softmax_op.cc 2.8 KB
Newer Older
N
nhzlx 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* 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. */

S
Shang Zhizhou 已提交
15
#include <algorithm>
N
nhzlx 已提交
16 17
#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"

S
Shang Zhizhou 已提交
18 19 20 21 22 23 24 25 26
namespace paddle {
namespace framework {
class Scope;
namespace proto {
class OpDesc;
}  // namespace proto
}  // namespace framework
}  // namespace paddle

N
nhzlx 已提交
27 28 29 30 31 32 33 34 35 36 37
namespace paddle {
namespace inference {
namespace tensorrt {

/*
 * SoftMaxOp, ISoftMaxLayer in TRT. This Layer doesn't has weights.
 */
class SoftMaxOpConverter : public OpConverter {
 public:
  void operator()(const framework::proto::OpDesc& op,
                  const framework::Scope& scope, bool test_mode) override {
38
    VLOG(3)
N
nhzlx 已提交
39 40 41 42
        << "convert a fluid softmax op to tensorrt softmax layer without bias";
    framework::OpDesc op_desc(op, nullptr);
    // Declare inputs
    auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]);
S
Shang Zhizhou 已提交
43 44 45 46 47
    nvinfer1::Dims input_shape = input1->getDimensions();
    int input_dims = input_shape.nbDims;
    int axis =
        op_desc.HasAttr("axis") ? boost::get<int>(op_desc.GetAttr("axis")) : -1;

N
nhzlx 已提交
48 49
    auto* layer = TRT_ENGINE_ADD_LAYER(engine_, SoftMax,
                                       *const_cast<nvinfer1::ITensor*>(input1));
S
Shang Zhizhou 已提交
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
    uint32_t axes = std::max(0, input_dims - 3);
    // TODO(cryoco): Poor workaround. Fix padded dims problem when TRT layers
    // support Nd.
    int padded_dims = 0;
    int explicit_batch = 0;
    if (engine_->with_dynamic_shape()) explicit_batch = 1;
    for (int i = input_dims - 1; i > explicit_batch; i--) {
      if (input_shape.d[i] == 1) {
        padded_dims += 1;
      } else {
        break;
      }
    }
    if (!engine_->with_dynamic_shape()) {
      if (axis == -1) {
        axes = input_dims - 1 - padded_dims;
      } else {
        axes = axis;
      }
    } else {
      if (axis == -1) {
        axes = input_dims - 1 - padded_dims;
      } else {
        axes = axis + 1;
      }
    }
    layer->setAxes(1 << axes);
N
nhzlx 已提交
77
    auto output_name = op_desc.Output("Out")[0];
78 79
    RreplenishLayerAndOutput(layer, "softmax", {output_name}, test_mode);

80 81
    // The trt will not run int for softmax.
    engine_->SetTensorDynamicRange(input1, 1.0);
N
nhzlx 已提交
82 83 84 85 86 87 88 89 90
  }
};

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

USE_OP(softmax);
REGISTER_TRT_OP_CONVERTER(softmax, SoftMaxOpConverter);