spectral_norm_op.cc 8.8 KB
Newer Older
1
/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
D
dengkaipeng 已提交
2 3 4 5 6 7 8 9 10 11 12
   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/spectral_norm_op.h"
Z
zhhsplendid 已提交
13 14 15

#include <memory>

D
dengkaipeng 已提交
16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38
#include "paddle/fluid/framework/op_registry.h"

namespace paddle {
namespace operators {

using framework::Tensor;

class SpectralNormOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

 protected:
  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("Weight"),
                   "Input(Weight) of SpectralNormOp should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("U"),
                   "Input(U) of SpectralNormOp should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("V"),
                   "Input(V) of SpectralNormOp should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput("Out"),
                   "Output(Out) of SpectralNormOp should not be null.");

    auto dim_weight = ctx->GetInputDim("Weight");
D
dengkaipeng 已提交
39 40 41
    auto rank_weight = dim_weight.size();
    PADDLE_ENFORCE(rank_weight >= 2 && rank_weight <= 5,
                   "The rank of Input(Weights) can only be 2, 3,"
D
dengkaipeng 已提交
42 43 44 45
                   "4, 5 for fc, conv1d, conv2d, conv3d layers.");

    int dim = ctx->Attrs().Get<int>("dim");
    int power_iters = ctx->Attrs().Get<int>("power_iters");
D
dengkaipeng 已提交
46
    PADDLE_ENFORCE(dim == 0 || dim == 1, "Attr(dim) can only be 0 or 1");
D
dengkaipeng 已提交
47 48 49
    PADDLE_ENFORCE(power_iters >= 0,
                   "Attr(power_iters) should be larger equal then 0");

D
dengkaipeng 已提交
50 51 52 53 54 55 56 57 58
    int h = dim_weight[dim];
    int w = 1;
    for (int i = 0; i < rank_weight; i++) {
      if (i != dim) {
        w *= dim_weight[i];
      }
    }
    auto dim_u = ctx->GetInputDim("U");
    auto dim_v = ctx->GetInputDim("V");
59 60 61 62 63 64 65 66 67 68 69 70 71

    if (ctx->IsRuntime() || (dim_u[0] > 0 && h > 0)) {
      PADDLE_ENFORCE_EQ(dim_u[0], h,
                        "Input(U) dims[0] should be equal to "
                        "Input(Weight) dims[Attr(dim)]");
    }

    if (ctx->IsRuntime() || (dim_v[0] > 0 && w > 0)) {
      PADDLE_ENFORCE_EQ(
          dim_v[0], w,
          "Input(V) dims[0] should be equal to "
          "the product of Input(Weight) dims except dims[Attr(dim)]");
    }
D
dengkaipeng 已提交
72

D
dengkaipeng 已提交
73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89
    ctx->SetOutputDim("Out", dim_weight);
    ctx->ShareLoD("Weight", /*->*/ "Out");
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(ctx.Input<Tensor>("Weight")->type(),
                                   ctx.GetPlace());
  }
};

class SpectralNormOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  void Make() override {
    AddInput("Weight",
             "The input weight tensor of spectral_norm operator, "
D
dengkaipeng 已提交
90
             "This can be a 2-D, 3-D, 4-D, 5-D tensor which is the "
K
Kaipeng Deng 已提交
91 92
             "weights of fc, conv1d, conv2d, conv3d layer. "
             "The data type is float32 or float64.");
D
dengkaipeng 已提交
93 94 95 96
    AddInput("U",
             "The weight_u tensor of spectral_norm operator, "
             "This can be a 1-D tensor in shape [H, 1],"
             "H is the 1st dimentions of Weight after reshape"
97 98
             "corresponding by Attr(dim). As for Attr(dim) = 1"
             "in conv2d layer with weight shape [M, C, K1, K2]"
D
dengkaipeng 已提交
99
             "Weight will be reshape to [C, M*K1*K2], U will"
100
             "be in shape [C, 1].");
D
dengkaipeng 已提交
101
    AddInput("V",
102
             "The weight_v tensor of spectral_norm operator, "
D
dengkaipeng 已提交
103 104 105 106 107
             "This can be a 1-D tensor in shape [W, 1], "
             "W is the 2nd dimentions of Weight after reshape "
             "corresponding by Attr(dim). As for Attr(dim) = 1 "
             "in conv2d layer with weight shape [M, C, K1, K2] "
             "Weight will be reshape to [C, M*K1*K2], V will "
108
             "be in shape [M*K1*K2, 1].");
D
dengkaipeng 已提交
109 110 111 112 113
    AddOutput("Out",
              "The output weight tensor of spectral_norm operator, "
              "This tensor is in same shape with Input(Weight).");

