momentum_op.h 13.0 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
S
sidgoyal78 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15

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
dzhwinter 已提交
16
#include <string>
Y
Yi Wang 已提交
17 18
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
D
dzhwinter 已提交
19 20 21
#include "paddle/fluid/operators/math/algorithm.h"
#include "paddle/fluid/operators/math/selected_rows_functor.h"
#include "paddle/fluid/platform/for_range.h"
S
sidgoyal78 已提交
22 23 24 25

namespace paddle {
namespace operators {

D
dzhwinter 已提交
26 27 28 29 30
using framework::Tensor;
using framework::SelectedRows;
struct NoNesterov;
struct UseNesterov;

31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78
class MomentumOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

 protected:
  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("Param"),
                   "Input(param) of Momentum should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("Grad"),
                   "Input(grad) of Momentum should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("Velocity"),
                   "Input(velocity) of Momentum should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("LearningRate"),
                   "Input(LearningRate) of Momentum should not be null.");
    PADDLE_ENFORCE(
        ctx->GetInputsVarType("Param").front() ==
            framework::proto::VarType::LOD_TENSOR,
        "The input var's type should be LoDTensor, but the received is %s",
        ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front());

    PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
                   "Output(ParamOut) of Momentum should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput("VelocityOut"),
                   "Output(VelocityOut) of Momentum should not be null.");

    auto param_dim = ctx->GetInputDim("Param");
    if (ctx->GetInputsVarType("Grad")[0] ==
        framework::proto::VarType::LOD_TENSOR) {
      PADDLE_ENFORCE_EQ(
          param_dim, ctx->GetInputDim("Grad"),
          "Param and Grad input of MomentumOp should have the same dimension.");
      PADDLE_ENFORCE_EQ(
          param_dim, ctx->GetInputDim("Velocity"),
          "Param and Velocity of MomentumOp should have the same dimension.");
    }
    PADDLE_ENFORCE_EQ(framework::product(ctx->GetInputDim("LearningRate")), 1,
                      "Learning_rate should be a scalar");

    ctx->SetOutputDim("ParamOut", param_dim);
    ctx->SetOutputDim("VelocityOut", param_dim);
  }
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    auto input_data_type = framework::GetDataTypeOfVar(ctx.InputVar("Param"));
    return framework::OpKernelType(input_data_type, ctx.GetPlace());
  }
};

D
dzhwinter 已提交
79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207
template <typename T>
class CPUDenseMomentumFunctor {
 private:
  const Tensor* param;
  const Tensor* grad;
  const Tensor* velocity;
  const Tensor* learning_rate;
  const T mu;
  const T use_nesterov;
  Tensor* param_out;
  Tensor* velocity_out;

 public:
  CPUDenseMomentumFunctor(const Tensor* param, const Tensor* grad,
                          const Tensor* velocity, const Tensor* learning_rate,
                          const T mu, const bool use_nesterov,
                          Tensor* param_out, Tensor* velocity_out)
      : param(param),
        grad(grad),
        velocity(velocity),
        learning_rate(learning_rate),
        mu(mu),
        use_nesterov(use_nesterov),
        param_out(param_out),
        velocity_out(velocity_out) {}

  inline void operator()() {
    auto p_out = framework::EigenVector<T>::Flatten(*param_out);
    auto v_out = framework::EigenVector<T>::Flatten(*velocity_out);

    auto p = framework::EigenVector<T>::Flatten(*param);
    auto v = framework::EigenVector<T>::Flatten(*velocity);
    auto g = framework::EigenVector<T>::Flatten(*grad);
    auto* lr = learning_rate->data<T>();

    v_out = v * mu + g;
    if (use_nesterov) {
      p_out = p - (g + v_out * mu) * lr[0];
    } else {
      p_out = p - lr[0] * v_out;
    }
  }
};

template <typename T, typename UpdateMethod>
class DenseMomentumFunctor;

// NOTE(dzh) for performance.
// avoid if/else in inside kernel, implement GPU UseNesterov/NoNesterov as two
// functor.
template <typename T>
class DenseMomentumFunctor<T, UseNesterov> {
 private:
  const T* p_;
  const T* g_;
  const T* v_;
  const T* lr_;
  const T mu_;
  const int64_t num_;
  T* p_out_;
  T* v_out_;

 public:
  DenseMomentumFunctor(const T* p, const T* g, const T* v,
                       const T* learning_rate, const T mu, const int64_t num,
                       T* p_out, T* v_out)
      : p_(p),
        g_(g),
        v_(v),
        lr_(learning_rate),
        mu_(mu),
        num_(num),
        p_out_(p_out),
        v_out_(v_out) {}
  inline HOSTDEVICE void operator()(size_t i) const {
    // put memory access in register
    const T p = p_[i];
    const T g = g_[i];
    const T lr = lr_[0];
    const T v = v_[i];
    T v_out = v * mu_ + g;
    T p_out = p - (g + v_out * mu_) * lr;
    // write reigster to memory
    v_out_[i] = v_out;
    p_out_[i] = p_out;
  }
};

