pool2d_op.cc 6.3 KB
Newer Older
N
nhzlx 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* 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. */

#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
N
nhzlx 已提交
16
#include "paddle/fluid/inference/tensorrt/plugin/avg_pool_op_plugin.h"
N
nhzlx 已提交
17 18 19 20 21

namespace paddle {
namespace inference {
namespace tensorrt {

N
nhzlx 已提交
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
void DealCeilMode(const nvinfer1::Dims &input_shape, std::vector<int> ksize,
                  std::vector<int> strides, std::vector<int> paddings,
                  nvinfer1::DimsHW *pre_pad, nvinfer1::DimsHW *post_pad,
                  int input_dims) {
  int input_height = input_shape.d[input_dims - 2];
  int input_width = input_shape.d[input_dims - 1];
  int floor_h_output_size =
      (input_height - ksize[0] + 2 * paddings[0]) / strides[0] + 1;
  int ceil_h_output_size =
      (input_height - ksize[0] + 2 * paddings[0] + strides[0] - 1) /
          strides[0] +
      1;

  int floor_w_output_size =
      (input_width - ksize[1] + 2 * paddings[1]) / strides[1] + 1;
  int ceil_w_output_size =
      (input_width - ksize[1] + 2 * paddings[1] + strides[1] - 1) / strides[1] +
      1;
  if (floor_h_output_size != ceil_h_output_size) {
    post_pad->h() = strides[0] - 1;
  }

  if (floor_w_output_size != ceil_w_output_size) {
    post_pad->w() = strides[1] - 1;
  }
}

N
nhzlx 已提交
49 50 51 52 53
/*
 * Pool2dOp, IPoolingLayer in TRT. This Layer doesn't has weights.
 */
class Pool2dOpConverter : public OpConverter {
 public:
N
nhzlx 已提交
54 55
  void operator()(const framework::proto::OpDesc &op,
                  const framework::Scope &scope, bool test_mode) override {
M
minqiyang 已提交
56
    VLOG(4)
N
nhzlx 已提交
57 58 59
        << "convert a fluid pool2d op to tensorrt pool2d layer without bias";
    framework::OpDesc op_desc(op, nullptr);
    // Declare inputs
N
nhzlx 已提交
60 61
    PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
    PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
N
nhzlx 已提交
62 63 64 65 66
    auto *input1 = engine_->GetITensor(op_desc.Input("X")[0]);
    nvinfer1::Dims input_shape = input1->getDimensions();
    int input_dims = input_shape.nbDims;

    PADDLE_ENFORCE_EQ(input_dims, 3UL);
N
nhzlx 已提交
67

N
nhzlx 已提交
68
    bool global_pooling = boost::get<bool>(op_desc.GetAttr("global_pooling"));
N
nhzlx 已提交
69 70 71 72 73 74 75 76
    std::string pool_type =
        boost::get<std::string>(op_desc.GetAttr("pooling_type"));
    std::vector<int> ksize =
        boost::get<std::vector<int>>(op_desc.GetAttr("ksize"));
    std::vector<int> strides =
        boost::get<std::vector<int>>(op_desc.GetAttr("strides"));
    std::vector<int> paddings =
        boost::get<std::vector<int>>(op_desc.GetAttr("paddings"));
77
    bool ceil_mode = boost::get<bool>(op_desc.GetAttr("ceil_mode"));
N
nhzlx 已提交
78

N
nhzlx 已提交
79
    nvinfer1::PoolingType nv_pool_type = nvinfer1::PoolingType::kMAX;
N
nhzlx 已提交
80
    if (pool_type == "max") {
N
nhzlx 已提交
81
      nv_pool_type = nvinfer1::PoolingType::kMAX;
N
nhzlx 已提交
82
    } else if (pool_type == "avg") {
N
nhzlx 已提交
83
      nv_pool_type = nvinfer1::PoolingType::kAVERAGE;
N
nhzlx 已提交
84 85 86 87
    } else {
      PADDLE_THROW("TensorRT unsupported pooling type!");
    }

