pyramid_hash_op.cc 16.5 KB
Newer Older
A
Aurelius84 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 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 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338
/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.

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. */

#include <xxhash.h>
#include <algorithm>
#include <cmath>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/search_compute.h"

extern "C" {
#include "math/bloomfilter.h"
// void* memcpy1(void* dst, void* src, uint32_t length);
}

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
using LoD = framework::LoD;

class PyramidHashOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  void Make() override {
    AddInput("X",
             "X (Tensor, MUST be Tensor<!!!_int32_!!!>) Input variable which "
             "should contain lod information.");
    AddInput("W", "W (Tensor)");
    AddInput("WhiteList", "WhiteList (Tensor)");
    AddInput("BlackList", "BlackList (Tensor)");
    AddAttr<int>("num_emb", "num_emb").SetDefault(0).EqualGreaterThan(0);
    AddAttr<int>("space_len", "space_len").SetDefault(0).EqualGreaterThan(0);
    AddAttr<int>("pyramid_layer", "pyramid_layer (must be >= 2)")
        .SetDefault(2)
        .EqualGreaterThan(2);
    AddAttr<int>("rand_len", "rand_len").SetDefault(0).EqualGreaterThan(0);
    AddAttr<float>("drop_out_percent", "drop_out_percent")
        .SetDefault(0)
        .EqualGreaterThan(0);
    AddAttr<int>("is_training", "is_training")
        .SetDefault(0)
        .EqualGreaterThan(0);
    AddAttr<bool>("use_filter", "use_filter").SetDefault(true);
    AddAttr<int>("white_list_len", "white_list_len")
        .SetDefault(0)
        .EqualGreaterThan(0);
    AddAttr<int>("black_list_len", "black_list_len")
        .SetDefault(0)
        .EqualGreaterThan(0);
    AddAttr<int>("seed", "seed").SetDefault(0).EqualGreaterThan(0);
    AddAttr<float>("lr", "learning rate").SetDefault(0.0).EqualGreaterThan(0.0);

    AddOutput("Out", "Out (Tensor, default Tensor<float>) Output variable");
    AddOutput("DropPos", "Out (Tensor, Tensor<int>) Output variable");
    AddOutput("X_Temp_Out", "Out (Tensor, Tensor<int>) Output variable")
        .AsIntermediate();

    AddComment(R"DOC(
      PyramidHash

      NOTE: only support 'float32' data type now.

    )DOC");
  }
};

class PyramidHashOP : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE_EQ(ctx->HasInput("X"), true, "X(Input) should not be null.");
    PADDLE_ENFORCE_EQ(ctx->HasInput("W"), true, "W(Input) should not be null.");
    PADDLE_ENFORCE_EQ(ctx->HasOutput("Out"), true,
                      "Out(Output) should not be null.");
    PADDLE_ENFORCE_EQ(ctx->HasOutput("DropPos"), true,
                      "DropPos(TMP Output) should not be null.");

    auto x_dims = ctx->GetInputDim("X");
    PADDLE_ENFORCE_EQ(x_dims.size(), 2, "The rank of X(Input) should be 2.");

    auto w_dims = ctx->GetInputDim("W");
    PADDLE_ENFORCE_EQ(w_dims.size(), 2, "W should be 2-D tensor");

    int space_len = ctx->Attrs().Get<int>("space_len");
    int rand_len = ctx->Attrs().Get<int>("rand_len");

    PADDLE_ENFORCE_EQ(w_dims[0], space_len + rand_len,
                      "w_dims[0] should be equal to (space_len + rand_len)");
    PADDLE_ENFORCE_EQ(w_dims[1], 1, "w_dims[1] should be equal to 1");

    int num_emb = ctx->Attrs().Get<int>("num_emb");
    PADDLE_ENFORCE_EQ(num_emb % rand_len, 0,
                      "random length should mod embedding size");

    int white_list_len = ctx->Attrs().Get<int>("white_list_len");
    if (white_list_len > 0) {
      PADDLE_ENFORCE_EQ(
          ctx->HasInput("WhiteList"), true,
          "WhiteList(Input) should not be null when white_list_len > 0");
      auto wl_dims = ctx->GetInputDim("WhiteList");
      PADDLE_ENFORCE_EQ(wl_dims.size(), 2, "WhiteList should be 2-D tensor");
      PADDLE_ENFORCE_EQ(wl_dims[0], white_list_len,
                        "wl_dims[0] should be equal to white_list_len");
      PADDLE_ENFORCE_EQ(wl_dims[1], 1, "wl_dims[1] should be equal to 1");
    }

