ContextProjectionOp.cpp 12.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* 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. */

15
#include "ContextProjectionOp.h"
16 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
#include "paddle/math/Matrix.h"
#include "paddle/math/Vector.h"

namespace paddle {

template <>
void ContextProjectionForward<DEVICE_TYPE_CPU>(Tensor& output,
                                               const Tensor& input,
                                               const Tensor& weight,
                                               const Tensor& sequence,
                                               size_t context_length,
                                               int context_start,
                                               size_t begin_pad,
                                               bool is_padding) {
  CHECK(output.getData() && input.getData() && sequence.getData());
  CHECK_EQ(output.dims_.size(), 2);
  CHECK_EQ(input.dims_.size(), 2);
  CHECK_EQ(weight.dims_.size(), 2);
  CHECK_EQ(sequence.dims_.size(), 1);

  auto out_mat = std::make_shared<CpuMatrix>(
      output.getData(), output.dims_[0], output.dims_[1]);
  const auto in_mat = std::make_shared<CpuMatrix>(
      input.getData(), input.dims_[0], input.dims_[1]);
  const auto weight_mat =
      !weight.getData()
          ? nullptr
          : std::make_shared<CpuMatrix>(
44
                weight.getData(), weight.dims_[0], weight.dims_[1]);
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
  CpuIVector seq_vec(sequence.dims_[0],
                     reinterpret_cast<int*>(sequence.getData()));
  CHECK_EQ(out_mat->getWidth(), in_mat->getWidth() * context_length);

  const int* starts = seq_vec.getData();
  const size_t num_sequences = seq_vec.getSize() - 1;
  for (size_t i = 0; i < num_sequences; ++i) {
    for (size_t j = 0; j < context_length; ++j) {
      int begin = starts[i] + context_start + j;
      int end = starts[i + 1] + context_start + j;
      int dst_begin = starts[i];
      int dst_end = starts[i + 1];
      if (begin < starts[i]) {
        int64_t pad_size =
            std::min(starts[i] - begin, starts[i + 1] - starts[i]);
        MatrixPtr mat = out_mat->subMatrix(starts[i], pad_size);
        if (is_padding && weight_mat) {
          MatrixPtr sub = weight_mat->subMatrix(j, pad_size);
          mat->addAtOffset(*sub, j * in_mat->getWidth());
        }
        dst_begin = starts[i] + pad_size;
        begin = starts[i];
      }
      if (end > starts[i + 1]) {
        int64_t pad_size =
            std::min(end - starts[i + 1], starts[i + 1] - starts[i]);
        MatrixPtr mat = out_mat->subMatrix(starts[i + 1] - pad_size, pad_size);
        if (is_padding && weight_mat) {
          MatrixPtr sub = weight_mat->subMatrix(
              begin_pad + context_start + j - pad_size, pad_size);
          mat->addAtOffset(*sub, j * in_mat->getWidth());
        }
        dst_end = starts[i + 1] - pad_size;
        end = starts[i + 1];
      }
      if (end <= begin) continue;
      MatrixPtr src = in_mat->subMatrix(begin, end - begin);
      MatrixPtr dst = out_mat->subMatrix(dst_begin, dst_end - dst_begin);
      dst->addAtOffset(*src, j * in_mat->getWidth());
    }
  }
}

/**
 * \param inputs[0] input value.
 * \param inputs[1] input weight.
 * \param inputs[2] input sequence.
 * \param outputs[0] output value.
 */
template <DeviceType Device>
class ContextProjectionForwardFunc : public FunctionBase {
public:
  void init(const FuncConfig& config) override {
    context_length_ = config.get<size_t>("context_length");
    context_start_ = config.get<int>("context_start");
    begin_pad_ = config.get<size_t>("begin_pad");
    is_padding_ = config.get<bool>("is_padding");
  }

  void calc(const Arguments& inputs,
            const Arguments& outputs,
            const Arguments& inouts) override {
    CHECK_EQ(3, inputs.size());
    CHECK_EQ(1, outputs.size());
    CHECK_EQ(0, inouts.size());

