch_size x 1]) The score for "
|
||||
"one item X1 to be ranked, from pairwise ranking model.");
|
||||
AddInput("X2",
|
||||
"(2-D tensor with shape [batch_size x 1]) The score for "
|
||||
"another item X2 to be ranked, from pairwise ranking model.");
|
||||
AddInput("Label",
|
||||
"(2-D tensor with shape [batch_size x 1]) "
|
||||
"The label indicating X1 ranked higher than X2 or not, "
|
||||
"can only be +1 or -1.");
|
||||
AddAttr<T>("margin", "(scalar, default 0) Margin for MarginRankLossOp.")
|
||||
.SetDefault(static_cast<T>(0));
|
||||
AddOutput("Activated",
|
||||
"(2-D tensor with shape [batch_size x 1]) Intermediate tensor "
|
||||
"to indicate whether each element of Output(Out) is activated.")
|
||||
.AsIntermediate();
|
||||
AddOutput("Out",
|
||||
"(2-D tensor with shape [batch_size x 1]) "
|
||||
"The output loss of MarginRankLoss operator.");
|
||||
AddComment(R"DOC(
|
||||
|
||||
MarginRankLoss operator measures the loss given a pair of training sample
|
||||
{`X1`, `X2`} and the `Label` with attribute `margin`, where `Label = +1`
|
||||
indicating X1 is ranked higher than `X2`, otherwise `Label = -1`. The loss
|
||||
turns out
|
||||
|
||||
loss(X1, X2, Label) = max(0, -Label * (X1 - X2) + margin).
|
||||
|
||||
The attribute `margin` involved here helps make the predictions more robust.
|
||||
Denote the item ranked higher as the positive sample, otherwise the negative
|
||||
sample. If the score of the two samples satisfies
|
||||
|
||||
positive sample - negative sample < margin,
|
||||
|
||||
the pair of samples will contribute to the final loss, which will backpropogate
|
||||
and train the ranking model to enlarge the difference of the two score.
|
||||
|
||||
For batch input with size `batch_size`, `X1`, `X2` and `Label`
|
||||
all have the same shape [batch_size x 1].
|
||||
|
||||
)DOC");
|
||||
}
|
||||
};
|
||||
|
||||
class MarginRankLossGradOp : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
protected:
|
||||
void InferShape(framework::InferShapeContext *ctx) const override {
|
||||
PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) shouldn't be null.");
|
||||
PADDLE_ENFORCE(ctx->HasInput("X1"), "Input(X1) shouldn't be null.");
|
||||
PADDLE_ENFORCE(ctx->HasInput("X2"), "Input(X2) shouldn't be null.");
|
||||
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
|
||||
"Input(Out@GRAD) shouldn't be null.");
|
||||
PADDLE_ENFORCE(ctx->HasInput("Activated"),
|
||||
"Intermediate(Activated) shouldn't be null.");
|
||||
auto dims = ctx->GetInputDim("Label");
|
||||
ctx->SetOutputDim(framework::GradVarName("X1"), dims);
|
||||
ctx->SetOutputDim(framework::GradVarName("X2"), dims);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
namespace ops = paddle::operators;
|
||||
|
||||
REGISTER_OP(margin_rank_loss, ops::MarginRankLossOp,
|
||||
ops::MarginRankLossOpMaker<float>, margin_rank_loss_grad,
|
||||
ops::MarginRankLossGradOp);
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
margin_rank_loss,
|
||||
ops::MarginRankLossKernel<paddle::platform::CPUPlace, float>);
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
margin_rank_loss_grad,
|
||||
ops::MarginRankLossGradKernel<paddle::platform::CPUPlace, float>);
|
||||
@ -0,0 +1,24 @@
|
||||
/* 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/margin_rank_loss_op.h"
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
|
||||
REGISTER_OP_GPU_KERNEL(
|
||||
margin_rank_loss,
|
||||
ops::MarginRankLossKernel<paddle::platform::GPUPlace, float>);
|
||||
REGISTER_OP_GPU_KERNEL(
|
||||
margin_rank_loss_grad,
|
||||
ops::MarginRankLossGradKernel<paddle::platform::GPUPlace, float>);
|
||||
@ -0,0 +1,98 @@
|
||||
/* 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/eigen.h"
|
||||
#include "paddle/framework/op_registry.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
template <typename T>
|
||||
struct ReLU {
|
||||
HOSTDEVICE T operator()(const T& val) const {
|
||||
return val > 0 ? val : static_cast<T>(0);
|
||||
}
|
||||
};
|
||||
|
||||
template <typename T>
|
||||
struct Heaviside {
|
||||
HOSTDEVICE T operator()(const T& val) const {
|
||||