    int black_list_len = ctx->Attrs().Get<int>("black_list_len");
    if (black_list_len > 0) {
      PADDLE_ENFORCE_EQ(
          ctx->HasInput("BlackList"), true,
          "BlackList(Input) should not be null when black_list_len > 0");
      auto bl_dims = ctx->GetInputDim("BlackList");
      PADDLE_ENFORCE_EQ(bl_dims.size(), 2, "BlackList should be 2-D tensor");
      PADDLE_ENFORCE_EQ(bl_dims[0], black_list_len,
                        "bl_dims[0] should be equal to black_list_len");
      PADDLE_ENFORCE_EQ(bl_dims[1], 1, "bl_dims[1] should be equal to 1");
    }

    if (ctx->IsRuntime()) {
      // something to do in runtime.
    } else {
      // compile time
      ctx->SetOutputDim("Out", framework::make_ddim({-1, num_emb}));
      ctx->SetOutputDim("X_Temp_Out", x_dims);
      ctx->ShareLoD("X", /*->*/ "Out");
    }
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(
        OperatorWithKernel::IndicateVarDataType(ctx, "W"), ctx.GetPlace());
  }
};

template <typename DeviceContext, typename T>
class CPUPyramidHashOPKernel : public framework::OpKernel<T> {
 public:
  bool should_use_term(math::bloomfilter* _filter,
                       math::bloomfilter* _black_filter, const T* word_repr,
                       int len) const {
    return (!_filter ||
            1 == math::bloomfilter_get(_filter, word_repr, len * sizeof(T))) &&
           (!_black_filter ||
            0 == math::bloomfilter_get(_black_filter, word_repr,
                                       len * sizeof(T)));
  }

  void hash_embedding_ff(const T* hash_id, int len, T* top_pos,
                         const T* weights, int _num_emb, int _rand_len,
                         int _space_len) const {
    for (unsigned int j = 0; j != _num_emb; j += _rand_len) {
      unsigned int pos = XXH32(hash_id, len * sizeof(T), j) % _space_len;
      memcpy(top_pos + j, const_cast<float*>(weights + pos),
             _rand_len * sizeof(T));
    }
  }

  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* bottom = ctx.Input<LoDTensor>("X");
    auto* _blobs_0 = ctx.Input<Tensor>("W");
    auto* _blobs_1 = ctx.Input<Tensor>("WhiteList");
    auto* _blobs_2 = ctx.Input<Tensor>("BlackList");
    auto* top = ctx.Output<LoDTensor>("Out");
    auto* drop_pos = ctx.Output<LoDTensor>("DropPos");

    int _num_emb = ctx.Attr<int>("num_emb");
    bool use_filter = ctx.Attr<bool>("use_filter");
    int white_list_len = ctx.Attr<int>("white_list_len");
    int black_list_len = ctx.Attr<int>("black_list_len");
    int _pyramid_layer = ctx.Attr<int>("pyramid_layer");
    int _is_training = ctx.Attr<int>("is_training");
    int seed = ctx.Attr<int>("seed");
    unsigned int _seed = (unsigned int)seed;
    int _rand_len = ctx.Attr<int>("rand_len");
    int _space_len = ctx.Attr<int>("space_len");
    float _drop_out_percent = ctx.Attr<float>("drop_out_percent");

    const auto& offset = bottom->lod()[0];
    const auto* bottom_data_ori = bottom->data<int32_t>();
    auto* buff = ctx.Output<LoDTensor>("X_Temp_Out");
    buff->Resize(framework::make_ddim({bottom->dims()[0], bottom->dims()[1]}));
    T* bottom_data = buff->mutable_data<T>(ctx.GetPlace());
    for (size_t i = 0; i < bottom->dims()[0]; i++) {
      bottom_data[i] = bottom_data_ori[i];
    }

    const auto* weights = _blobs_0->data<T>();

    std::vector<size_t> top_offset;
    top_offset.resize(offset.size());
    top_offset[0] = 0;

    math::bloomfilter* _filter = NULL;
    math::bloomfilter* _black_filter = NULL;
    if (use_filter) {
      if (white_list_len != 0) {
        _filter = (math::bloomfilter*)_blobs_1->data<T>();
        PADDLE_ENFORCE_EQ(math::bloomfilter_check(_filter), 1,
                          "white filter not load");
      }
      if (black_list_len != 0) {
        _black_filter = (math::bloomfilter*)_blobs_2->data<T>();
        PADDLE_ENFORCE_EQ(math::bloomfilter_check(_black_filter), 1,
                          "black filter not load");
      }
    }

