lookup_table_op.h 9.8 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
L
Luo Tao 已提交
2 3 4 5 6 7 8 9 10 11 12 13

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. */
14 15 16

#pragma once

17 18 19
#include <string>
#include <vector>

Y
Yi Wang 已提交
20 21 22
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_registry.h"
23
#include "paddle/fluid/framework/selected_rows_utils.h"
24
#include "paddle/phi/kernels/funcs/blas/blas.h"
25 26 27 28

namespace paddle {
namespace operators {

C
chengduoZH 已提交
29
using Tensor = framework::Tensor;
F
fengjiayi 已提交
30
using LoDTensor = framework::LoDTensor;
31
using SelectedRows = phi::SelectedRows;
32 33
using DDim = framework::DDim;

Q
qiaolongfei 已提交
34
constexpr int64_t kNoPadding = -1;
35 36

template <typename T>
Y
Yu Yang 已提交
37
class LookupTableKernel : public framework::OpKernel<T> {
38
 public:
39
  void Compute(const framework::ExecutionContext &context) const override {
40 41
    auto *ids_t = context.Input<LoDTensor>("Ids");      // int tensor
    auto *output_t = context.Output<LoDTensor>("Out");  // float tensor
42
    auto *table_var = context.InputVar("W");
43

H
hong 已提交
44 45 46
    auto id_name = context.InputNames("Ids").front();
    auto embedding_name = context.InputNames("W").front();
    auto out_name = context.OutputNames("Out").front();
Q
Qiao Longfei 已提交
47

48 49
    int64_t padding_idx = context.Attr<int64_t>("padding_idx");
    bool is_test = context.Attr<bool>("is_test");
Q
Qiao Longfei 已提交
50

51 52
    int64_t *ids = const_cast<int64_t *>(ids_t->data<int64_t>());
    int64_t ids_numel = ids_t->numel();
Q
Qiao Longfei 已提交
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
    if (table_var->IsType<LoDTensor>()) {
      auto *table_t = context.Input<LoDTensor>("W");
      int64_t row_number = table_t->dims()[0];
      int64_t row_width = table_t->dims()[1];

      auto *table = table_t->data<T>();
      auto *output = output_t->mutable_data<T>(context.GetPlace());

      for (int64_t i = 0; i < ids_numel; ++i) {
        if (padding_idx != kNoPadding && ids[i] == padding_idx) {
          memset(output + i * row_width, 0, row_width * sizeof(T));
        } else {
          PADDLE_ENFORCE_LT(
              ids[i], row_number,
              platform::errors::InvalidArgument(
                  "Variable value (input) of OP(fluid.layers.embedding) "
                  "expected >= 0 and < %ld, but got %ld. Please check input "
                  "value.",
                  row_number, ids[i]));
          PADDLE_ENFORCE_GE(
              ids[i], 0,
              platform::errors::InvalidArgument(
                  "Variable value (input) of OP(fluid.layers.embedding) "
                  "expected >= 0 and < %ld, but got %ld. Please check input "
                  "value.",
                  row_number, ids[i]));
          memcpy(output + i * row_width, table + ids[i] * row_width,
                 row_width * sizeof(T));
82
        }
83 84
      }

85 86
    } else if (table_var->IsType<phi::SelectedRows>()) {
      const auto &table_t = table_var->Get<phi::SelectedRows>();
87 88 89
      int64_t row_width = table_t.value().dims()[1];
      const auto *table = table_t.value().data<T>();
      auto *output = output_t->mutable_data<T>(context.GetPlace());
90 91
      auto input_data_type =
          framework::TransToProtoVarType(table_t.value().dtype());
92 93 94 95 96 97 98 99 100 101 102 103 104 105
      for (int64_t i = 0; i < ids_numel; ++i) {
        if (padding_idx != kNoPadding && ids[i] == padding_idx) {
          memset(output + i * row_width, 0, row_width * sizeof(T));
        } else {
          PADDLE_ENFORCE_GE(
              ids[i], 0,
              platform::errors::InvalidArgument(
                  "Variable value (input) of OP(fluid.layers.embedding) "
                  "expected >= 0. But received %ld",
                  ids[i]));
          if (is_test) {
            auto id_index = table_t.GetIndexFromId(ids[i]);

            if (id_index != -1) {
106
              if (input_data_type == framework::proto::VarType::INT8 ||
107
                  input_data_type == framework::proto::VarType::INT16 ||
108
                  input_data_type == framework::proto::VarType::BF16) {
109 110 111
                memcpy(output + i * row_width, table + id_index * row_width,
                       row_width * sizeof(T));
              } else {
112 113
                auto blas =
                    phi::funcs::GetBlas<platform::CPUDeviceContext, T>(context);
114 115 116 117 118 119
                blas.VCOPY(row_width, table + id_index * row_width,
                           output + i * row_width);
              }
            } else {
              memset(output + i * row_width, 0, row_width * sizeof(T));
            }
Q
Qiao Longfei 已提交
120
          } else {
121
            auto id_index = table_t.Index(ids[i]);
122 123
            PADDLE_ENFORCE_GE(
                ids[i], 0,
124 125 126 127
                platform::errors::InvalidArgument(
                    "Variable value (input) of OP(fluid.layers.embedding) "
                    "expected >= 0. But received %ld",
                    ids[i]));
128
            PADDLE_ENFORCE_GE(
129 130 131 132
                id_index, 0,
                platform::errors::InvalidArgument(
                    "the input key should be exists. But received %d.",
                    id_index));
133