return static_cast<T>(val > 0 ? 1 : 0);
|
||||
}
|
||||
};
|
||||
|
||||
template <typename Place, typename T>
|
||||
class MarginRankLossKernel : public framework::OpKernel<T> {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& ctx) const {
|
||||
auto* out_t = ctx.Output<framework::Tensor>("Out");
|
||||
auto* act_t = ctx.Output<framework::Tensor>("Activated");
|
||||
|
||||
auto* label_t = ctx.Input<framework::Tensor>("Label");
|
||||
auto* x1_t = ctx.Input<framework::Tensor>("X1");
|
||||
auto* x2_t = ctx.Input<framework::Tensor>("X2");
|
||||
|
||||
out_t->mutable_data<T>(ctx.GetPlace());
|
||||
act_t->mutable_data<T>(ctx.GetPlace());
|
||||
|
||||
auto margin = static_cast<T>(ctx.Attr<T>("margin"));
|
||||
auto out = framework::EigenVector<T>::Flatten(*out_t);
|
||||
auto act = framework::EigenVector<T>::Flatten(*act_t);
|
||||
|
||||
auto label = framework::EigenVector<T>::Flatten(*label_t);
|
||||
auto x1 = framework::EigenVector<T>::Flatten(*x1_t);
|
||||
auto x2 = framework::EigenVector<T>::Flatten(*x2_t);
|
||||
|
||||
auto& dev = ctx.GetEigenDevice<Place>();
|
||||
out.device(dev) = (-label * (x1 - x2) + margin).unaryExpr(ReLU<T>());
|
||||
act.device(dev) = out.unaryExpr(Heaviside<T>());
|
||||
}
|
||||
};
|
||||
|
||||
template <typename Place, typename T>
|
||||
class MarginRankLossGradKernel : public framework::OpKernel<T> {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& ctx) const {
|
||||
auto* d_x1_t =
|
||||
ctx.Output<framework::LoDTensor>(framework::GradVarName("X1"));
|
||||
auto* d_x2_t =
|
||||
ctx.Output<framework::LoDTensor>(framework::GradVarName("X2"));
|
||||
|
||||
auto* act_t = ctx.Input<framework::Tensor>("Activated");
|
||||
auto* d_out_t = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
|
||||
auto* label_t = ctx.Input<framework::Tensor>("Label");
|
||||
|
||||
auto d_out = framework::EigenVector<T>::Flatten(*d_out_t);
|
||||
auto act = framework::EigenVector<T>::Flatten(*act_t);
|
||||
auto label = framework::EigenVector<T>::Flatten(*label_t);
|
||||
auto& dev = ctx.GetEigenDevice<Place>();
|
||||
|
||||
// compute d_x1
|
||||
if (d_x1_t) {
|
||||
d_x1_t->mutable_data<T>(ctx.GetPlace());
|
||||
auto d_x1 = framework::EigenVector<T>::Flatten(*d_x1_t);
|
||||
d_x1.device(dev) = -d_out * act * label;
|
||||
}
|
||||
// compute d_x2
|
||||
if (d_x2_t) {
|
||||
d_x2_t->mutable_data<T>(ctx.GetPlace());
|
||||
auto d_x2 = framework::EigenVector<T>::Flatten(*d_x2_t);
|
||||
d_x2.device(dev) = d_out * act * label;
|
||||
}
|
||||
}
|
||||
};
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,129 @@
|
||||
/* 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_concat_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
class SequenceConcatOp : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
protected:
|
||||
void InferShape(framework::InferShapeContext* ctx) const override {
|
||||
PADDLE_ENFORCE(ctx->HasInputs("X"),
|
||||
"Inputs(X) of SequenceConcatOp should not be null.");
|
||||
PADDLE_ENFORCE(ctx->HasOutput("Out"),
|
||||
"Output(Out) of SequenceConcatOp should not be null.");
|
||||
const size_t level = static_cast<size_t>(ctx->Attrs().Get<int>("level"));
|
||||
const size_t axis = static_cast<size_t>(ctx->Attrs().Get<int>("axis"));
|
||||
PADDLE_ENFORCE(level == 0UL || level == 1UL,
|
||||
"The sequence_concat operator only accepts sequence "
|
||||
"or a nested sequence as its input.");
|
||||
auto ins_dims = ctx->GetInputsDim("X");
|
||||
framework::DDim out_dims = ins_dims[0];
|
||||
const size_t n = ins_dims.size();
|
||||
for (size_t i = 1; i < n; ++i) {
|
||||
out_dims[axis] += ins_dims[i][axis];
|
||||
}
|
||||
ctx->SetOutputDim("Out", out_dims);
|
||||
}
|
||||
};
|
||||
|
||||
class SequenceConcatOpMaker : public framework::OpProtoAndCheckerMaker {
|
||||
public:
|
||||
SequenceConcatOpMaker(framework::OpProto* proto,
|
||||
framework::OpAttrChecker* op_checker)
|
||||
: OpProtoAndCheckerMaker(proto, op_checker) {
|
||||
AddInput("X",
|
||||
"(A vector of LoDTensor), the input is a vector of LoDTensor, "
|
||||
"each of which is a variable-length sequence or nested sequence.")