    drop_pos->Resize(framework::make_ddim(
        {bottom->dims()[0] * bottom->dims()[1] * _pyramid_layer, 1}));
    std::vector<size_t> drop_pos_offset;
    drop_pos_offset.resize(offset.size());
    drop_pos_offset[0] = 0;
    int* iter = drop_pos->mutable_data<int>(ctx.GetPlace());
    int* iter_end = iter;

    for (int i = 0; i < top_offset.size() - 1; ++i) {
      int w = offset[i + 1] - offset[i];
      int nsentense_with_pyramid = 0;
      if (w < 2) {
        nsentense_with_pyramid = 0;
      } else {
        for (int ilayer = 1; ilayer < _pyramid_layer && ilayer < w; ++ilayer) {
          for (int l = 0; l < w - ilayer; ++l) {
            if (should_use_term(_filter, _black_filter,
                                (const T*)(bottom_data + offset[i] + l),
                                ilayer + 1)) {
              if (_is_training != 0) {
                unsigned int rand_val = rand_r(&_seed);
                T rate = static_cast<T>(rand_val) / (RAND_MAX);
                *(iter_end++) = (rate < _drop_out_percent ? 0 : 1);
              } else {
                *(iter_end++) = 1;
              }
            } else {
              *(iter_end++) = 0;
            }
          }
        }
        nsentense_with_pyramid = std::count(iter, iter_end, 1);
        iter = iter_end;
      }
      drop_pos_offset[i + 1] = drop_pos_offset[i] + nsentense_with_pyramid;
      top_offset[i + 1] =
          top_offset[i] +
          (nsentense_with_pyramid == 0 ? 1 : nsentense_with_pyramid);
    }

    int top_l = top_offset[top_offset.size() - 1];

    framework::LoD top_lod;
    top_lod.push_back(top_offset);
    top->set_lod(top_lod);
    top->Resize(framework::make_ddim({top_l, _num_emb}));
    auto* top_data = top->mutable_data<T>(ctx.GetPlace());

    framework::LoD drop_pos_lod;
    drop_pos_lod.push_back(drop_pos_offset);
    drop_pos->set_lod(drop_pos_lod);

    iter = drop_pos->mutable_data<int>(ctx.GetPlace());
    int top_counter = 0;
    for (int i = 0; i < offset.size() - 1; ++i) {
      int w_drop = drop_pos_offset[i + 1] - drop_pos_offset[i];
      int w = offset[i + 1] - offset[i];
      if (w_drop == 0) {
        if (w >= 2) {
          for (int ilayer = 1; ilayer < _pyramid_layer && ilayer < w;
               ++ilayer) {
            for (int l = 0; l < w - ilayer; ++l) {
              iter++;
            }
          }
        }
        auto* top_pos = top_data + top_counter++ * _num_emb;
        memset(top_pos, 0, _num_emb * sizeof(T));
        continue;
      }
      if (w >= 2) {
        for (int ilayer = 1; ilayer < _pyramid_layer && ilayer < w; ++ilayer) {
          for (int l = 0; l < w - ilayer; ++l) {
            if (*(iter++) == 0) {
              // do nothing
            } else {
              auto* top_pos = top_data + top_counter++ * _num_emb;
              hash_embedding_ff((const T*)(bottom_data + offset[i] + l),
                                ilayer + 1, top_pos, weights, _num_emb,
                                _rand_len, _space_len);
            }
          }
        }
      }
    }
    if (iter != iter_end) {
      exit(1);
    }
    if (_is_training == 0) {
      avx_axpy_noadd(top_data, top_data, top->dims()[0] * top->dims()[1],
                     _drop_out_percent);
    }
  }
};

class PyramidHashOpGrad : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE_EQ(ctx->HasInput("X"), true, "Input(X) should not be null.");
    PADDLE_ENFORCE_EQ(ctx->HasInput("W"), true, "Input(W) should not be null.");
    PADDLE_ENFORCE_EQ(ctx->HasInput("DropPos"), true,
                      "Input(DropPos) should not be null.");
    PADDLE_ENFORCE_EQ(
        ctx->HasInput(framework::GradVarName("Out")), true,
        "Input(Out@GRAD) of PyramidHashGradOp should not be null.");
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(
        OperatorWithKernel::IndicateVarDataType(ctx, "W"), ctx.GetPlace());
  }
};

