heqiaozhi
6 years ago
parent
1c8b34ddd2
commit
54dddee37e
@ -0,0 +1,163 @@
|
|||||||
|
/* Copyright (c) 2018 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/fluid/operators/cvm_op.h"
|
||||||
|
#include "paddle/fluid/operators/math/math_function.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
|
||||||
|
using Tensor = framework::Tensor;
|
||||||
|
|
||||||
|
class CVMOp : public framework::OperatorWithKernel {
|
||||||
|
public:
|
||||||
|
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||||
|
|
||||||
|
void InferShape(framework::InferShapeContext* ctx) const override {
|
||||||
|
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
|
||||||
|
PADDLE_ENFORCE(ctx->HasInput("CVM"), "Input(CVM) should be not null.");
|
||||||
|
PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) should be not null.");
|
||||||
|
|
||||||
|
auto x_dims = ctx->GetInputDim("X");
|
||||||
|
auto cvm_dims = ctx->GetInputDim("CVM");
|
||||||
|
PADDLE_ENFORCE_EQ(x_dims.size(), 2UL, "Input(X)'s rank should be 2.");
|
||||||
|
PADDLE_ENFORCE_EQ(cvm_dims.size(), 2UL, "Input(CVM)'s rank should be 2.");
|
||||||
|
PADDLE_ENFORCE_EQ(cvm_dims[1], 2UL,
|
||||||
|
"The 2nd dimension of "
|
||||||
|
"Input(CVM) should be 2.");
|
||||||
|
|
||||||
|
if (ctx->Attrs().Get<bool>("use_cvm")) {
|
||||||
|
ctx->SetOutputDim("Y", {x_dims[0], x_dims[1]});
|
||||||
|
} else {
|
||||||
|
ctx->SetOutputDim("Y", {x_dims[0], x_dims[1] - 2});
|
||||||
|
}
|
||||||
|
ctx->ShareLoD("X", /*->*/ "Y");
|
||||||
|
}
|
||||||
|
|
||||||
|
protected:
|
||||||
|
// Explicitly set that the data type of computation kernel of
|
||||||
|
// cvm
|
||||||
|
// is determined by its input "X".
|
||||||
|
framework::OpKernelType GetExpectedKernelType(
|
||||||
|
const framework::ExecutionContext& ctx) const override {
|
||||||
|
return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
|
||||||
|
ctx.device_context());
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
class CVMGradientOp : public framework::OperatorWithKernel {
|
||||||
|
public:
|
||||||
|
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||||
|
|
||||||
|
void InferShape(framework::InferShapeContext* ctx) const override {
|
||||||
|
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
|
||||||
|
PADDLE_ENFORCE(ctx->HasInput("CVM"), "Input(CVM) should be not null.");
|
||||||
|
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")),
|
||||||
|
"Input(Y@GRAD) should be not null.");
|
||||||
|
PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
|
||||||
|
"Output(X@GRAD) should be not null.");
|
||||||
|
|
||||||
|
auto x_dims = ctx->GetInputDim("X");
|
||||||
|
auto cvm_dims = ctx->GetInputDim("CVM");
|
||||||
|
auto dy_dims = ctx->GetInputDim(framework::GradVarName("Y"));
|
||||||
|
PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
|
||||||
|
PADDLE_ENFORCE_EQ(dy_dims.size(), 2, "Input(Y@Grad)'s rank should be 2.");
|
||||||
|
PADDLE_ENFORCE_EQ(cvm_dims.size(), 2, "Input(CVM)'s rank should be 2.");
|
||||||
|
|
||||||
|
PADDLE_ENFORCE_EQ(x_dims[0], dy_dims[0],
|
||||||
|
"The 1st dimension of Input(X) and Input(Y@Grad) should "
|
||||||
|
"be equal.");
|
||||||
|
|
||||||
|
PADDLE_ENFORCE_EQ(cvm_dims[1], 2,
|
||||||
|
"When Attr(soft_label) == false, the 2nd dimension of "
|
||||||
|
"Input(CVM) should be 2.");
|
||||||
|
ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
|
||||||
|
ctx->ShareLoD("X", framework::GradVarName("X"));
|
||||||
|
}
|
||||||
|
|
||||||
|
protected:
|
||||||
|
// Explicitly set that the data type of computation kernel of
|
||||||
|
// cvm
|
||||||
|
// is determined by its input "X".