    ContextProjectionForward<Device>((Tensor&)outputs[0],
                                     inputs[0],
                                     inputs[1],
                                     inputs[2],
                                     context_length_,
                                     context_start_,
                                     begin_pad_,
                                     is_padding_);
  }

private:
  size_t context_length_;
  int context_start_;
  size_t begin_pad_;
  bool is_padding_;
};

128 129
template <>
void ContextProjectionBackward<DEVICE_TYPE_CPU>(Tensor& out_grad,
130 131
                                                Tensor& in_grad,
                                                Tensor& w_grad,
132 133 134 135
                                                const Tensor& sequence,
                                                size_t context_length,
                                                int context_start,
                                                size_t begin_pad,
136 137
                                                bool is_padding,
                                                size_t total_pad) {
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
  CHECK(out_grad.getData() && sequence.getData());
  CHECK_EQ(out_grad.dims_.size(), 2);
  CHECK_EQ(in_grad.dims_.size(), 2);
  CHECK_EQ(w_grad.dims_.size(), 2);
  CHECK_EQ(sequence.dims_.size(), 1);

  auto out_grad_mat = std::make_shared<CpuMatrix>(
      out_grad.getData(), out_grad.dims_[0], out_grad.dims_[1]);
  const auto in_grad_mat =
      !in_grad.getData()
          ? nullptr
          : std::make_shared<CpuMatrix>(
                in_grad.getData(), in_grad.dims_[0], in_grad.dims_[1]);
  const auto w_grad_mat =
      !w_grad.getData()
          ? nullptr
          : std::make_shared<CpuMatrix>(
                w_grad.getData(), w_grad.dims_[0], w_grad.dims_[1]);
  CpuIVector seq_vec(sequence.dims_[0],
                     reinterpret_cast<int*>(sequence.getData()));
  CHECK_EQ(out_grad_mat->getWidth(), in_grad_mat->getWidth() * context_length);

  size_t input_dim = in_grad_mat ? in_grad_mat->getWidth()
                                 : w_grad_mat ? w_grad_mat->getWidth() : 0;
  CHECK_EQ(out_grad_mat->getWidth(), input_dim * context_length);

  const int* starts = seq_vec.getData();
  size_t num_sequences = seq_vec.getSize() - 1;
  for (size_t i = 0; i < num_sequences; ++i) {
    for (size_t j = 0; j < context_length; ++j) {
      int begin = starts[i] + context_start + j;
      int end = starts[i + 1] + context_start + j;
      int dst_begin = starts[i];
      int dst_end = starts[i + 1];
      if (begin < starts[i]) {
        int64_t pad_size =
            std::min(starts[i] - begin, starts[i + 1] - starts[i]);
        if (is_padding && w_grad_mat) {
          MatrixPtr mat = out_grad_mat->subMatrix(starts[i], pad_size);
          MatrixPtr sub = w_grad_mat->subMatrix(j, pad_size);
          sub->addAtOffset(*mat, j * input_dim);
        }
        dst_begin = starts[i] + pad_size;
        begin = starts[i];
      }
      if (end > starts[i + 1]) {
        int64_t pad_size =
            std::min(end - starts[i + 1], starts[i + 1] - starts[i]);
        if (is_padding && w_grad_mat) {
          MatrixPtr mat =
              out_grad_mat->subMatrix(starts[i + 1] - pad_size, pad_size);
          MatrixPtr sub = w_grad_mat->subMatrix(
              begin_pad + context_start + j - pad_size, pad_size);
          sub->addAtOffset(*mat, j * input_dim);
        }
        dst_end = starts[i + 1] - pad_size;
        end = starts[i + 1];
      }
      if (end <= begin) continue;
      if (!in_grad_mat) continue;
      MatrixPtr src = in_grad_mat->subMatrix(begin, end - begin);
      MatrixPtr dst = out_grad_mat->subMatrix(dst_begin, dst_end - dst_begin);
      src->addAtOffset(*dst, j * input_dim);
    }
  }
}

