spectral_norm_op.h 9.9 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. */

#pragma once
D
dengkaipeng 已提交
13
#include <vector>
D
dengkaipeng 已提交
14 15
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
16
#include "paddle/pten/kernels/funcs/blas/blas.h"
17
#include "paddle/pten/kernels/funcs/math_function.h"
D
dengkaipeng 已提交
18 19 20 21 22 23 24 25 26 27 28 29 30

namespace paddle {
namespace operators {

template <typename T, size_t D, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
using Tensor = framework::Tensor;

using Array1 = Eigen::DSizes<int64_t, 1>;
using Array2 = Eigen::DSizes<int64_t, 2>;
using IndexPair = Eigen::IndexPair<int>;

D
dengkaipeng 已提交
31 32 33 34 35
template <typename DeviceContext, typename T>
static inline void TransCompute(const int rank, const Tensor& in, Tensor* out,
                                const std::vector<int>& perm,
                                const DeviceContext& dev_ctx) {
  if (rank <= 1 || rank > 5) {
36 37 38
    PADDLE_THROW(paddle::platform::errors::Fatal(
        "Weight rank of SpectralNorm should be in range [2, 5], but got %d.",
        rank));
D
dengkaipeng 已提交
39 40 41 42
  }

  switch (rank) {
    case 2:
43
      pten::funcs::Transpose<DeviceContext, T, 2> trans2;
D
dengkaipeng 已提交
44 45 46
      trans2(dev_ctx, in, out, perm);
      break;
    case 3:
47
      pten::funcs::Transpose<DeviceContext, T, 3> trans3;
D
dengkaipeng 已提交
48 49 50
      trans3(dev_ctx, in, out, perm);
      break;
    case 4:
51
      pten::funcs::Transpose<DeviceContext, T, 4> trans4;
D
dengkaipeng 已提交
52 53 54
      trans4(dev_ctx, in, out, perm);
      break;
    case 5:
55
      pten::funcs::Transpose<DeviceContext, T, 5> trans5;
D
dengkaipeng 已提交
56 57 58 59
      trans5(dev_ctx, in, out, perm);
      break;
    default:
      break;
D
dengkaipeng 已提交
60 61 62 63 64 65 66 67
  }
}

template <typename DeviceContext, typename T>
static inline void CalcMatrixSigmaAndNormWeight(
    Tensor* sigma, Tensor* u, Tensor* v, Tensor* weight, const int power_iters,
    const float eps, const framework::ExecutionContext& ctx) {
  auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
68
  auto blas = pten::funcs::GetBlas<DeviceContext, T>(ctx);
D
dengkaipeng 已提交
69 70
  auto sigma_t = EigenTensor<T, 2>::From(*sigma);
  auto weight_t = EigenTensor<T, 2>::From(*weight);
71 72
  auto u_t = EigenTensor<T, 2>::From(*u);
  auto v_t = EigenTensor<T, 2>::From(*v);
D
dengkaipeng 已提交
73 74 75 76 77

  const int h = weight->dims()[0];
  const int w = weight->dims()[1];

  for (int i = 0; i < power_iters; i++) {
D
dengkaipeng 已提交
78
    // V = W^T * U / ||W^T * U||_2
79
    blas.MatMul(*weight, true, *u, false, T(1), v, T(0));
D
dengkaipeng 已提交
80 81 82 83
    auto v_t_norm =
        v_t.square().sum().sqrt().eval().reshape(Array1(1)).broadcast(
            Array1(w));
    v_t.device(place) = v_t / (v_t_norm + v_t_norm.constant(eps));
D
dengkaipeng 已提交
84
    // U = W^T * V / ||W^T * V||_2
85
    blas.MatMul(*weight, false, *v, false, T(1), u, T(0));
D
dengkaipeng 已提交
86 87 88 89 90
    auto u_t_norm =
        u_t.square().sum().sqrt().eval().reshape(Array1(1)).broadcast(
            Array1(h));
    u_t.device(place) = u_t / (u_t_norm + u_t_norm.constant(eps));
  }
91 92 93 94 95
  Tensor weight_v;
  weight_v.mutable_data<T>({h, 1}, ctx.GetPlace());
  blas.MatMul(*weight, false, *v, false, T(1), &weight_v, T(0));
  auto weight_v_t = EigenTensor<T, 2>::From(weight_v);
  sigma_t.device(place) = (u_t * weight_v_t)
D
dengkaipeng 已提交
96 97 98 99 100 101 102 103 104 105 106
                              .sum()
                              .eval()
                              .reshape(Array2(1, 1))
                              .broadcast(Array2(h, w));
  weight_t.device(place) = weight_t / sigma_t;
}