|
||||||
|
framework::OpKernelType GetExpectedKernelType(
|
||||||
|
const framework::ExecutionContext& ctx) const override {
|
||||||
|
return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
|
||||||
|
ctx.device_context());
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
class CVMOpMaker : public framework::OpProtoAndCheckerMaker {
|
||||||
|
public:
|
||||||
|
void Make() override {
|
||||||
|
AddInput("X",
|
||||||
|
"(LodTensor, default LodTensor<float>), a 2-D tensor with shape "
|
||||||
|
"[N x D],"
|
||||||
|
" where N is the batch size and D is the emebdding dim. ");
|
||||||
|
AddInput("CVM",
|
||||||
|
"(Tensor), a 2-D Tensor with shape [N x 2], where N is the batch "
|
||||||
|
"size, 2 is show and click.");
|
||||||
|
AddOutput("Y",
|
||||||
|
"(LodTensor, default LodTensor<float>), a 2-D tensor with shape "
|
||||||
|
"[N x K].");
|
||||||
|
AddAttr<bool>("use_cvm", "bool, use cvm or not").SetDefault(true);
|
||||||
|
AddComment(R"DOC(
|
||||||
|
CVM Operator.
|
||||||
|
|
||||||
|
example:
|
||||||
|
input = fluid.layers.data(name=\"input\", shape=[-1, 1], lod_level=1, append_batch_size=False, dtype=\"int64\")
|
||||||
|
label = fluid.layers.data(name=\"label\", shape=[-1, 1], append_batch_size=False, dtype=\"int64\")
|
||||||
|
|
||||||
|
embed = fluid.layers.embedding(
|
||||||
|
input=input,
|
||||||
|
size=[100, 11],
|
||||||
|
dtype='float32')
|
||||||
|
|
||||||
|
ones = fluid.layers.fill_constant_batch_size_like(input=label, shape=[-1, 1], dtype=\"int64\", value=1)
|
||||||
|
show_clk = fluid.layers.cast(fluid.layers.concat([ones, label], axis=1), dtype='float32')
|
||||||
|
show_clk.stop_gradient = True
|
||||||
|
|
||||||
|
input_with_cvm = fluid.layers.continuous_value_model(embed, show_clk, True)
|
||||||
|
|
||||||
|
)DOC");
|
||||||
|
}
|
||||||
|
};
|
||||||
|
class CVMGradOpDescMaker : public framework::SingleGradOpDescMaker {
|
||||||
|
public:
|
||||||
|
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
|
||||||
|
|
||||||
|
protected:
|
||||||
|
std::unique_ptr<framework::OpDesc> Apply() const override {
|
||||||
|
std::unique_ptr<framework::OpDesc> op(new framework::OpDesc());
|
||||||
|
op->SetType("cvm_grad");
|
||||||
|
op->SetInput("X", Input("X"));
|
||||||
|
op->SetInput("CVM", Input("CVM"));
|
||||||
|
op->SetInput(framework::GradVarName("Y"), OutputGrad("Y"));
|
||||||
|
op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
|
||||||
|
op->SetOutput(framework::GradVarName("CVM"), InputGrad("CVM"));
|
||||||
|
op->SetAttrMap(Attrs());
|
||||||
|
return op;
|
||||||
|
}
|
||||||
|
};
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
||||||
|
|
||||||
|
namespace ops = paddle::operators;
|
||||||
|
REGISTER_OPERATOR(cvm, ops::CVMOp, ops::CVMOpMaker, ops::CVMGradOpDescMaker);
|
||||||
|
|
||||||
|
REGISTER_OPERATOR(cvm_grad, ops::CVMGradientOp);
|
||||||
|
|
||||||
|
REGISTER_OP_CPU_KERNEL(cvm, ops::CVMOpKernel<float>, ops::CVMOpKernel<double>);
|
||||||
|
|
||||||
|
REGISTER_OP_CPU_KERNEL(cvm_grad, ops::CVMGradOpKernel<float>,
|
||||||
|
ops::CVMGradOpKernel<double>);
|
||||||
@ -0,0 +1,105 @@
|
|||||||
|
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
||||||
|
|
||||||
|
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/fluid/framework/eigen.h"
|
||||||
|
#include "paddle/fluid/framework/op_registry.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