|
||||
.AsDuplicable();
|
||||
AddOutput("Out",
|
||||
"(A LoDTensor), the variable-length output of "
|
||||
"sequence_concat Op.");
|
||||
AddAttr<int>("axis",
|
||||
"(int, default 0)"
|
||||
"The axis which the inputs will be joined with. "
|
||||
"If axis is 0, the inputs will be joined with LoD index.")
|
||||
.SetDefault(0);
|
||||
AddAttr<int>("level",
|
||||
"(int, default 0)"
|
||||
"The level at which the inputs will be joined. "
|
||||
"If the level is 0, the inputs will be joined at the nested "
|
||||
"sequence level. "
|
||||
"If the level is 1, the inputs will be joined at the "
|
||||
"sequence level. "
|
||||
"The level should be less than the level number of inputs.")
|
||||
.SetDefault(0);
|
||||
AddComment(R"DOC(
|
||||
The sequence_concat operator concatenates multiple LoDTensors.
|
||||
It only supports sequence (LoD Tensor with level number is 1)
|
||||
or a nested sequence (LoD tensor with level number is 2) as its input.
|
||||
- Case1:
|
||||
If the axis is other than 0(here, axis is 1 and level is 1),
|
||||
each input should have the same LoD information and the LoD
|
||||
information of the output keeps the same as the input.
|
||||
|
||||
LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)
|
||||
LoD(x1) = {{0,2,4}, {0,1,2,3,4}}; Dims(x1) = (4,4,4)
|
||||
LoD(Out) = {{0,2,4}, {0,1,2,3,4}}; Dims(Out) = (4,7,4)
|
||||
|
||||
- Case2:
|
||||
If the axis is 0(here, leve is 0), the inputs are concatenated along
|
||||
time steps, the LoD information of the output need to re-compute.
|
||||
|
||||
LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)
|
||||
LoD(x1) = {{0,3,5}, {0,1,2,3,5}}; Dims(x1) = (5,3,4)
|
||||
LoD(Out) = {{0,5,9}, {0,1,2,3,4,5,6,7,9}}; Dims(Out) = (9,3,4)
|
||||
|
||||
- Case3:
|
||||
If the axis is 0(here, level is 1).
|
||||
|
||||
LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)
|
||||
LoD(x1) = {{0,3,5}, {0,1,3,4,5}}; Dims(x1) = (5,3,4)
|
||||
LoD(Out) = {{0,5,9}, {0,2,5,7,9}}; Dims(Out) = (9,3,4)
|
||||
|
||||
NOTE: The levels of all the inputs should be the same.