/**
206 207
 * \param inputs[0] input grad.
 * \param inputs[1] weight grad.
208 209 210 211 212 213 214 215 216 217 218
 * \param inputs[2] input sequence.
 * \param outputs[0] output value.
 */
template <DeviceType Device>
class ContextProjectionBackwardFunc : public FunctionBase {
public:
  void init(const FuncConfig& config) override {
    context_length_ = config.get<size_t>("context_length");
    context_start_ = config.get<int>("context_start");
    begin_pad_ = config.get<size_t>("begin_pad");
    is_padding_ = config.get<bool>("is_padding");
219
    total_pad_ = config.get<size_t>("total_pad");
220 221 222 223 224 225 226 227 228 229
  }

  void calc(const Arguments& inputs,
            const Arguments& outputs,
            const Arguments& inouts) override {
    CHECK_EQ(3, inputs.size());
    CHECK_EQ(1, outputs.size());
    CHECK_EQ(0, inouts.size());

    ContextProjectionBackward<Device>((Tensor&)outputs[0],
230 231
                                      (Tensor&)inputs[0],
                                      (Tensor&)inputs[1],
232 233 234 235
                                      inputs[2],
                                      context_length_,
                                      context_start_,
                                      begin_pad_,
236 237
                                      is_padding_,
                                      total_pad_);
238 239 240 241 242 243 244
  }

private:
  size_t context_length_;
  int context_start_;
  size_t begin_pad_;
  bool is_padding_;
245
  size_t total_pad_;
246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317
};

/**
 * \param inputs[0] input grad.
 * \param inputs[1] input sequence.
 * \param outputs[0] output grad.
 */
template <DeviceType Device>
class ContextProjectionBackwardDataFunc : public FunctionBase {
public:
  void init(const FuncConfig& config) override {
    context_length_ = config.get<size_t>("context_length");
    context_start_ = config.get<int>("context_start");
  }

  void calc(const Arguments& inputs,
            const Arguments& outputs,
            const Arguments& inouts) override {
    CHECK_EQ(2, inputs.size());
    CHECK_EQ(1, outputs.size());
    CHECK_EQ(0, inouts.size());

    ContextProjectionBackwardData<Device>((Tensor&)outputs[0],
                                          (Tensor&)inputs[0],
                                          inputs[1],
                                          context_length_,
                                          context_start_);
  }

private:
  size_t context_length_;
  int context_start_;
};

/**
 * \param inputs[0] weight grad.
 * \param inputs[1] input sequence.
 * \param outputs[0] output grad.
 */
template <DeviceType Device>
class ContextProjectionBackwardWeightFunc : public FunctionBase {
public:
  void init(const FuncConfig& config) override {
    context_length_ = config.get<size_t>("context_length");
    context_start_ = config.get<int>("context_start");
    begin_pad_ = config.get<size_t>("begin_pad");
    total_pad_ = config.get<size_t>("total_pad");
  }

  void calc(const Arguments& inputs,
            const Arguments& outputs,
            const Arguments& inouts) override {
    CHECK_EQ(2, inputs.size());
    CHECK_EQ(1, outputs.size());
    CHECK_EQ(0, inouts.size());

    ContextProjectionBackwardWeight<Device>((Tensor&)outputs[0],
                                            (Tensor&)inputs[0],
                                            inputs[1],
                                            context_length_,
                                            context_start_,
                                            total_pad_,
                                            begin_pad_);
  }

private:
  size_t context_length_;
  int context_start_;
  size_t begin_pad_;
  size_t total_pad_;
};

318 319 320
REGISTER_TYPED_FUNC(ContextProjectionForward,
                    CPU,
                    ContextProjectionForwardFunc);
321 322 323
REGISTER_TYPED_FUNC(ContextProjectionBackward,
                    CPU,
                    ContextProjectionBackwardFunc);
324 325 326 327
#ifndef PADDLE_ONLY_CPU
REGISTER_TYPED_FUNC(ContextProjectionForward,
                    GPU,
                    ContextProjectionForwardFunc);
328 329 330
REGISTER_TYPED_FUNC(ContextProjectionBackward,
                    GPU,
                    ContextProjectionBackwardFunc);
331 332 333 334 335 336
REGISTER_TYPED_FUNC(ContextProjectionBackwardData,
                    GPU,
                    ContextProjectionBackwardDataFunc);
REGISTER_TYPED_FUNC(ContextProjectionBackwardWeight,
                    GPU,
                    ContextProjectionBackwardWeightFunc);
337 338
#endif
}  // namespace paddle