|
||||||
|
using Tensor = framework::Tensor;
|
||||||
|
using LoDTensor = framework::LoDTensor;
|
||||||
|
|
||||||
|
template <typename T>
|
||||||
|
class CVMOpKernel : public framework::OpKernel<T> {
|
||||||
|
public:
|
||||||
|
void Compute(const framework::ExecutionContext& context) const override {
|
||||||
|
const LoDTensor* x = context.Input<LoDTensor>("X");
|
||||||
|
const T* x_data = x->data<T>();
|
||||||
|
auto lod = x->lod()[0];
|
||||||
|
int64_t item_size = x->numel() / x->dims()[0];
|
||||||
|
int offset = 2;
|
||||||
|
if (!context.Attr<bool>("use_cvm")) {
|
||||||
|
item_size -= offset;
|
||||||
|
}
|
||||||
|
LoDTensor* y = context.Output<LoDTensor>("Y");
|
||||||
|
T* y_data = y->mutable_data<T>(context.GetPlace());
|
||||||
|
|
||||||
|
int seq_num = static_cast<int>(lod.size()) - 1;
|
||||||
|
for (int i = 0; i < seq_num; ++i) {
|
||||||
|
int64_t seq_len = static_cast<int64_t>(lod[i + 1] - lod[i]);
|
||||||
|
|
||||||
|
for (int j = 0; j < seq_len; ++j) {
|
||||||
|
if (context.Attr<bool>("use_cvm")) {
|
||||||
|
std::memcpy(y_data, x_data, item_size * sizeof(T));
|
||||||
|
y_data[0] = log(y_data[0] + 1);
|
||||||
|
y_data[1] = log(y_data[1] + 1) - y_data[0];
|
||||||
|
x_data += item_size;
|
||||||
|
y_data += item_size;
|
||||||
|
} else {
|
||||||
|
std::memcpy(y_data, x_data + offset, item_size * sizeof(T));
|
||||||
|
x_data += item_size + offset;
|
||||||
|
y_data += item_size;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
}
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
template <typename T>
|
||||||
|
class CVMGradOpKernel : public framework::OpKernel<T> {
|
||||||
|
public:
|
||||||
|
void Compute(const framework::ExecutionContext& context) const override {
|
||||||
|
LoDTensor* dx = context.Output<LoDTensor>(framework::GradVarName("X"));
|
||||||
|
T* dx_data = dx->mutable_data<T>(context.GetPlace());
|
||||||
|
|
||||||
|
const Tensor* cvm = context.Input<Tensor>("CVM");
|
||||||
|
const T* cvm_data = cvm->data<T>();
|
||||||
|
int offset = 2;
|
||||||
|
const framework::LoDTensor* dOut =
|
||||||
|
context.Input<framework::LoDTensor>(framework::GradVarName("Y"));
|
||||||
|
const T* dout_data = dOut->data<T>();
|
||||||
|
|
||||||
|
auto lod = dx->lod()[0];
|
||||||
|
int64_t item_size = dx->numel() / dx->dims()[0];
|
||||||
|
if (!context.Attr<bool>("use_cvm")) {
|
||||||
|
item_size -= offset;
|
||||||
|
}
|
||||||
|
|
||||||
|
int seq_num = static_cast<int>(lod.size()) - 1;
|
||||||
|
for (int i = 0; i < seq_num; ++i) {
|
||||||
|
int64_t seq_len = static_cast<int64_t>(lod[i + 1] - lod[i]);
|
||||||
|
|
||||||
|
for (int j = 0; j < seq_len; ++j) {
|
||||||
|
if (context.Attr<bool>("use_cvm")) {
|
||||||
|
std::memcpy(dx_data, dout_data, item_size * sizeof(T));
|
||||||
|
dx_data[0] = cvm_data[0];
|
||||||
|
dx_data[1] = cvm_data[1];
|
||||||
|
dx_data += item_size;
|
||||||
|
dout_data += item_size;
|
||||||
|
} else {
|
||||||
|
std::memcpy(dx_data + offset, dout_data, item_size * sizeof(T));
|
||||||
|
dx_data[0] = cvm_data[0];
|
||||||
|
dx_data[1] = cvm_data[1];
|
||||||
|
dx_data += item_size + offset;
|
||||||
|
dout_data += item_size;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
cvm_data += offset;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
};
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
||||||
Loading…
Reference in new issue