template <typename DeviceContext, typename T>
class SpectralNormKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
D
dengkaipeng 已提交
107
    auto& dev_ctx = ctx.template device_context<DeviceContext>();
D
dengkaipeng 已提交
108 109 110 111 112 113 114 115 116
    auto weight = ctx.Input<Tensor>("Weight");
    auto u = ctx.Input<Tensor>("U");
    auto v = ctx.Input<Tensor>("V");
    auto out = ctx.Output<Tensor>("Out");

    int dim = ctx.Attr<int>("dim");
    int power_iters = ctx.Attr<int>("power_iters");
    float eps = ctx.Attr<float>("eps");

D
dengkaipeng 已提交
117 118 119
    const int h = u->dims()[0];
    const int w = v->dims()[0];

D
dengkaipeng 已提交
120
    Tensor weight_mat;
D
dengkaipeng 已提交
121 122 123 124 125 126 127 128 129 130 131 132 133
    auto dims = weight->dims();
    const int rank = dims.size();
    std::vector<int> real_dims;
    if (dim != 0) {
      std::vector<int> perm;
      perm.push_back(dim);
      real_dims.push_back(dims[dim]);
      for (int i = 0; i < rank; i++) {
        if (i != dim) {
          perm.push_back(i);
          real_dims.push_back(dims[i]);
        }
      }
134
      weight_mat.mutable_data<T>(pten::make_ddim(real_dims), ctx.GetPlace());
D
dengkaipeng 已提交
135 136 137 138 139
      TransCompute<DeviceContext, T>(rank, *weight, &weight_mat, perm, dev_ctx);
    } else {
      for (int i = 0; i < rank; i++) {
        real_dims.push_back(i);
      }
140
      paddle::framework::TensorCopySync(*weight, ctx.GetPlace(), &weight_mat);
D
dengkaipeng 已提交
141
    }
142
    weight_mat = weight_mat.Resize({h, w});
D
dengkaipeng 已提交
143 144

    Tensor sigma;
145
    sigma.mutable_data<T>(weight_mat.dims(), ctx.GetPlace());
D
dengkaipeng 已提交
146
    Tensor uu, vv;
147 148
    paddle::framework::TensorCopySync(*u, ctx.GetPlace(), &uu);
    paddle::framework::TensorCopySync(*v, ctx.GetPlace(), &vv);
D
dengkaipeng 已提交
149
    CalcMatrixSigmaAndNormWeight<DeviceContext, T>(
150 151
        &sigma, &(uu.Resize({h, 1})), &(vv.Resize({w, 1})), &weight_mat,
        power_iters, eps, ctx);
D
dengkaipeng 已提交
152 153 154 155 156 157 158 159 160 161 162 163 164 165

    if (dim != 0) {
      std::vector<int> perm;
      for (int i = 0; i < rank; i++) {
        if (i < dim) {
          perm.push_back(i + 1);
        } else if (i == dim) {
          perm.push_back(0);
        } else {
          perm.push_back(i);
        }
      }
      out->mutable_data<T>(dims, ctx.GetPlace());
      TransCompute<DeviceContext, T>(
166
          rank, weight_mat.Resize(pten::make_ddim(real_dims)), out, perm,
D
dengkaipeng 已提交
167 168
          dev_ctx);
    } else {
169 170
      paddle::framework::TensorCopySync(weight_mat.Resize(dims), ctx.GetPlace(),
                                        out);
D
dengkaipeng 已提交
171
    }
D
dengkaipeng 已提交
172 173 174 175 176 177
  }
};

template <typename DeviceContext, typename T>
class SpectralNormGradKernel : public framework::OpKernel<T> {
 public:
178 179
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
D
dengkaipeng 已提交
180
    auto& dev_ctx = ctx.template device_context<DeviceContext>();
181
    auto blas = pten::funcs::GetBlas<DeviceContext, T>(ctx);
182 183 184 185 186 187 188 189 190 191
    auto weight = ctx.Input<Tensor>("Weight");
    auto u = ctx.Input<Tensor>("U");
    auto v = ctx.Input<Tensor>("V");
    auto out_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
    auto weight_grad = ctx.Output<Tensor>(framework::GradVarName("Weight"));

    int dim = ctx.Attr<int>("dim");
    int power_iters = ctx.Attr<int>("power_iters");
    float eps = ctx.Attr<float>("eps");