|
||||
)DOC");
|
||||
}
|
||||
};
|
||||
|
||||
class SequenceConcatGradOp : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
protected:
|
||||
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"));
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
REGISTER_OP(sequence_concat, ops::SequenceConcatOp, ops::SequenceConcatOpMaker,
|
||||
sequence_concat_grad, ops::SequenceConcatGradOp);
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
sequence_concat,
|
||||
ops::SequenceConcatOpKernel<paddle::platform::CPUPlace, float>);
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
sequence_concat_grad,
|
||||
ops::SequenceConcatGradOpKernel<paddle::platform::CPUPlace, float>);
|
||||
@ -0,0 +1,25 @@
|
||||
/* 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. */
|
||||
|
||||
#define EIGEN_USE_GPU
|
||||
|
||||
#include "paddle/operators/sequence_concat_op.h"
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
REGISTER_OP_GPU_KERNEL(
|
||||
sequence_concat,
|
||||
ops::SequenceConcatOpKernel<paddle::platform::GPUPlace, float>);
|
||||
REGISTER_OP_GPU_KERNEL(
|
||||
sequence_concat_grad,
|
||||
ops::SequenceConcatGradOpKernel<paddle::platform::GPUPlace, float>);
|
||||
@ -0,0 +1,155 @@
|
||||
/* 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/strided_memcpy.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
using Tensor = framework::Tensor;
|
||||
using LoDTensor = framework::LoDTensor;
|
||||
using LoD = framework::LoD;
|
||||
|
||||
template <typename T>
|
||||
LoD concatLoD(const std::vector<const T*> ins, const size_t axis,
|
||||
const size_t level) {
|
||||
auto out_lod = ins[0]->lod();
|
||||
const size_t n = ins.size();
|
||||
if (axis == 0UL) {
|
||||
for (size_t i = 1; i < n; ++i) {
|
||||
for (size_t j = 0; j < ins[i]->lod()[0].size(); ++j) {
|
||||
out_lod[0][j] += ins[i]->lod()[0][j];
|
||||
}
|
||||
|
||||
if (ins[0]->NumLevels() == 2) {
|
||||
for (size_t j = 1; j < ins[i]->lod()[1].size(); ++j) {
|
||||
if (level == 0UL) {
|
||||
out_lod[1].push_back(out_lod[1].back() + ins[i]->lod()[1][j] -
|
||||
ins[i]->lod()[1][j - 1]);
|
||||
} else if (level == 1UL) {
|
||||
out_lod[1][j] += ins[1]->lod()[1][j];
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
return out_lod;
|
||||
}
|
||||
|
||||
template <typename Place, typename T>
|
||||
class SequenceConcatOpKernel : public framework::OpKernel<T> {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& ctx) const override {
|
||||
auto ins = ctx.MultiInput<LoDTensor>("X");
|
||||
auto* out = ctx.Output<LoDTensor>("Out");
|
||||
const size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
|
||||
const size_t level = static_cast<size_t>(ctx.Attr<int>("level"));
|
||||
const size_t n = ins.size();
|
||||
|
||||
for (size_t i = 1; i < n; ++i) {
|
||||
PADDLE_ENFORCE_EQ(ins[0]->NumLevels(), ins[i]->NumLevels(),
|
||||
"The levels of all the input LoDTensors "
|
||||
"should be the same.");
|
||||
PADDLE_ENFORCE_EQ(ins[0]->dims().size(), ins[i]->dims().size(),
|
||||
"The dimension size of all the input LoDTensors "
|
||||
"should be the same.");
|
||||
|
||||
const size_t dims_size = ins[i]->dims().size();
|
||||
for (size_t j = 0; j < dims_size; ++j) {
|
||||
if (j == axis) continue;
|
||||
PADDLE_ENFORCE_EQ(ins[0]->dims()[j], ins[i]->dims()[j],
|
||||
"Except for the dimension of the specified "
|
||||
"axis along which all the inputs are concatenated, "
|
||||
"dimensions of all the other axises of the input "
|
||||
"LoDTensors should be the same.");
|
||||
}
|
||||
}
|
||||
PADDLE_ENFORCE_GT(ins[0]->NumLevels(), level,
|
||||
"The levels of all the input LoDTensors "