H
hong 已提交
339 340
template <typename T>
class PyramidHashGradOpMaker : public framework::SingleGradOpMaker<T> {
A
Aurelius84 已提交
341
 public:
H
hong 已提交
342
  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
A
Aurelius84 已提交
343 344

 protected:
H
hong 已提交
345 346
  std::unique_ptr<T> Apply() const override {
    auto* op_desc_ptr = new T();
A
Aurelius84 已提交
347
    op_desc_ptr->SetType("pyramid_hash_grad");
H
hong 已提交
348 349 350 351 352 353 354 355 356
    op_desc_ptr->SetInput("X", this->Input("X"));
    op_desc_ptr->SetInput("W", this->Input("W"));
    op_desc_ptr->SetInput("DropPos", this->Output("DropPos"));

    op_desc_ptr->SetInput(framework::GradVarName("Out"),
                          this->OutputGrad("Out"));
    op_desc_ptr->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
    op_desc_ptr->SetAttrMap(this->Attrs());
    return std::unique_ptr<T>(op_desc_ptr);
A
Aurelius84 已提交
357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439
  }
};

template <typename DeviceContext, typename T>
class CPUPyramidHashOPGradKernel : public framework::OpKernel<T> {
 public:
  void hash_embedding_bp(const T* hash_id, int len, const T* top_pos,
                         T* weights, T mlr, int _num_emb, int _rand_len,
                         int _space_len) const {
    for (unsigned int j = 0; j != _num_emb; j += _rand_len) {
      unsigned int pos = XXH32(hash_id, len * sizeof(T), j) % _space_len;
      avx_axpy(top_pos + j, weights + pos, _rand_len, mlr);
    }
  }

  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* bottom = ctx.Input<LoDTensor>("X");
    auto* _blobs = ctx.Input<Tensor>("W");
    auto* drop_pos = ctx.Input<LoDTensor>("DropPos");
    auto* top = ctx.Input<LoDTensor>(framework::GradVarName("Out"));

    int _num_emb = ctx.Attr<int>("num_emb");
    float _lr = ctx.Attr<float>("lr");
    int _rand_len = ctx.Attr<int>("rand_len");
    int _space_len = ctx.Attr<int>("space_len");
    int _pyramid_layer = ctx.Attr<int>("pyramid_layer");

    const auto* bottom_data_ori = bottom->data<int32_t>();
    Tensor buff;
    buff.Resize(framework::make_ddim({bottom->dims()[0], bottom->dims()[1]}));
    T* bottom_data = buff.mutable_data<T>(ctx.GetPlace());
    for (size_t i = 0; i < bottom->dims()[0]; i++) {
      bottom_data[i] = bottom_data_ori[i];
    }

    int _slot_len = bottom->dims()[0];
    if (_slot_len == bottom->lod()[0].size() - 1 &&
        std::count(bottom_data, bottom_data + _slot_len, -1) == _slot_len) {
      return;
    }

    auto& offset = bottom->lod()[0];
    auto& drop_pos_offset = drop_pos->lod()[0];

    const auto* top_diff = top->data<T>();
    T* weights = const_cast<T*>(_blobs->data<T>());
    T mlr = -1.0 * _lr;

    const int* iter = drop_pos->data<int>();
    int top_counter = 0;
    for (int i = 0; i < offset.size() - 1; ++i) {
      int w = offset[i + 1] - offset[i];
      int w_drop = drop_pos_offset[i + 1] - drop_pos_offset[i];
      if (w_drop == 0) {
        top_counter++;
      }
      if (w > 1) {
        for (int ilayer = 1; ilayer < _pyramid_layer && ilayer < w; ++ilayer) {
          for (int l = 0; l < w - ilayer; ++l) {
            if (*(iter++) == 0) {
              // do nothing
            } else {
              const T* top_pos = top_diff + top_counter++ * _num_emb;
              hash_embedding_bp((const T*)(bottom_data + offset[i] + l),
                                ilayer + 1, top_pos, weights, mlr, _num_emb,
                                _rand_len, _space_len);
            }
          }
        }
      } else {
        // do nothing
      }
    }
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
namespace plt = paddle::platform;
namespace frm = paddle::framework;
REGISTER_OPERATOR(pyramid_hash, ops::PyramidHashOP, ops::PyramidHashOpMaker,
H
hong 已提交
440 441
                  ops::PyramidHashGradOpMaker<paddle::framework::OpDesc>,
                  ops::PyramidHashGradOpMaker<paddle::imperative::OpBase>);
A
Aurelius84 已提交
442 443 444 445 446 447 448
REGISTER_OPERATOR(pyramid_hash_grad, ops::PyramidHashOpGrad);

REGISTER_OP_CPU_KERNEL(
    pyramid_hash, ops::CPUPyramidHashOPKernel<plt::CPUDeviceContext, float>);
REGISTER_OP_CPU_KERNEL(
    pyramid_hash_grad,
    ops::CPUPyramidHashOPGradKernel<plt::CPUDeviceContext, float>);