N
nhzlx 已提交
88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104
    nvinfer1::DimsHW nv_ksize(ksize[0], ksize[1]);
    nvinfer1::DimsHW nv_strides(strides[0], strides[1]);
    nvinfer1::DimsHW nv_paddings(paddings[0], paddings[1]);

    nvinfer1::ILayer *layer = nullptr;

    if (global_pooling == true) {
      nv_ksize.d[0] = input_shape.d[input_dims - 2];
      nv_ksize.d[1] = input_shape.d[input_dims - 1];
      auto *layer = TRT_ENGINE_ADD_LAYER(
          engine_, Pooling, *const_cast<nvinfer1::ITensor *>(input1),
          nv_pool_type, nv_ksize);
      PADDLE_ENFORCE_NOT_NULL(layer, "pool layer could not be created.");
      auto output_name = op_desc.Output("Out")[0];
      layer->setName(("pool2d (Output: " + output_name + ")").c_str());
      layer->getOutput(0)->setName(output_name.c_str());
      engine_->SetITensor(output_name, layer->getOutput(0));
N
nhzlx 已提交
105
      if (test_mode) {
N
nhzlx 已提交
106
        engine_->DeclareOutput(output_name);
107
      }
N
nhzlx 已提交
108 109
      return;
    }
110

N
nhzlx 已提交
111
    if (pool_type == "max") {
N
nhzlx 已提交
112 113 114 115
      // Under ceil mode, the pre_pad and post_pad are used to
      // record the the padding size. In some ceil mode cases,
      // we do not need padding, so we initialize the two vars to 0.

N
nhzlx 已提交
116 117
      nvinfer1::DimsHW pre_pad(0, 0);
      nvinfer1::DimsHW post_pad(0, 0);
N
nhzlx 已提交
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
      if (ceil_mode) {
        // If ceil mode is true, we will pad the appropriate size to the input.
        DealCeilMode(input_shape, ksize, strides, paddings, &pre_pad, &post_pad,
                     input_dims);
        auto *pad_layer = TRT_ENGINE_ADD_LAYER(
            engine_, Padding, *const_cast<nvinfer1::ITensor *>(input1), pre_pad,
            post_pad);
        PADDLE_ENFORCE_NOT_NULL(
            pad_layer, "pad layer in poolOp converter could not be created.");
        input1 = pad_layer->getOutput(0);
      }
      auto *pool_layer = TRT_ENGINE_ADD_LAYER(
          engine_, Pooling, *const_cast<nvinfer1::ITensor *>(input1),
          nv_pool_type, nv_ksize);
      PADDLE_ENFORCE_NOT_NULL(pool_layer, "pool layer could not be created.");
      pool_layer->setStride(nv_strides);
      pool_layer->setPadding(nv_paddings);
      layer = pool_layer;
    } else {
      // Average pooling needs to exclude the padding pixels from the average
      // mean.
      // It is not supported well by TRT, we use a plugin here.
      std::vector<int> input_shape_v;
      for (int i = 0; i < input_dims; i++) {
        input_shape_v.push_back(input_shape.d[i]);
143
      }
N
nhzlx 已提交
144 145
      plugin::AvgPoolPlugin *plugin = new plugin::AvgPoolPlugin(
          ceil_mode, ksize, strides, paddings, input_shape_v);
N
nhzlx 已提交
146 147
      auto *avg_pool_layer = engine_->AddPlugin(&input1, 1, plugin);
      layer = avg_pool_layer;
148
    }
N
nhzlx 已提交
149 150

    auto output_name = op_desc.Output("Out")[0];
151 152
    layer->setName(("pool2d (Output: " + output_name + ")").c_str());
    layer->getOutput(0)->setName(output_name.c_str());
N
nhzlx 已提交
153
    engine_->SetITensor(output_name, layer->getOutput(0));
N
nhzlx 已提交
154
    if (test_mode) {
N
nhzlx 已提交
155 156 157 158 159 160 161 162 163 164 165
      engine_->DeclareOutput(output_name);
    }
  }
};

}  // namespace tensorrt
}  // namespace inference
}  // namespace paddle

USE_OP(pool2d);
REGISTER_TRT_OP_CONVERTER(pool2d, Pool2dOpConverter);