lstm_unit_op.h 4.6 KB
Newer Older
Z
zchen0211 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

   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
Z
lstm  
zchen0211 已提交
16
#include "glog/logging.h"
Z
zchen0211 已提交
17 18 19 20 21 22 23 24 25 26 27 28 29 30 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 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
#include "paddle/framework/op_registry.h"

namespace paddle {
namespace operators {

using framework::LoDTensor;
using framework::Tensor;

template <typename T>
inline T sigmoid(T x) {
  return 1. / (1. + exp(-x));
}

template <typename T>
inline T tanh(T x) {
  return 2. * sigmoid(2. * x) - 1.;
}

template <typename Place, typename T, typename AttrType = T>
class LstmUnitKernel : public framework::OpKernel {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
                   "It must use CPUPlace.");

    auto* x_tensor = ctx.Input<framework::Tensor>("X");
    auto* c_prev_tensor = ctx.Input<framework::Tensor>("C_prev");
    auto* c_tensor = ctx.Output<framework::Tensor>("C");
    auto* h_tensor = ctx.Output<framework::Tensor>("H");

    auto forget_bias = static_cast<T>(ctx.Attr<AttrType>("forget_bias"));

    int b_size = c_tensor->dims()[0];
    int D = c_tensor->dims()[1];

    T* C = c_tensor->mutable_data<T>(ctx.GetPlace());
    T* H = h_tensor->mutable_data<T>(ctx.GetPlace());

    const T* X = x_tensor->data<T>();
    const T* C_prev = c_prev_tensor->data<T>();

    for (int n = 0; n < b_size; ++n) {
      for (int d = 0; d < D; ++d) {
        const T i = sigmoid(X[d]);
        const T f = sigmoid(X[1 * D + d] + forget_bias);
        const T o = sigmoid(X[2 * D + d]);
        const T g = tanh(X[3 * D + d]);
        const T c_prev = C_prev[d];
        const T c = f * c_prev + i * g;
        C[d] = c;
        const T tanh_c = tanh(c);
        H[d] = o * tanh_c;
      }
      C_prev += D;
      X += 4 * D;
      C += D;
      H += D;
    }
  }
};

template <typename Place, typename T, typename AttrType = T>
class LstmUnitGradKernel : public framework::OpKernel {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
                   "It must use CPUPlace.");

    auto x_tensor = ctx.Input<Tensor>("X");
    auto c_prev_tensor = ctx.Input<Tensor>("C_prev");
    auto c_tensor = ctx.Input<Tensor>("C");
    auto h_tensor = ctx.Input<Tensor>("H");

    auto hdiff_tensor = ctx.Input<Tensor>(framework::GradVarName("H"));
    auto cdiff_tensor = ctx.Input<Tensor>(framework::GradVarName("C"));

    auto xdiff_tensor = ctx.Output<Tensor>(framework::GradVarName("X"));
    auto c_prev_diff_tensor =
        ctx.Output<Tensor>(framework::GradVarName("C_prev"));

    auto* X = x_tensor->data<T>();
    auto* C_prev = c_prev_tensor->data<T>();
    auto* C = c_tensor->data<T>();
    auto* H = h_tensor->data<T>();

    auto* H_diff = hdiff_tensor->data<T>();
    auto* C_diff = cdiff_tensor->data<T>();

    auto* C_prev_diff = c_prev_diff_tensor->mutable_data<T>(ctx.GetPlace());
    auto* X_diff = xdiff_tensor->mutable_data<T>(ctx.GetPlace());

    int N = c_tensor->dims()[0];
    int D = c_tensor->dims()[1];

    auto forget_bias = static_cast<T>(ctx.Attr<AttrType>("forget_bias"));

    for (int n = 0; n < N; ++n) {
      for (int d = 0; d < D; ++d) {
        T* c_prev_diff = C_prev_diff + d;
        T* i_diff = X_diff + d;
        T* f_diff = X_diff + 1 * D + d;
        T* o_diff = X_diff + 2 * D + d;
        T* g_diff = X_diff + 3 * D + d;

        const T i = sigmoid(X[d]);
        const T f = sigmoid(X[1 * D + d] + forget_bias);
        const T o = sigmoid(X[2 * D + d]);
        const T g = tanh(X[3 * D + d]);
        const T c_prev = C_prev[d];
        const T c = C[d];
        const T tanh_c = tanh(c);
        const T c_term_diff = C_diff[d] + H_diff[d] * o * (1 - tanh_c * tanh_c);
        *c_prev_diff = c_term_diff * f;
        *i_diff = c_term_diff * g * i * (1 - i);
        *f_diff = c_term_diff * c_prev * f * (1 - f);
        *o_diff = H_diff[d] * tanh_c * o * (1 - o);
        *g_diff = c_term_diff * i * (1 - g * g);
      }
      C_prev += D;
      X += 4 * D;
      C += D;
      H += D;
      C_diff += D;
      H_diff += D;
      X_diff += 4 * D;
      C_prev_diff += D;
    }
  }
};

}  // namespace operators
}  // namespace paddle