未验证 提交 552cdc12 编写于 作者: Z Zhaolong Xing 提交者: GitHub

Merge pull request #13422 from NHZlX/add_dropout_simoid_trt

Add dropout sigmoid op converter for trt
...@@ -69,8 +69,9 @@ class DfgPassManagerImpl final : public DfgPassManager { ...@@ -69,8 +69,9 @@ class DfgPassManagerImpl final : public DfgPassManager {
if (FLAGS_IA_enable_tensorrt_subgraph_engine) { if (FLAGS_IA_enable_tensorrt_subgraph_engine) {
auto trt_teller = [&](const Node* node) { auto trt_teller = [&](const Node* node) {
std::unordered_set<std::string> teller_set( std::unordered_set<std::string> teller_set(
{"elementwise_add", "mul", "conv2d", "pool2d", "relu", "softmax", {"mul", "conv2d", "pool2d", "relu", "softmax", "sigmoid",
"depthwise_conv2d", "batch_norm", "concat"}); "depthwise_conv2d", "batch_norm", "concat", "tanh",
"elementwise_add", "dropout"});
if (!node->IsFunction()) return false; if (!node->IsFunction()) return false;
const auto* func = static_cast<const Function*>(node); const auto* func = static_cast<const Function*>(node);
......
...@@ -153,11 +153,21 @@ CreatePaddlePredictor<TensorRTConfig, PaddleEngineKind::kAutoMixedTensorRT>( ...@@ -153,11 +153,21 @@ CreatePaddlePredictor<TensorRTConfig, PaddleEngineKind::kAutoMixedTensorRT>(
} // namespace paddle } // namespace paddle
USE_TRT_CONVERTER(elementwise_add_weight); USE_TRT_CONVERTER(elementwise_add_weight);
USE_TRT_CONVERTER(elementwise_add_tensor);
USE_TRT_CONVERTER(elementwise_sub_tensor);
USE_TRT_CONVERTER(elementwise_div_tensor);
USE_TRT_CONVERTER(elementwise_mul_tensor);
USE_TRT_CONVERTER(elementwise_max_tensor);
USE_TRT_CONVERTER(elementwise_min_tensor);
USE_TRT_CONVERTER(elementwise_pow_tensor);
USE_TRT_CONVERTER(mul); USE_TRT_CONVERTER(mul);
USE_TRT_CONVERTER(conv2d); USE_TRT_CONVERTER(conv2d);
USE_TRT_CONVERTER(relu); USE_TRT_CONVERTER(relu);
USE_TRT_CONVERTER(sigmoid);
USE_TRT_CONVERTER(tanh);
USE_TRT_CONVERTER(fc); USE_TRT_CONVERTER(fc);
USE_TRT_CONVERTER(pool2d); USE_TRT_CONVERTER(pool2d);
USE_TRT_CONVERTER(softmax); USE_TRT_CONVERTER(softmax);
USE_TRT_CONVERTER(batch_norm); USE_TRT_CONVERTER(batch_norm);
USE_TRT_CONVERTER(concat); USE_TRT_CONVERTER(concat);
USE_TRT_CONVERTER(dropout);
# Add TRT tests # Add TRT tests
nv_library(tensorrt_converter nv_library(tensorrt_converter
SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc
batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc dropout_op.cc
DEPS tensorrt_engine operator scope framework_proto op_registry) DEPS tensorrt_engine operator scope framework_proto op_registry)
nv_test(test_op_converter SRCS test_op_converter.cc DEPS nv_test(test_op_converter SRCS test_op_converter.cc DEPS
...@@ -24,6 +24,8 @@ nv_test(test_trt_softmax_op SRCS test_softmax_op.cc softmax_op.cc ...@@ -24,6 +24,8 @@ nv_test(test_trt_softmax_op SRCS test_softmax_op.cc softmax_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine softmax_op SERIAL) DEPS ${FLUID_CORE_MODULES} tensorrt_engine softmax_op SERIAL)
nv_test(test_trt_batch_norm_op SRCS test_batch_norm_op.cc batch_norm_op.cc nv_test(test_trt_batch_norm_op SRCS test_batch_norm_op.cc batch_norm_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine batch_norm_op SERIAL) DEPS ${FLUID_CORE_MODULES} tensorrt_engine batch_norm_op SERIAL)
nv_test(test_trt_concat_op SRCS test_concat_op.cc concat_op.cc nv_test(test_trt_concat_op SRCS test_concat_op.cc concat_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine concat_op SERIAL) DEPS ${FLUID_CORE_MODULES} tensorrt_engine concat_op SERIAL)
