“da158c2279008010996687d32b88c76e4e8ec1b7”上不存在“paddle/legacy/trainer/tests/sample_trainer_config.conf”
split_ids_op.h 4.6 KB
Newer Older
Q
Qiao Longfei 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* Copyright (c) 2018 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. */

#pragma once

S
seiriosPlus 已提交
17 18
#include <iterator>
#include <set>
Q
qiaolongfei 已提交
19
#include <unordered_map>
Q
Qiao Longfei 已提交
20 21 22 23 24 25 26 27 28 29
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/selected_rows_functor.h"

namespace paddle {
namespace operators {

template <typename DeviceContext, typename T>
class SplitIdsOpKernel : public framework::OpKernel<T> {
 public:
30
  void Compute(const framework::ExecutionContext &ctx) const override {
Q
Qiao Longfei 已提交
31 32 33 34 35
    auto place = ctx.GetPlace();
    if (!platform::is_cpu_place(place)) {
      PADDLE_THROW("SplitIds do not support GPU kernel");
    }

S
seiriosPlus 已提交
36 37 38 39 40
    const auto ids_vars = ctx.MultiInputVar("Ids");

    PADDLE_ENFORCE_GT(ids_vars.size(), 0, "The number of Ids should > 0");
    auto *ids_var = ids_vars[0];

41
    if (ids_var->IsType<framework::LoDTensor>()) {
S
seiriosPlus 已提交
42 43 44 45 46
      int batch_size = 0;
      const auto ids_tensors = ctx.MultiInput<framework::LoDTensor>("Ids");
      for (size_t i = 0; i < ids_tensors.size(); ++i) {
        batch_size += ids_tensors[i]->dims()[0];
      }
M
minqiyang 已提交
47
      VLOG(4) << "Get Total BatchSize is: " << batch_size;
S
seiriosPlus 已提交
48 49 50 51 52 53 54 55 56 57 58 59 60

      std::vector<T> all_ids(batch_size);
      int offset = 0;
      for (size_t i = 0; i < ids_tensors.size(); ++i) {
        const auto *ids = ids_tensors[i];
        std::memcpy(all_ids.data() + offset, ids->data<T>(),
                    ids->numel() * sizeof(T));
        offset += ids->numel();
      }

      std::set<T> st(all_ids.begin(), all_ids.end());
      all_ids.assign(st.begin(), st.end());

61 62 63 64
      auto outs = ctx.MultiOutput<framework::LoDTensor>("Out");
      const size_t shard_num = outs.size();
      std::vector<std::vector<T>> out_ids;
      out_ids.resize(outs.size());
Q
Qiao Longfei 已提交
65

66
      // split id by their shard_num.
T
Tao Luo 已提交
67
      for (size_t i = 0; i < all_ids.size(); ++i) {
S
seiriosPlus 已提交
68
        T id = all_ids[i];
69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
        size_t shard_id = static_cast<size_t>(id) % shard_num;
        out_ids[shard_id].push_back(id);
      }

      // create tensor for each shard and send to parameter server
      for (size_t i = 0; i < out_ids.size(); ++i) {
        auto *shard_t = outs[i];
        std::vector<T> ids = out_ids[i];
        auto *shard_data = shard_t->mutable_data<T>(
            framework::make_ddim({static_cast<int64_t>(ids.size()), 1}), place);
        for (size_t i = 0; i < ids.size(); ++i) {
          shard_data[i] = ids[i];
        }
      }
    } else if (ids_var->IsType<framework::SelectedRows>()) {
      const auto *ids_selected_rows = ctx.Input<framework::SelectedRows>("Ids");
      auto &ids_dims = ids_selected_rows->value().dims();
A
Abhinav Arora 已提交
86 87 88
      PADDLE_ENFORCE_EQ(ids_dims[0],
                        static_cast<int64_t>(ids_selected_rows->rows().size()),
                        "");
S
seiriosPlus 已提交
89
      const T *ids_data = ids_selected_rows->value().data<T>();
90 91 92
      const auto &ids_rows = ids_selected_rows->rows();
      auto outs = ctx.MultiOutput<framework::SelectedRows>("Out");
      const size_t shard_num = outs.size();
93 94 95
      for (auto &out : outs) {
        out->mutable_rows()->clear();
      }
96
      // get rows for outputs
Q
qiaolongfei 已提交
97
      std::unordered_map<int64_t, size_t> id_to_index;
Q
qiaolongfei 已提交
98 99 100 101
      for (size_t i = 0; i < ids_rows.size(); ++i) {
        id_to_index[ids_rows[i]] = i;
        size_t shard_id = static_cast<size_t>(ids_rows[i]) % shard_num;
        outs[shard_id]->mutable_rows()->push_back(ids_rows[i]);
102
      }
Q
Qiao Longfei 已提交
103

104 105 106 107 108 109
      int64_t row_width = ids_dims[1];
      for (auto &out : outs) {
        out->set_height(ids_selected_rows->height());
        framework::DDim ddim = framework::make_ddim(
            {static_cast<int64_t>(out->rows().size()), row_width});
        T *output = out->mutable_value()->mutable_data<T>(ddim, place);
A
Abhinav Arora 已提交
110
        for (int64_t i = 0; i < ddim[0]; ++i) {
Q
qiaolongfei 已提交
111
          memcpy(output + i * row_width,
S
seiriosPlus 已提交
112
                 ids_data + id_to_index[out->rows()[i]] * row_width,
113 114
                 row_width * sizeof(T));
        }
Q
Qiao Longfei 已提交
115
      }
C
chengduo 已提交
116 117 118
    } else {
      PADDLE_THROW(
          "% should be LoDTensor or SelectedRows, but the received type is %s",
S
sneaxiy 已提交
119
          ctx.Inputs("Ids")[0], framework::ToTypeName(ids_var->Type()));
Q
Qiao Longfei 已提交
120 121 122 123 124 125
    }
  }
};

}  // namespace operators
}  // namespace paddle