134
            if (input_data_type == framework::proto::VarType::INT8 ||
135
                input_data_type == framework::proto::VarType::INT16 ||
136
                input_data_type == framework::proto::VarType::BF16) {
137 138 139
              memcpy(output + i * row_width, table + id_index * row_width,
                     row_width * sizeof(T));
            } else {
140
              auto blas =
141
                  phi::funcs::GetBlas<platform::CPUDeviceContext, T>(context);
142 143 144
              blas.VCOPY(row_width, table + id_index * row_width,
                         output + i * row_width);
            }
Q
Qiao Longfei 已提交
145
          }
146 147
        }
      }
148 149 150 151 152
    }
  }
};

template <typename T>
Y
Yu Yang 已提交
153
class LookupTableGradKernel : public framework::OpKernel<T> {
154
 public:
155
  void Compute(const framework::ExecutionContext &context) const override {
Q
qiaolongfei 已提交
156 157 158 159
    auto *table_var = context.InputVar("W");
    DDim table_dim;
    if (table_var->IsType<LoDTensor>()) {
      table_dim = context.Input<LoDTensor>("W")->dims();
160 161
    } else if (table_var->IsType<phi::SelectedRows>()) {
      auto *table_t = context.Input<phi::SelectedRows>("W");
Q
qiaolongfei 已提交
162 163
      table_dim = table_t->value().dims();
    } else {
164
      PADDLE_THROW(platform::errors::InvalidArgument(
Q
qiaolongfei 已提交
165
          "The parameter W of a LookupTable "
166
          "must be either LoDTensor or SelectedRows"));
Q
qiaolongfei 已提交
167 168
    }

169
    int64_t padding_idx = context.Attr<int64_t>("padding_idx");
170
    bool is_sparse = context.Attr<bool>("is_sparse");
171 172
    // Since paddings are not trainable and fixed in forward, the gradient of
    // paddings makes no sense and we don't deal with it in backward.
173
    if (is_sparse) {
174 175
      auto *ids = context.Input<LoDTensor>("Ids");
      auto *d_output = context.Input<LoDTensor>(framework::GradVarName("Out"));
176
      auto *d_table =
177
          context.Output<phi::SelectedRows>(framework::GradVarName("W"));
178

179
      auto *ids_data = ids->data<int64_t>();
180
      int64_t ids_num = ids->numel();
181

M
minqiyang 已提交
182
      std::vector<int64_t> new_rows;
M
minqiyang 已提交
183 184
      new_rows.resize(ids_num);
      std::memcpy(&new_rows[0], ids_data, ids_num * sizeof(int64_t));
185
      d_table->set_rows(new_rows);
186

187
      auto *d_table_value = d_table->mutable_value();
188
      d_table_value->Resize({ids_num, table_dim[1]});
189 190 191 192 193 194 195 196
      d_table_value->mutable_data<T>(context.GetPlace());
      d_table->set_height(table_dim[0]);

      auto *d_output_data = d_output->data<T>();
      auto *d_table_data = d_table_value->data<T>();

      auto d_output_dims = d_output->dims();
      auto d_output_dims_2d =
197
          phi::flatten_to_2d(d_output_dims, d_output_dims.size() - 1);
198 199 200 201 202 203 204 205
      PADDLE_ENFORCE_EQ(d_table_value->dims(), d_output_dims_2d,
                        platform::errors::InvalidArgument(
                            "ShapeError: The shape of lookup_table@Grad and "
                            "output@Grad should be same. "
                            "But received lookup_table@Grad's shape = [%s], "
                            "output@Grad's shape = [%s].",
                            d_table_value->dims(), d_output_dims_2d));
      memcpy(d_table_data, d_output_data, sizeof(T) * d_output->numel());
206
    } else {
207 208 209
      auto *ids = context.Input<LoDTensor>("Ids");
      auto *d_output = context.Input<LoDTensor>(framework::GradVarName("Out"));
      auto *d_table = context.Output<LoDTensor>(framework::GradVarName("W"));
210

211
      auto *ids_data = ids->data<int64_t>();
212

213 214
      int64_t N = table_dim[0];
      int64_t D = table_dim[1];
215

216 217
      auto *d_output_data = d_output->data<T>();
      auto *d_table_data = d_table->mutable_data<T>(context.GetPlace());
218

219 220
      memset(d_table_data, 0, d_table->numel() * sizeof(T));

221
      for (int64_t i = 0; i < ids->numel(); ++i) {
Q
Qiao Longfei 已提交
222 223 224 225
        if (padding_idx != kNoPadding && ids_data[i] == padding_idx) {
          // the gradient of padding_idx should be 0, already done by memset, so
          // do nothing.
        } else {
226 227
          PADDLE_ENFORCE_LT(
              ids_data[i], N,
228 229 230 231 232
              platform::errors::InvalidArgument(
                  "Variable value (input) of OP(fluid.layers.embedding) "
                  "expected >= 0 and < %ld, but got %ld. Please check input "
                  "value.",
                  N, ids_data[i]));
233 234
          PADDLE_ENFORCE_GE(
              ids_data[i], 0,
235 236 237 238 239
              platform::errors::InvalidArgument(
                  "Variable value (input) of OP(fluid.layers.embedding) "
                  "expected >= 0 and < %ld, but got %ld. Please check input"
                  "value.",
                  N, ids_data[i]));
240 241 242
          for (int j = 0; j < D; ++j) {
            d_table_data[ids_data[i] * D + j] += d_output_data[i * D + j];
          }
243
        }
244 245 246 247 248 249 250
      }
    }
  }
};

}  // namespace operators
}  // namespace paddle