fc_op.cc 5.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* 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. */

#include "paddle/fluid/operators/fc_op.h"
16
#include <vector>
17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32

namespace paddle {
namespace operators {

void FCOp::InferShape(framework::InferShapeContext* ctx) const {
  PADDLE_ENFORCE(ctx->HasInput("Input"),
                 "X(Input) of Fully Connected should not be null.");
  PADDLE_ENFORCE(ctx->HasOutput("Out"),
                 "Out(Output) of Fully Connected should not be null.");
  PADDLE_ENFORCE(ctx->HasInput("W"),
                 "W(Input) of Fully Connected should not be null.");

  auto in_dims = ctx->GetInputDim("Input");
  auto w_dims = ctx->GetInputDim("W");
  std::vector<int64_t> output_shape({in_dims[0], w_dims[1]});

T
tensor-tang 已提交
33 34 35 36 37 38
  if (ctx->HasInput("Bias")) {
    auto bias_dims = ctx->GetInputDim("Bias");
    PADDLE_ENFORCE_EQ(bias_dims[0], 1, "The shape of Bias must be [1, dim].");
    PADDLE_ENFORCE_EQ(bias_dims[1], framework::product(w_dims) / w_dims[0],
                      "The shape of Bias must be [1, dim].");
  }
39
  PADDLE_ENFORCE(in_dims.size() == 2 || in_dims.size() == 4,
M
mozga-intel 已提交
40
                 "Fully Connected input should be 2-D or 4-D tensor.");
41

42 43
  PADDLE_ENFORCE(w_dims.size() == 2 || w_dims.size() == 4,
                 "Fully Connected input should be 2-D or 4-D tensor.");
44

T
tensor-tang 已提交
45 46 47 48
  PADDLE_ENFORCE_EQ(framework::product(w_dims) / w_dims[0],
                    framework::product(in_dims) / in_dims[0],
                    "Fully Connected input and weigth size do not match.");

49 50 51 52 53 54
  ctx->SetOutputDim("Out", framework::make_ddim(output_shape));
  ctx->ShareLoD("Input", "Out");
}

framework::OpKernelType FCOp::GetExpectedKernelType(
    const framework::ExecutionContext& ctx) const {
T
tensor-tang 已提交
55 56 57 58 59 60
  framework::LibraryType library = framework::LibraryType::kPlain;
  framework::DataLayout layout = framework::DataLayout::kAnyLayout;
  if (ctx.Attr<bool>("use_mkldnn");) {
    library = framework::LibraryType::kMKLDNN;
    layout = framework::DataLayout::kMKLDNN;
  }
61 62 63 64 65 66 67 68 69 70 71 72 73 74 75
  return framework::OpKernelType(
      framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
      layout, library);
}

void FCOpGrad::InferShape(framework::InferShapeContext* ctx) const {
  auto in_dims = ctx->GetInputDim("Input");
  auto w_dims = ctx->GetInputDim("W");

  if (ctx->HasOutput(framework::GradVarName("Input"))) {
    ctx->SetOutputDim(framework::GradVarName("Input"), in_dims);
  }
  if (ctx->HasOutput(framework::GradVarName("W"))) {
    ctx->SetOutputDim(framework::GradVarName("W"), w_dims);
  }
T
tensor-tang 已提交
76 77 78 79 80 81

  if (ctx->HasInput("Bias")) {
    auto bias_dims = ctx->GetInputDim("Bias");
    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Bias"));
    ctx->SetOutputDim(framework::GradVarName("Bias"), bias_dims);
  }
82 83 84 85
}

framework::OpKernelType FCOpGrad::GetExpectedKernelType(
    const framework::ExecutionContext& ctx) const {
T
tensor-tang 已提交
86 87 88 89 90 91
  framework::LibraryType library = framework::LibraryType::kPlain;
  framework::DataLayout layout = framework::DataLayout::kAnyLayout;
  if (ctx.Attr<bool>("use_mkldnn");) {
    library = framework::LibraryType::kMKLDNN;
    layout = framework::DataLayout::kMKLDNN;
  }
92 93 94 95 96
  return framework::OpKernelType(
      framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
      layout, library);
}

Y
Yu Yang 已提交
97
void FCOpMaker::Make() {
98
  AddInput("Input", "(Tensor) The input tensor of fully connected operator. ");
99
  AddInput("W", "(Tensor), The second input tensor of fc op.");
T
tensor-tang 已提交
100 101
  AddInput("Bias", "(Tensor, optional) Bias vector with shape (1 x D")
      .AsDispensable();
102
  AddOutput("Out", "(Tensor) The output tensor of fully connected operator. ");
103 104 105 106 107 108 109 110 111 112 113 114
  AddAttr<bool>("use_mkldnn",
                "(bool, default false) Only used in mkldnn kernel")
      .SetDefault(false);
  AddComment(R"DOC(
  Fully Connected Operator.

  The fully connected operation calculates the output based on the input, weights and bias attribute.
  The size of each dimension of the parameters checked in the infer-shape.
  The matrix of bias is generated by the mkldnn framework, when the bias_attr is True.
  Additional parametrs are use_mkldnn and bias_attr.
  The input(X) size and output(Out) size may be diffrent.

M
mozga-intel 已提交
115
  The fully connected layer only supports MKLDNN version
116 117 118
)DOC");
}

T
tensor-tang 已提交
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145
template <typename T>
class FCOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
    PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
                   "It must use CPUPlace.");
    auto& dev_ctx = ctx.template device_context<CPUDeviceContext>();
    auto blas = math::GetBlas<CPUDeviceContext, T>(dev_ctx);
    auto input = ctx.Input<Tensor>("Input");
    auto w = ctx.Input<Tensor>("W");
    auto b = ctx.Input<Tensor>("Bias");

    const T* input_data = input->data<T>();
    const T* w_data = w->data<T>();
    auto output = ctx.Output<Tensor>("Out");
    T* output_data = output->mutable_data<T>(ctx.GetPlace());

    auto in_dims = ctx->GetInputDim("Input");
    auto w_dims = ctx->GetInputDim("W");
    std::vector<int64_t> output_shape({in_dims[0], w_dims[1]});

    if (bias) {
      const T* bias_data = bias->data<T>();
    }
  }
};

146 147 148
}  // namespace operators
}  // namespace paddle

T
tensor-tang 已提交
149 150
namespace ops = paddle::operators;
REGISTER_OPERATOR(fc, ops::FCOp, ops::FCOpMaker,
151
                  paddle::framework::DefaultGradOpDescMaker<true>);
T
tensor-tang 已提交
152 153 154
REGISTER_OPERATOR(fc_grad, ops::FCOpGrad);
REGISTER_OP_CPU_KERNEL(fc, ops::FCMKLDNNOpKernel<float>,
                       ops::FCMKLDNNOpKernel<double>);