template <typename T>
class DenseMomentumFunctor<T, NoNesterov> {
 private:
  const T* p_;
  const T* g_;
  const T* v_;
  const T* lr_;
  const T mu_;
  const int64_t num_;
  T* p_out_;
  T* v_out_;

 public:
  DenseMomentumFunctor(const T* p, const T* g, const T* v,
                       const T* learning_rate, const T mu, const int64_t num,
                       T* p_out, T* v_out)
      : p_(p),
        g_(g),
        v_(v),
        lr_(learning_rate),
        mu_(mu),
        num_(num),
        p_out_(p_out),
        v_out_(v_out) {}
  inline HOSTDEVICE void operator()(size_t i) const {
    // put memory access in register
    const T p = p_[i];
    const T g = g_[i];
    const T lr = lr_[0];
    const T v = v_[i];
    T v_out = v * mu_ + g;
    T p_out = p - lr * v_out;
    // write reigster to memory
    v_out_[i] = v_out;
    p_out_[i] = p_out;
  }
};

template <typename T, typename UpdateMethod>
class SparseMomentumFunctor;

208
template <typename T>
D
dzhwinter 已提交
209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298
class SparseMomentumFunctor<T, UseNesterov> {
 private:
  const T* p_;
  const T* g_;
  const T* v_;
  const T* lr_;
  const T mu_;
  const int64_t* rows_;
  const int64_t row_numel_;
  const int64_t row_height_;
  T* p_out_;
  T* v_out_;

 public:
  SparseMomentumFunctor(const T* p, const T* g, const T* v, const T* lr,
                        const T mu, const int64_t* rows, int64_t row_numel,
                        int64_t row_height, T* p_out, T* v_out)
      : p_(p),
        g_(g),
        v_(v),
        lr_(lr),
        mu_(mu),
        rows_(rows),
        row_numel_(row_numel),
        row_height_(row_height),
        p_out_(p_out),
        v_out_(v_out) {}

  inline HOSTDEVICE void operator()(size_t i) {
    auto row_idx =
        math::BinarySearch<int64_t>(rows_, row_height_, i / row_numel_);
    T g = row_idx >= 0 ? g_[row_idx * row_numel_ + i % row_numel_] : 0;
    // put memory access in register
    const T p = p_[i];
    const T lr = lr_[0];
    const T v = v_[i];
    T v_out = v * mu_ + g;
    T p_out = p - (g + v_out * mu_) * lr;
    // write reigster to memory
    v_out_[i] = v_out;
    p_out_[i] = p_out;
  }
};

template <typename T>
class SparseMomentumFunctor<T, NoNesterov> {
 private:
  const T* p_;
  const T* g_;
  const T* v_;
  const T* lr_;
  const T mu_;
  const int64_t* rows_;
  const int64_t row_numel_;
  const int64_t row_height_;
  T* p_out_;
  T* v_out_;

 public:
  SparseMomentumFunctor(const T* p, const T* g, const T* v, const T* lr,
                        const T mu, const int64_t* rows, int64_t row_numel,
                        int64_t row_height, T* p_out, T* v_out)
      : p_(p),
        g_(g),
        v_(v),
        lr_(lr),
        mu_(mu),
        rows_(rows),
        row_numel_(row_numel),
        row_height_(row_height),
        p_out_(p_out),
        v_out_(v_out) {}

  inline HOSTDEVICE void operator()(size_t i) {
    auto row_idx =
        math::BinarySearch<int64_t>(rows_, row_height_, i / row_numel_);
    T g = row_idx >= 0 ? g_[row_idx * row_numel_ + i % row_numel_] : 0;
    // put memory access in register
    const T p = p_[i];
    const T lr = lr_[0];
    const T v = v_[i];
    T v_out = v * mu_ + g;
    T p_out = p - v_out * lr;
    // write reigster to memory
    v_out_[i] = v_out;
    p_out_[i] = p_out;
  }
};

template <typename DeviceContext, typename T>
S
sidgoyal78 已提交
299 300 301
class MomentumOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
302
    T mu = static_cast<T>(ctx.Attr<float>("mu"));
303
    bool use_nesterov = ctx.Attr<bool>("use_nesterov");
S
sidgoyal78 已提交
304

305 306 307
    auto learning_rate = ctx.Input<framework::Tensor>("LearningRate");
    auto param = ctx.Input<framework::Tensor>("Param");
    auto param_out = ctx.Output<framework::Tensor>("ParamOut");
D
dzhwinter 已提交
308 309 310 311 312
    auto* velocity = ctx.Input<framework::Tensor>("Velocity");
    auto velocity_out = ctx.Output<framework::Tensor>("VelocityOut");
    param_out->mutable_data<T>(ctx.GetPlace());
    velocity_out->mutable_data<T>(ctx.GetPlace());