D
dengkaipeng 已提交
192 193 194
    const int h = u->dims()[0];
    const int w = v->dims()[0];

195
    Tensor weight_mat, out_grad_mat;
D
dengkaipeng 已提交
196 197 198 199 200 201 202 203 204 205 206 207 208
    auto dims = weight->dims();
    const int rank = dims.size();
    std::vector<int> real_dims;
    if (dim != 0) {
      std::vector<int> perm;
      perm.push_back(dim);
      real_dims.push_back(dims[dim]);
      for (int i = 0; i < rank; i++) {
        if (i != dim) {
          perm.push_back(i);
          real_dims.push_back(dims[i]);
        }
      }
209 210
      weight_mat.mutable_data<T>(pten::make_ddim(real_dims), ctx.GetPlace());
      out_grad_mat.mutable_data<T>(pten::make_ddim(real_dims), ctx.GetPlace());
D
dengkaipeng 已提交
211 212 213 214 215 216 217
      TransCompute<DeviceContext, T>(rank, *weight, &weight_mat, perm, dev_ctx);
      TransCompute<DeviceContext, T>(rank, *out_grad, &out_grad_mat, perm,
                                     dev_ctx);
    } else {
      for (int i = 0; i < rank; i++) {
        real_dims.push_back(i);
      }
218 219 220
      paddle::framework::TensorCopySync(*weight, ctx.GetPlace(), &weight_mat);
      paddle::framework::TensorCopySync(*out_grad, ctx.GetPlace(),
                                        &out_grad_mat);
D
dengkaipeng 已提交
221
    }
222 223 224 225 226 227
    weight_mat = weight_mat.Resize({h, w});
    out_grad_mat = out_grad_mat.Resize({h, w});

    Tensor sigma;
    sigma.mutable_data<T>(weight_mat.dims(), ctx.GetPlace());
    Tensor uu, vv;
228 229
    paddle::framework::TensorCopySync(*u, ctx.GetPlace(), &uu);
    paddle::framework::TensorCopySync(*v, ctx.GetPlace(), &vv);
230 231 232 233 234 235 236 237 238
    CalcMatrixSigmaAndNormWeight<DeviceContext, T>(
        &sigma, &(uu.Resize({h, 1})), &(vv.Resize({w, 1})), &weight_mat,
        power_iters, eps, ctx);

    Tensor uv;
    uv.mutable_data<T>({h, w}, ctx.GetPlace());
    blas.MatMul(uu.Resize({h, 1}), false, vv.Resize({w, 1}), false, T(1), &uv,
                T(0));

D
dengkaipeng 已提交
239
    Tensor weight_grad_mat;
240 241 242 243 244 245 246 247 248
    weight_grad_mat.mutable_data<T>({h, w}, ctx.GetPlace());
    auto weight_grad_mat_t = EigenTensor<T, 2>::From(weight_grad_mat);
    auto weight_mat_t = EigenTensor<T, 2>::From(weight_mat);
    auto out_grad_mat_t = EigenTensor<T, 2>::From(out_grad_mat);
    auto sigma_t = EigenTensor<T, 2>::From(sigma);
    auto uv_t = EigenTensor<T, 2>::From(uv);
    weight_mat_t.device(place) =
        weight_mat_t.sum().eval().reshape(Array2(1, 1)).broadcast(Array2(h, w));
    weight_grad_mat_t.device(place) =
D
dengkaipeng 已提交
249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
        out_grad_mat_t * (out_grad_mat_t.constant(1.0) - uv_t * weight_mat_t) /
        sigma_t;

    if (dim != 0) {
      std::vector<int> perm;
      for (int i = 0; i < rank; i++) {
        if (i < dim) {
          perm.push_back(i + 1);
        } else if (i == dim) {
          perm.push_back(0);
        } else {
          perm.push_back(i);
        }
      }
      weight_grad->mutable_data<T>(dims, ctx.GetPlace());
      TransCompute<DeviceContext, T>(
265 266
          rank, weight_grad_mat.Resize(pten::make_ddim(real_dims)), weight_grad,
          perm, dev_ctx);
D
dengkaipeng 已提交
267
    } else {
268 269
      paddle::framework::TensorCopySync(weight_grad_mat.Resize(dims),
                                        ctx.GetPlace(), weight_grad);
D
dengkaipeng 已提交
270
    }
271
  }
D
dengkaipeng 已提交
272 273 274 275
};

}  // namespace operators
}  // namespace paddle