dropout_op.cc 3.0 KB
Newer Older
N
nhzlx 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
/* 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"

namespace paddle {
namespace inference {
namespace tensorrt {

/*
 * DropoutOp. This Layer doesn't has weights.
 */
class DropoutOpConverter : public OpConverter {
 public:
  void operator()(const framework::proto::OpDesc& op,
                  const framework::Scope& scope, bool test_mode) override {
28
    VLOG(3) << "convert a fluid dropout op to tensorrt dropout layer";
N
nhzlx 已提交
29 30 31 32 33
    framework::OpDesc op_desc(op, nullptr);
    // Declare inputs
    auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]);
    float dropout_prob = boost::get<float>(op_desc.GetAttr("dropout_prob"));

34 35 36 37 38 39 40 41 42 43 44 45 46 47
    std::string downgrade_in_infer = "";
    if (op_desc.HasAttr("dropout_implementation")) {
      downgrade_in_infer =
          boost::get<std::string>(op_desc.GetAttr("dropout_implementation"));
    }

    if (!downgrade_in_infer.empty() &&
        downgrade_in_infer == "upscale_in_train") {
      auto* layer = TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *input1);
      auto output_name = op_desc.Output("Out")[0];
      RreplenishLayerAndOutput(layer, "dropout", {output_name}, test_mode);
      return;
    }

N
nhzlx 已提交
48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68
    platform::CPUPlace cpu_place;
    std::unique_ptr<framework::LoDTensor> weight_tensor(
        new framework::LoDTensor());
    weight_tensor->Resize(framework::make_ddim({1}));
    auto* weight_data =
        weight_tensor->mutable_data<float>(platform::CPUPlace());
    weight_data[0] = 1 - dropout_prob;

    TensorRTEngine::Weight scale_weights{
        nvinfer1::DataType::kFLOAT, static_cast<void*>(weight_data),
        weight_tensor->memory_size() / sizeof(float)};
    TensorRTEngine::Weight shift_weights{nvinfer1::DataType::kFLOAT, nullptr,
                                         0};
    TensorRTEngine::Weight power_weights{nvinfer1::DataType::kFLOAT, nullptr,
                                         0};

    auto* layer = TRT_ENGINE_ADD_LAYER(
        engine_, Scale, *const_cast<nvinfer1::ITensor*>(input1),
        nvinfer1::ScaleMode::kUNIFORM, shift_weights.get(), scale_weights.get(),
        power_weights.get());

69 70
    engine_->SetWeights(op_desc.Output("Out").front() + "_dropout",
                        std::move(weight_tensor));
N
nhzlx 已提交
71
    auto output_name = op_desc.Output("Out")[0];
72 73

    RreplenishLayerAndOutput(layer, "dropout", {output_name}, test_mode);
N
nhzlx 已提交
74 75 76 77 78 79 80 81 82
  }
};

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

USE_OP(dropout);
REGISTER_TRT_OP_CONVERTER(dropout, DropoutOpConverter);