313 314 315
    auto* grad_var = ctx.InputVar("Grad");
    if (grad_var->IsType<framework::LoDTensor>()) {
      auto grad = ctx.Input<framework::Tensor>("Grad");
D
dzhwinter 已提交
316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340
      if (platform::is_cpu_place(ctx.GetPlace())) {
        CPUDenseMomentumFunctor<T> functor(param, grad, velocity, learning_rate,
                                           mu, use_nesterov, param_out,
                                           velocity_out);
        functor();
      } else if (platform::is_gpu_place(ctx.GetPlace())) {
        platform::ForRange<DeviceContext> for_range(
            static_cast<const DeviceContext&>(ctx.device_context()),
            param->numel());
        if (use_nesterov) {
          DenseMomentumFunctor<T, UseNesterov> functor(
              param->data<T>(), grad->data<T>(), velocity->data<T>(),
              learning_rate->data<T>(), mu, param->numel(),
              param_out->mutable_data<T>(ctx.GetPlace()),
              velocity_out->mutable_data<T>(ctx.GetPlace()));
          for_range(functor);

        } else {
          DenseMomentumFunctor<T, NoNesterov> functor(
              param->data<T>(), grad->data<T>(), velocity->data<T>(),
              learning_rate->data<T>(), mu, param->numel(),
              param_out->mutable_data<T>(ctx.GetPlace()),
              velocity_out->mutable_data<T>(ctx.GetPlace()));
          for_range(functor);
        }
341
      }
D
dzhwinter 已提交
342

343 344 345
    } else if (grad_var->IsType<framework::SelectedRows>()) {
      // sparse update embedding with selectedrows
      auto grad = ctx.Input<framework::SelectedRows>("Grad");
S
sidgoyal78 已提交
346

347 348
      // sparse update maybe empty.
      if (grad->rows().size() == 0) {
349
        VLOG(30) << "Grad SelectedRows contains no data!";
350 351
        return;
      }
D
dzhwinter 已提交
352 353 354 355 356 357 358 359
      auto* merged_grad = const_cast<framework::Scope&>(ctx.scope())
                              .Var()
                              ->GetMutable<framework::SelectedRows>();
      math::scatter::MergeAdd<DeviceContext, T> merge_func;
      merge_func(ctx.template device_context<DeviceContext>(), *grad,
                 merged_grad);

      const int64_t* rows = nullptr;
D
dzhwinter 已提交
360
#ifdef PADDLE_WITH_CUDA
D
dzhwinter 已提交
361 362 363
      if (platform::is_gpu_place(ctx.GetPlace())) {
        rows = merged_grad->rows().CUDAData(ctx.GetPlace());
      } else {
D
dzhwinter 已提交
364
#endif
D
dzhwinter 已提交
365
        rows = merged_grad->rows().data();
D
dzhwinter 已提交
366
#ifdef PADDLE_WITH_CUDA
D
dzhwinter 已提交
367
      }
D
dzhwinter 已提交
368 369 370 371 372 373
#endif
      int64_t row_numel =
          merged_grad->value().numel() / merged_grad->rows().size();
      platform::ForRange<DeviceContext> for_range(
          static_cast<const DeviceContext&>(ctx.device_context()),
          param->numel());
D
dzhwinter 已提交
374 375 376
      if (use_nesterov) {
        SparseMomentumFunctor<T, UseNesterov> functor(
            param->data<T>(), merged_grad->value().data<T>(),
D
dzhwinter 已提交
377
            velocity->data<T>(), learning_rate->data<T>(), mu, rows, row_numel,
D
dzhwinter 已提交
378 379 380 381 382 383 384 385
            static_cast<int64_t>(merged_grad->rows().size()),
            param_out->mutable_data<T>(ctx.GetPlace()),
            velocity_out->mutable_data<T>(ctx.GetPlace()));
        for_range(functor);

      } else {
        SparseMomentumFunctor<T, NoNesterov> functor(
            param->data<T>(), merged_grad->value().data<T>(),
D
dzhwinter 已提交
386
            velocity->data<T>(), learning_rate->data<T>(), mu, rows, row_numel,
D
dzhwinter 已提交
387 388 389 390
            static_cast<int64_t>(merged_grad->rows().size()),
            param_out->mutable_data<T>(ctx.GetPlace()),
            velocity_out->mutable_data<T>(ctx.GetPlace()));
        for_range(functor);
391
      }
K
kavyasrinet 已提交
392
    } else {
D
dzhwinter 已提交
393 394 395 396
      PADDLE_THROW(
          string::Sprintf("MomentumOp only supports LoDTensor or SelectedRows "
                          "gradient, but the received Variable Type is %s",
                          grad_var->Type().name()));
K
kavyasrinet 已提交
397
    }
S
sidgoyal78 已提交
398 399 400 401 402
  }
};

}  // namespace operators
}  // namespace paddle