|
||||
"should be greater than the specify level");
|
||||
|
||||
out->mutable_data<T>(ctx.GetPlace());
|
||||
auto out_lod = concatLoD<LoDTensor>(ins, axis, level);
|
||||
out->set_lod(out_lod);
|
||||
|
||||
auto out_lod_level = out_lod[level];
|
||||
for (size_t i = 0; i < out_lod_level.size() - 1; ++i) {
|
||||
Tensor out_t = out->Slice<T>(static_cast<int>(out_lod_level[i]),
|
||||
static_cast<int>(out_lod_level[i + 1]));
|
||||
auto out_stride = framework::stride(out_t.dims());
|
||||
size_t offset = 0;
|
||||
|
||||
for (size_t j = 0; j < n; ++j) {
|
||||
auto in_lod_level = ins[j]->lod()[level];
|
||||
auto in_stride = framework::stride(ins[j]->dims());
|
||||
Tensor in_t = ins[j]->Slice<T>(static_cast<int>(in_lod_level[i]),
|
||||
static_cast<int>(in_lod_level[i + 1]));
|
||||
size_t axis_dim = in_t.dims()[axis];
|
||||
StridedMemcpy<T>(ctx.device_context(), in_t.data<T>(), in_stride,
|
||||
in_t.dims(), out_stride, out_t.data<T>() + offset);
|
||||
offset += axis_dim * in_stride[axis];
|
||||
}
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
template <typename Place, typename T>
|
||||
class SequenceConcatGradOpKernel : public framework::OpKernel<T> {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& ctx) const override {
|
||||
auto ins = ctx.MultiInput<framework::LoDTensor>("X");
|
||||
auto* out_grad =
|
||||
ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"));
|
||||
auto x_grads =
|
||||
ctx.MultiOutput<framework::LoDTensor>(framework::GradVarName("X"));
|
||||
size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
|
||||
size_t level = static_cast<size_t>(ctx.Attr<int>("level"));
|
||||
const size_t n = x_grads.size();
|
||||
|
||||
// Set Grad(X) LoD as X
|
||||
for (size_t i = 0; i < n; i++) {
|
||||
x_grads[i]->set_lod(ins[i]->lod());
|
||||
x_grads[i]->mutable_data<T>(ctx.GetPlace());
|
||||
}
|
||||
|
||||
auto out_lod = concatLoD<LoDTensor>(ins, axis, level);
|
||||
auto out_lod_level = out_lod[level];
|
||||
|
||||
for (size_t i = 0; i < out_lod_level.size() - 1; ++i) {
|
||||
Tensor out_grad_t =
|
||||
out_grad->Slice<T>(static_cast<int>(out_lod_level[i]),
|
||||
static_cast<int>(out_lod_level[i + 1]));
|
||||
auto out_grad_stride = framework::stride(out_grad_t.dims());
|
||||
size_t offset = 0;
|
||||
|
||||
for (size_t j = 0; j < n; ++j) {
|
||||
auto x_grad_lod_level = x_grads[j]->lod()[level];
|
||||
auto x_grad_stride = framework::stride(x_grads[j]->dims());
|
||||
Tensor x_grad_t =
|
||||
x_grads[j]->Slice<T>(static_cast<int>(x_grad_lod_level[i]),
|
||||
static_cast<int>(x_grad_lod_level[i + 1]));
|
||||
size_t axis_dim = x_grad_t.dims()[axis];
|
||||
StridedMemcpy<T>(ctx.device_context(), out_grad_t.data<T>() + offset,
|
||||
out_grad_stride, out_grad_t.dims(), x_grad_stride,
|
||||
x_grad_t.data<T>());
|
||||
offset += axis_dim * out_grad_stride[axis];
|
||||
}
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,39 @@
|
||||
import unittest
|
||||
import numpy as np
|
||||
from op_test import OpTest
|
||||
|
||||
|
||||
class TestMarginRankLossOp(OpTest):
|
||||
def setUp(self):
|
||||
self.op_type = "margin_rank_loss"
|
||||
batch_size = 5
|
||||
margin = 0.5
|
||||
# labels_{i} = {-1, 1}
|
||||
label = 2 * np.random.randint(
|
||||
0, 2, size=(batch_size, 1)).astype("float32") - 1
|
||||
x1 = np.random.random((batch_size, 1)).astype("float32")
|
||||
x2 = np.random.random((batch_size, 1)).astype("float32")
|
||||
# loss = max(0, -label * (x1 - x2) + margin)
|
||||
loss = -label * (x1 - x2) + margin
|
||||
loss = np.where(loss > 0, loss, 0)
|
||||
act = np.where(loss > 0, 1., 0.)
