adamw_op.h 6.4 KB
Newer Older
Z
zhaoyingli 已提交
1
/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
R
Roc 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

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/operators/optimizers/adam_op.h>

namespace paddle {
namespace operators {

class AdamWOp : public AdamOp {
  using AdamOp::AdamOp;
};

Z
zhaoyingli 已提交
25
struct GPUAdamW;
R
Roc 已提交
26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49
struct CPUAdamW;

template <typename T, typename Flavour>
class AdamWFunctor;

template <typename T>
class AdamWFunctor<T, CPUAdamW> {
 private:
  const float coeff_;
  const float learning_rate_;
  T* param_;

 public:
  AdamWFunctor(const float& coeff, const float& learning_rate, T* param)
      : coeff_(coeff), learning_rate_(learning_rate), param_(param) {}

  inline HOSTDEVICE void operator()(size_t numel) const {
    Eigen::Map<Eigen::Array<T, 1, Eigen::Dynamic>> param{
        param_, static_cast<Eigen::Index>(numel)};
    // Calculation
    param = param * (1.0f - learning_rate_ * coeff_);
  }
};

Z
zhaoyingli 已提交
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 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
template <typename T, typename Flavour, typename MT = T>
class SparseAdamWFunctor;

template <typename T, typename MT>
class SparseAdamWFunctor<T, GPUAdamW, MT> {
 private:
  MT beta1_;
  MT beta2_;
  MT epsilon_;
  MT coeff_;

  const MT* beta1_pow_;
  const MT* beta2_pow_;
  const MT* moment1_;
  MT* moment1_out_;
  const MT* moment2_;
  MT* moment2_out_;
  const MT* lr_;
  const T* grad_;
  const T* param_;
  T* param_out_;
  const MT* master_param_;
  MT* master_param_out_;

  const int64_t* rows_;
  int64_t row_numel_;
  int64_t row_count_;
  bool lazy_mode_;

 public:
  SparseAdamWFunctor(MT beta1, MT beta2, MT epsilon, MT coeff,
                     const MT* beta1_pow, const MT* beta2_pow, const MT* mom1,
                     MT* mom1_out, const MT* mom2, MT* mom2_out, const MT* lr,
                     const T* grad, const T* param, T* param_out,
                     const MT* master_param, MT* master_param_out,
                     const int64_t* rows, int64_t row_numel, int64_t row_count,
                     bool lazy_mode)
      : beta1_(beta1),
        beta2_(beta2),
        epsilon_(epsilon),
        coeff_(coeff),
        beta1_pow_(beta1_pow),
        beta2_pow_(beta2_pow),
        moment1_(mom1),
        moment1_out_(mom1_out),
        moment2_(mom2),
        moment2_out_(mom2_out),
        lr_(lr),
        grad_(grad),
        param_(param),
        param_out_(param_out),
        master_param_(master_param),
        master_param_out_(master_param_out),
        rows_(rows),
        row_numel_(row_numel),
        row_count_(row_count),
        lazy_mode_(lazy_mode) {}

  inline HOSTDEVICE void adamw_update(size_t i, MT g) const {
    // The following code is the same as dense
    MT mom1 = moment1_[i];
    MT mom2 = moment2_[i];
    MT lr = *lr_;
    MT beta1_pow = *beta1_pow_;
    MT beta2_pow = *beta2_pow_;
    MT p = master_param_ ? master_param_[i] : static_cast<MT>(param_[i]);

    // Calculation
    MT wd = static_cast<MT>(1.0) - coeff_ * lr;
    lr *= sqrt(static_cast<MT>(1.0) - beta2_pow) /
          (static_cast<MT>(1.0) - beta1_pow);

    mom1 = beta1_ * mom1 + (static_cast<MT>(1.0) - beta1_) * g;
    mom2 = beta2_ * mom2 + (static_cast<MT>(1.0) - beta2_) * g * g;
    p = wd * p -
        lr * (mom1 /
              (sqrt(mom2) + epsilon_ * sqrt(static_cast<MT>(1.0) - beta2_pow)));

    // Write back to global memory
    moment1_out_[i] = mom1;
    moment2_out_[i] = mom2;
    param_out_[i] = static_cast<T>(p);
    if (master_param_out_) {
      master_param_out_[i] = p;
    }
  }

  inline HOSTDEVICE void operator()(size_t i) const {
    auto row_idx =
        math::BinarySearch<int64_t>(rows_, row_count_, i / row_numel_);
    if (lazy_mode_ && row_idx < 0) {
      return;
    } else {
      MT g = row_idx >= 0
                 ? static_cast<MT>(grad_[row_idx * row_numel_ + i % row_numel_])
                 : static_cast<MT>(0);
      adamw_update(i, g);
    }
  }
};

R
Roc 已提交
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 DeviceContext, typename T>
class AdamWOpKernel : public AdamOpKernel<DeviceContext, T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    const auto* param_var = ctx.InputVar("Param");
    PADDLE_ENFORCE_EQ(param_var->IsType<framework::LoDTensor>(), true,
                      platform::errors::InvalidArgument(
                          "The Var(%s)'s type should be LoDTensor, "
                          "but the received is %s",
                          ctx.InputNames("Param").front(),
                          framework::ToTypeName(param_var->Type())));

    using paddle::framework::LoDTensor;
    bool skip_update = false;
    // TODO(liupeng):
    if (ctx.HasInput("SkipUpdate")) {
      VLOG(3) << "Has SkipUpdate";
      auto* skip_update_tensor = ctx.Input<framework::Tensor>("SkipUpdate");
      PADDLE_ENFORCE_EQ(skip_update_tensor->numel(), 1,
                        platform::errors::InvalidArgument(
                            "Input(SkipUpdate) size must be 1, but get %d",
                            skip_update_tensor->numel()));
      std::vector<bool> skip_update_vec;
      TensorToVector(*skip_update_tensor, ctx.device_context(),
                     &skip_update_vec);
      skip_update = skip_update_vec[0];
    }
    VLOG(3) << "Skip update" << skip_update;
    bool with_decay = ctx.Attr<bool>("with_decay");

    if (skip_update || !with_decay) {
      AdamOpKernel<DeviceContext, T>::Compute(ctx);
      return;
    }

    float coeff = ctx.Attr<float>("coeff");
    auto* lr = ctx.Input<LoDTensor>("LearningRate");

    LoDTensor* param;

    if (ctx.HasInput("MasterParam")) {
      // TODO(liupeng): master
      param = const_cast<LoDTensor*>(ctx.Input<LoDTensor>("MasterParam"));
    } else {
      param = const_cast<LoDTensor*>(ctx.Input<LoDTensor>("Param"));
    }

    // AdamWFunctor(float coeff, const float* learning_rate, T* parma)
    AdamWFunctor<T, CPUAdamW> functor(coeff, *lr->data<float>(),
                                      param->data<T>());
    functor(param->numel());

    AdamOpKernel<DeviceContext, T>::Compute(ctx);
  }
};
}  // namespace operators
}  // namespace paddle