hierarchical_sigmoid_op.h 10.1 KB
Newer Older
Y
Yancey1989 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
W
weixing02 已提交
16
#include <iostream>
17
#include <iterator>
J
JiabinYang 已提交
18
#include <set>
19
#include <string>
W
weixing02 已提交
20
#include <vector>
J
JiabinYang 已提交
21
#include "paddle/fluid/framework/mixed_vector.h"
W
weixing02 已提交
22 23
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/clip_op.h"
J
JiabinYang 已提交
24
#include "paddle/fluid/operators/detail/safe_ref.h"
W
weixing02 已提交
25 26 27
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/math/matrix_bit_code.h"
#include "paddle/fluid/platform/transform.h"
J
JiabinYang 已提交
28

29 30 31 32
#ifdef PADDLE_WITH_DISTRIBUTE
#include "paddle/fluid/operators/distributed/parameter_prefetch.h"
#endif

Y
Yancey1989 已提交
33 34 35
namespace paddle {
namespace operators {

Y
Yancey1989 已提交
36 37 38
template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
Y
Yancey1989 已提交
39
using platform::Transform;
Y
Yancey1989 已提交
40

J
JiabinYang 已提交
41 42 43 44 45 46
static std::vector<int64_t> PathToRows(const framework::LoDTensor& path) {
  std::set<int64_t> rows;
  for (int64_t i = 0; i < path.numel(); ++i) {
    int64_t row = path.data<int64_t>()[i];
    if (row < 0) {
      continue;
J
JiabinYang 已提交
47
    }
J
JiabinYang 已提交
48
    rows.emplace(row);
J
JiabinYang 已提交
49
  }
J
JiabinYang 已提交
50
  return std::vector<int64_t>(rows.begin(), rows.end());
J
JiabinYang 已提交
51
}
Y
Yancey1989 已提交
52
template <typename DeviceContext, typename T>
Y
Yancey1989 已提交
53 54
class HierarchicalSigmoidOpKernel : public framework::OpKernel<T> {
 public:
Y
Yancey1989 已提交
55
  void Compute(const framework::ExecutionContext& ctx) const override {
56 57
    auto& in = detail::Ref(ctx.Input<framework::LoDTensor>("X"));
    auto& w = detail::Ref(ctx.Input<framework::LoDTensor>("W"));
58
    auto* path = ctx.Input<framework::LoDTensor>("PathTable");
J
JiabinYang 已提交
59
    auto* code = ctx.Input<framework::LoDTensor>("PathCode");
60
    auto& label = detail::Ref(ctx.Input<framework::LoDTensor>("Label"));
J
JiabinYang 已提交
61 62 63
    auto* bias = ctx.Input<framework::LoDTensor>("Bias");
    auto* out = ctx.Output<framework::LoDTensor>("Out");
    auto* pre_out = ctx.Output<framework::LoDTensor>("PreOut");
Y
Yancey1989 已提交
64
    size_t num_classes = static_cast<size_t>(ctx.Attr<int>("num_classes"));
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
    // for remote prefetch

    auto epmap = ctx.Attr<std::vector<std::string>>("epmap");
    if (!epmap.empty()) {
      // if epmap is not empty, then the parameter will be fetched from remote
      // parameter
      // server
      auto height_sections = ctx.Attr<std::vector<int>>("height_sections");
      auto table_names = ctx.Attr<std::vector<std::string>>("table_names");
      VLOG(3) << "path type is " << path->type().name();
      std::vector<int64_t> real_rows = PathToRows(*path);
      framework::Scope& local_scope = ctx.scope().NewScope();
      auto* ids = local_scope.Var("Ids@Prefetch");
      auto* x_tensor = ids->GetMutable<framework::LoDTensor>();

      x_tensor->mutable_data<int64_t>(
          framework::make_ddim({static_cast<int64_t>(real_rows.size()), 1}),
          ctx.GetPlace());
      // copy.

      std::memcpy(x_tensor->data<int64_t>(), real_rows.data(),
                  real_rows.size() * sizeof(int64_t));

      framework::DDim w_dims = ctx.Input<Tensor>("W")->dims();
      w_dims[0] = x_tensor->dims()[0];
      auto* w_tensor =
          local_scope.Var("W@Prefetch")->GetMutable<framework::LoDTensor>();
      w_tensor->Resize(w_dims);

#ifdef PADDLE_WITH_DISTRIBUTE
      // w_Out is set to used by prefetch, never change it in other cases
      auto* w_out = ctx.Output<framework::LoDTensor>("W_Out");
      operators::distributed::prefetch_with_reconstruct<T>(
          "Ids@Prefetch", "W@Prefetch", table_names, epmap, height_sections,
          ctx, local_scope, w_out);
#else
      PADDLE_THROW(
          "paddle is not compiled with distribute support, can not do "
          "parameter prefetch!");
#endif
    }