|
||||
|
||||
self.attrs = {'margin': margin}
|
||||
self.inputs = {'Label': label, 'X1': x1, 'X2': x2}
|
||||
self.outputs = {'Activated': act, 'Out': loss}
|
||||
|
||||
def test_check_output(self):
|
||||
self.check_output()
|
||||
|
||||
def test_check_grad(self):
|
||||
self.check_grad(["X1", "X2"], "Out")
|
||||
|
||||
def test_check_grad_ignore_x1(self):
|
||||
self.check_grad(["X2"], "Out", no_grad_set=set('X1'))
|
||||
|
||||
def test_check_grad_ignore_x2(self):
|
||||
self.check_grad(["X1"], "Out", no_grad_set=set('X2'))
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
unittest.main()
|
||||
@ -0,0 +1,77 @@
|
||||
import unittest
|
||||
import numpy as np
|
||||
from op_test import OpTest
|
||||
|
||||
|
||||
class TestConcatOp(OpTest):
|
||||
def set_data(self):
|
||||
# two level, batch size is 3
|
||||
x0 = np.random.random((4, 6, 3)).astype('float32')
|
||||
lod0 = [[0, 2, 4], [0, 1, 2, 3, 4]]
|
||||
x1 = np.random.random((4, 8, 3)).astype('float32')
|
||||
lod1 = [[0, 2, 4], [0, 1, 2, 3, 4]]
|
||||
axis = 1
|
||||
level = 1
|
||||
self.inputs = {'X': [('x0', (x0, lod0)), ('x1', (x1, lod1))]}
|
||||
self.attrs = {'axis': axis, 'level': level}
|
||||
outs = []
|
||||
for i in range(4):
|
||||
sub_x0 = x0[lod0[level][i]:lod0[level][i + 1], :]
|
||||
sub_x1 = x1[lod1[level][i]:lod1[level][i + 1], :]
|
||||
outs.append(np.concatenate((sub_x0, sub_x1), axis=axis))
|
||||
|
||||
self.outputs = {'Out': np.concatenate(outs, axis=0)}
|
||||
|
||||
def setUp(self):
|
||||
self.op_type = "sequence_concat"
|
||||
self.set_data()
|
||||
|
||||
def test_check_output(self):
|
||||
self.check_output()
|
||||
|
||||
def test_check_grad(self):
|
||||
self.check_grad(['x0'], 'Out')
|
||||
|
||||
|
||||
class TestConcatOpDiffLod(TestConcatOp):
|
||||
def set_data(self):
|
||||
# two level, batch size is 3
|
||||
x0 = np.random.random((4, 6, 3)).astype('float32')
|
||||
lod0 = [[0, 2, 4], [0, 1, 2, 3, 4]]
|
||||
x1 = np.random.random((5, 6, 3)).astype('float32')
|
||||
lod1 = [[0, 3, 5], [0, 1, 2, 3, 5]]
|
||||
axis = 0
|
||||
level = 1
|
||||
self.inputs = {'X': [('x0', (x0, lod0)), ('x1', (x1, lod1))]}
|
||||
self.attrs = {'axis': axis, 'level': level}
|
||||
outs = []
|
||||
for i in range(4):
|
||||
sub_x0 = x0[lod0[level][i]:lod0[level][i + 1], :]
|
||||
sub_x1 = x1[lod1[level][i]:lod1[level][i + 1], :]
|
||||
outs.append(np.concatenate((sub_x0, sub_x1), axis=axis))
|
||||
|
||||
self.outputs = {'Out': np.concatenate(outs, axis=0)}
|
||||
|
||||
|
||||
class TestConcatOpLevelZero(TestConcatOp):
|
||||
def set_data(self):
|
||||
# two level, batch size is 3
|
||||
x0 = np.random.random((4, 3, 4)).astype('float32')
|
||||
lod0 = [[0, 2, 4], [0, 1, 2, 3, 4]]
|
||||
x1 = np.random.random((5, 3, 4)).astype('float32')
|
||||
lod1 = [[0, 3, 5], [0, 1, 3, 4, 5]]
|
||||
axis = 0
|
||||
level = 0
|
||||
self.inputs = {'X': [('x0', (x0, lod0)), ('x1', (x1, lod1))]}
|
||||
self.attrs = {'axis': axis, 'level': level}
|
||||
outs = []
|
||||
for i in range(2):
|
||||
sub_x0 = x0[lod0[level][i]:lod0[level][i + 1], :]
|
||||
sub_x1 = x1[lod1[level][i]:lod1[level][i + 1], :]
|
||||
outs.append(np.concatenate((sub_x0, sub_x1), axis=axis))
|
||||
|
||||
self.outputs = {'Out': np.concatenate(outs, axis=0)}
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
unittest.main()
|
||||
Loading…
Reference in new issue