    AddAttr<int>("dim",
D
dengkaipeng 已提交
114 115
                 "The index of dimension which should be permuted "
                 "to the first before reshaping Input(Weight) to "
D
dengkaipeng 已提交
116 117
                 "matrix, it should be set as 0 if Input(Weight) is "
                 "the weight of fc layer, and should be set as 1 if "
D
dengkaipeng 已提交
118 119
                 "Input(Weight) is the weight of conv layer, "
                 "default 0.")
D
dengkaipeng 已提交
120 121
        .SetDefault(0);
    AddAttr<int>("power_iters",
D
dengkaipeng 已提交
122 123
                 "number of power iterations to calculate "
                 "spectral norm, default 1.")
D
dengkaipeng 已提交
124 125
        .SetDefault(1);
    AddAttr<float>("eps",
D
dengkaipeng 已提交
126
                   "epsilon for numerical stability in "
K
Kaipeng Deng 已提交
127 128 129
                   "calculating norms, it will be added to "
                   "the denominator to aviod divide zero. "
                   "Default 1e-12.")
D
dengkaipeng 已提交
130 131 132
        .SetDefault(1e-12);

    AddComment(R"DOC(
D
dengkaipeng 已提交
133
          This layer calculates the spectral normalization value of weight of
134 135
          fc, conv1d, conv2d, conv3d layers which should be 2-D, 3-D, 4-D, 5-D
          tensor.
D
dengkaipeng 已提交
136

137 138 139
          Spectral normalization stabilizes the training of critic in GANs
          (Generative Adversarial Networks). This layer rescaling weight tensor
          with spectral normalize value.
D
dengkaipeng 已提交
140

141
          For spectral normalization calculations, we rescaling weight
D
dengkaipeng 已提交
142
          tensor with :math:`\sigma`, while :math:`\sigma{\mathbf{W}}` is
143

D
dengkaipeng 已提交
144
            $$\sigma(\mathbf{W}) = \max_{\mathbf{h}: \mathbf{h} \ne 0} \\frac{\|\mathbf{W} \mathbf{h}\|_2}{\|\mathbf{h}\|_2}$$
145

D
dengkaipeng 已提交
146
          We calculate :math:`\sigma{\mathbf{W}}` through power iterations as
147

D
dengkaipeng 已提交
148
            $$
149
            \mathbf{v} = \mathbf{W}^{T} \mathbf{u}
D
dengkaipeng 已提交
150 151 152 153 154
            $$
            $$
            \mathbf{v} = \\frac{\mathbf{v}}{\|\mathbf{v}\|_2}
            $$
            $$
155
            \mathbf{u} = \mathbf{W}^{T} \mathbf{v}
D
dengkaipeng 已提交
156 157 158 159
            $$
            $$
            \mathbf{u} = \\frac{\mathbf{u}}{\|\mathbf{u}\|_2}
            $$
160

D
dengkaipeng 已提交
161
          And :math:`\sigma` should be
162

D
dengkaipeng 已提交
163
            $$\sigma{\mathbf{W}} = \mathbf{u}^{T} \mathbf{W} \mathbf{v}$$
164 165 166

          For details of spectral normalization, please refer to paper: 
          `Spectral Normalization <https://arxiv.org/abs/1802.05957>`_ .
D
dengkaipeng 已提交
167 168 169 170
         )DOC");
  }
};

Z
zhhsplendid 已提交
171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192
class SpectralNormGradOpDescMaker : 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("spectral_norm_grad");

    op->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
    op->SetInput("Weight", Input("Weight"));
    op->SetInput("U", Input("U"));
    op->SetInput("V", Input("V"));

    op->SetOutput(framework::GradVarName("Weight"), InputGrad("Weight"));

    op->SetAttrMap(Attrs());

    return op;
  }
};

D
dengkaipeng 已提交
193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221
class SpectralNormOpGrad : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

 protected:
  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("Weight"), "Input(Weight) should not be null");
    PADDLE_ENFORCE(ctx->HasInput("U"), "Input(U) should not be null");
    PADDLE_ENFORCE(ctx->HasInput("V"), "Input(V) should not be null");
    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
                   "Input(Out@GRAD) should not be null");
    auto dim_x = ctx->GetInputDim("Weight");
    if (ctx->HasOutput(framework::GradVarName("Weight"))) {
      ctx->SetOutputDim(framework::GradVarName("Weight"), dim_x);
    }
  }

  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(ctx.Input<Tensor>("Weight")->type(),
                                   ctx.GetPlace());
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
REGISTER_OPERATOR(spectral_norm, ops::SpectralNormOp, ops::SpectralNormOpMaker,
Z
zhhsplendid 已提交
222
                  ops::SpectralNormGradOpDescMaker);
D
dengkaipeng 已提交
223 224 225 226 227 228 229 230 231
REGISTER_OPERATOR(spectral_norm_grad, ops::SpectralNormOpGrad);
REGISTER_OP_CPU_KERNEL(
    spectral_norm,
    ops::SpectralNormKernel<paddle::platform::CPUDeviceContext, float>,
    ops::SpectralNormKernel<paddle::platform::CPUDeviceContext, double>);
REGISTER_OP_CPU_KERNEL(
    spectral_norm_grad,
    ops::SpectralNormGradKernel<paddle::platform::CPUDeviceContext, float>,
    ops::SpectralNormGradKernel<paddle::platform::CPUDeviceContext, double>);