107 108 109 110 111 112
    bool is_custom = false;
    if (path) {
      is_custom = true;
    }
    int64_t code_length =
        path ? path->dims()[1] : math::FindLastSet(num_classes - 1);
J
JiabinYang 已提交
113
    int64_t batch_size = in.dims()[0];
J
JiabinYang 已提交
114
    framework::LoDTensor sum;
W
weixing02 已提交
115
    auto& dev_ctx = ctx.template device_context<DeviceContext>();
G
guosheng 已提交
116
    auto* pre_out_data = pre_out->mutable_data<T>(
Y
Yancey1989 已提交
117
        framework::make_ddim({batch_size, code_length}), ctx.GetPlace());
W
weixing02 已提交
118
    auto pre_out_mat = EigenMatrix<T>::From(*pre_out);
G
guosheng 已提交
119 120
    // Not all class(leaf) nodes' path lengths equal code_length, thus init as
    // 0s can avoid out of path's loss.
121
    math::SetConstant<DeviceContext, T> zero;
W
weixing02 已提交
122
    zero(dev_ctx, pre_out, static_cast<T>(0.0));
Y
Yancey1989 已提交
123 124
    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
    math::RowwiseSum<DeviceContext, T> row_sum;
125 126 127 128

    std::unique_ptr<math::MatrixBitCodeFunctor<T>> bit_code;
    if (!is_custom) {
      bit_code.reset(new math::MatrixBitCodeFunctor<T>(num_classes,
J
JiabinYang 已提交
129
                                                       label.data<int64_t>()));
130
    } else {
J
JiabinYang 已提交
131 132
      bit_code.reset(new math::MatrixBitCodeFunctor<T>(*path, *code,
                                                       label.data<int64_t>()));
133
    }
Y
Yancey1989 已提交
134

Y
Yancey1989 已提交
135 136
    std::vector<int64_t> sum_dims({batch_size, 1UL});
    sum.mutable_data<T>(framework::make_ddim(sum_dims), ctx.GetPlace());
Y
Yancey1989 已提交
137
    auto sum_mat = EigenMatrix<T>::From(sum);
Y
Yancey1989 已提交
138
    out->mutable_data<T>(ctx.GetPlace());
Y
Yancey1989 已提交
139
    auto out_mat = framework::EigenVector<T>::Flatten(*out);
Y
Yancey1989 已提交
140
    if (bias) {
141
      bit_code->Add(*bias, pre_out);
Y
Yancey1989 已提交
142
    }
J
JiabinYang 已提交
143
    bit_code->Mul(pre_out, w, in);
G
guosheng 已提交
144
    // clip to [-40, 40]
Y
Yancey1989 已提交
145 146
    Transform<DeviceContext> trans;
    trans(ctx.template device_context<DeviceContext>(), pre_out_data,
W
weixing02 已提交
147
          pre_out_data + pre_out->numel(), pre_out_data,
Y
Yancey1989 已提交
148
          ClipFunctor<T>(static_cast<T>(-40.0), static_cast<T>(40.0)));
149
    bit_code->Sum(*pre_out, out, static_cast<T>(-1));
G
guosheng 已提交
150
    // use softrelu to calculate cross entropy
Y
Yancey1989 已提交
151
    pre_out_mat.device(place) = (static_cast<T>(1.0) + pre_out_mat.exp()).log();
W
weixing02 已提交
152
    row_sum(dev_ctx, *pre_out, &sum);
153 154 155 156
    // TODO(guosheng): Subtract the out of path's loss, since not all
    // class(leaf) nodes' path lengths equal code_length. But it won't break the
    // gradient check since both have the out of path's loss and will cancel out
    // each other.
Y
Yancey1989 已提交
157
    out_mat.device(place) = sum_mat + out_mat;
Y
Yancey1989 已提交
158
  }
Y
Yancey1989 已提交
159 160
};

Y
Yancey1989 已提交
161
template <typename DeviceContext, typename T>
Y
Yancey1989 已提交
162 163
class HierarchicalSigmoidGradOpKernel : public framework::OpKernel<T> {
 public:
Y
Yancey1989 已提交
164
  void Compute(const framework::ExecutionContext& ctx) const override {
165 166
    auto& in = detail::Ref(ctx.Input<framework::LoDTensor>("X"));
    auto& w = detail::Ref(ctx.Input<framework::LoDTensor>("W"));
167
    auto* path = ctx.Input<framework::LoDTensor>("PathTable");
J
JiabinYang 已提交
168
    auto* code = ctx.Input<framework::LoDTensor>("PathCode");
J
JiabinYang 已提交
169 170 171 172 173
    auto* in_grad =
        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
    bool is_sparse = ctx.Attr<bool>("is_sparse");
    auto& dev_ctx = ctx.template device_context<DeviceContext>();
    math::SetConstant<DeviceContext, T> zero;
174 175 176
    auto& label = detail::Ref(ctx.Input<framework::LoDTensor>("Label"));
    auto& pre_out = detail::Ref(ctx.Input<framework::LoDTensor>("PreOut"));
    auto& out_grad = detail::Ref(
J
JiabinYang 已提交
177
        ctx.Input<framework::LoDTensor>(framework::GradVarName("Out")));
J
JiabinYang 已提交
178
    framework::LoDTensor pre_out_grad;
179

J
JiabinYang 已提交
180
    pre_out_grad.mutable_data<T>(pre_out.dims(), ctx.GetPlace());
181 182
    in_grad->mutable_data<T>(ctx.GetPlace());
    zero(dev_ctx, in_grad, static_cast<T>(0.0));
W
weixing02 已提交
183

Y
Yancey1989 已提交
184
    size_t num_classes = static_cast<size_t>(ctx.Attr<int>("num_classes"));
185 186 187 188 189 190 191 192 193

    bool is_custom = false;
    if (path) {
      is_custom = true;
    }

    std::unique_ptr<math::MatrixBitCodeFunctor<T>> bit_code;
    if (!is_custom) {
      bit_code.reset(new math::MatrixBitCodeFunctor<T>(num_classes,
J
JiabinYang 已提交
194
                                                       label.data<int64_t>()));
195
    } else {
J
JiabinYang 已提交
196 197
      bit_code.reset(new math::MatrixBitCodeFunctor<T>(*path, *code,
                                                       label.data<int64_t>()));
198
    }
199

Y
Yancey1989 已提交
200
    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
J
JiabinYang 已提交
201
    auto pre_out_mat = EigenMatrix<T>::From(pre_out);
W
weixing02 已提交
202
    auto pre_out_grad_mat = EigenMatrix<T>::From(pre_out_grad);
J
JiabinYang 已提交
203
    auto out_grad_mat = EigenMatrix<T>::From(out_grad);
J
JiabinYang 已提交
204

J
JiabinYang 已提交
205
    Eigen::array<int, 2> bcast{1, static_cast<int>(pre_out_grad.dims()[1])};
206 207 208 209

    // softrelu derivative
    pre_out_grad_mat.device(place) =
        static_cast<T>(1.0) - static_cast<T>(1.0) / pre_out_mat.exp();
210
    bit_code->Sub(&pre_out_grad);  // the gradient of clip(w * x + b)
W
weixing02 已提交
211
    pre_out_grad_mat.device(place) =
212
        pre_out_grad_mat * out_grad_mat.broadcast(bcast);
G
guosheng 已提交
213 214
    // TODO(guosheng): multiply pre_out_grad with subgradient of clipping to
    // be consistent with the clipping in forward.
215 216 217 218 219 220 221
    auto* bias_grad =
        ctx.Output<framework::LoDTensor>(framework::GradVarName("Bias"));
    if (bias_grad) {
      bias_grad->mutable_data<T>(ctx.GetPlace());
      zero(dev_ctx, bias_grad, static_cast<T>(0.0));
      bit_code->AddGrad(pre_out_grad, bias_grad);
    }
J
JiabinYang 已提交
222 223 224 225 226
    if (!is_sparse) {
      auto* w_grad =
          ctx.Output<framework::LoDTensor>(framework::GradVarName("W"));
      w_grad->mutable_data<T>(ctx.GetPlace());
      zero(dev_ctx, w_grad, static_cast<T>(0.0));
J
JiabinYang 已提交
227
      bit_code->MulGradWeight(pre_out_grad, w_grad, in);
J
JiabinYang 已提交
228
    } else {
J
JiabinYang 已提交
229
      framework::Vector<int64_t> real_rows = PathToRows(*path);
J
JiabinYang 已提交
230 231 232
      auto* w_grad =
          ctx.Output<framework::SelectedRows>(framework::GradVarName("W"));
      w_grad->set_rows(real_rows);
233
      // Build a map of id -> row_index to speed up finding the index of one id
J
JiabinYang 已提交
234
      w_grad->set_height(w.dims()[0]);
J
JiabinYang 已提交
235
      auto* w_grad_value = w_grad->mutable_value();
J
JiabinYang 已提交
236
      framework::DDim temp_dim(w.dims());
J
JiabinYang 已提交
237 238 239 240
      set(temp_dim, 0, real_rows.size());

      w_grad_value->mutable_data<T>(temp_dim, ctx.GetPlace());
      zero(dev_ctx, w_grad_value, static_cast<T>(0.0));
J
JiabinYang 已提交
241
      bit_code->MulGradWeight(pre_out_grad, w_grad, in);
J
JiabinYang 已提交
242
    }
J
JiabinYang 已提交
243
    bit_code->MulGradError(pre_out_grad, w, in_grad);
Y
Yancey1989 已提交
244
  }
Y
Yancey1989 已提交
245 246 247 248
};

}  // namespace operators
}  // namespace paddle