nv_test(test_trt_dropout_op SRCS test_dropout_op.cc dropout_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine dropout_op SERIAL)
...@@ -19,23 +19,31 @@ namespace paddle { ...@@ -19,23 +19,31 @@ namespace paddle {
namespace inference { namespace inference {
namespace tensorrt { namespace tensorrt {
class ReluOpConverter : public OpConverter { class ActivationOpConverter : public OpConverter {
public: public:
ReluOpConverter() {} ActivationOpConverter() {}
void operator()(const framework::proto::OpDesc& op, void operator()(const framework::proto::OpDesc& op,
const framework::Scope& scope, bool test_mode) override { const framework::Scope& scope, bool test_mode) override {
// Here the two nullptr looks strange, that's because the // Here the two nullptr looks strange, that's because the
// framework::OpDesc's constructor is strange. // framework::OpDesc's constructor is strange.
framework::OpDesc op_desc(op, nullptr); framework::OpDesc op_desc(op, nullptr);
LOG(INFO) << "convert a fluid relu op to tensorrt activation layer whose " LOG(INFO)
"type is Relu"; << "convert a fluid Activation op to tensorrt activation layer whose "
"type is "
<< op_type_;
const nvinfer1::ITensor* input_tensor = const nvinfer1::ITensor* input_tensor =
engine_->GetITensor(op_desc.Input("X")[0]); engine_->GetITensor(op_desc.Input("X")[0]);
auto op_pair = ops.find(op_type_);
if (op_pair == ops.end()) {
PADDLE_THROW("Wrong activation op type!");
}
nvinfer1::IActivationLayer* layer = TRT_ENGINE_ADD_LAYER( nvinfer1::IActivationLayer* layer = TRT_ENGINE_ADD_LAYER(
engine_, Activation, *const_cast<nvinfer1::ITensor*>(input_tensor), engine_, Activation, *const_cast<nvinfer1::ITensor*>(input_tensor),
nvinfer1::ActivationType::kRELU); op_pair->second);
auto output_name = op_desc.Output("Out")[0]; auto output_name = op_desc.Output("Out")[0];
layer->setName(("relu (Output: " + output_name + ")").c_str()); layer->setName((op_type_ + " (Output: " + output_name + ")").c_str());
layer->getOutput(0)->setName(output_name.c_str()); layer->getOutput(0)->setName(output_name.c_str());
engine_->SetITensor(output_name, layer->getOutput(0)); engine_->SetITensor(output_name, layer->getOutput(0));
if (test_mode) { // the test framework can not determine which is the if (test_mode) { // the test framework can not determine which is the
...@@ -43,6 +51,32 @@ class ReluOpConverter : public OpConverter { ...@@ -43,6 +51,32 @@ class ReluOpConverter : public OpConverter {
engine_->DeclareOutput(output_name); engine_->DeclareOutput(output_name);
} }
} }
protected:
std::string op_type_;
static const std::unordered_map<std::string, nvinfer1::ActivationType> ops;
};
const std::unordered_map<std::string, nvinfer1::ActivationType>
ActivationOpConverter::ops = {
{"relu", nvinfer1::ActivationType::kRELU},
{"sigmoid", nvinfer1::ActivationType::kSIGMOID},
{"tanh", nvinfer1::ActivationType::kTANH},
};
class ReluOpConverter : public ActivationOpConverter {
public:
ReluOpConverter() { op_type_ = "relu"; }
};
class SigmoidOpConverter : public ActivationOpConverter {
public:
SigmoidOpConverter() { op_type_ = "sigmoid"; }
};
class TanhOpConverter : public ActivationOpConverter {
public:
TanhOpConverter() { op_type_ = "tanh"; }
}; };
} // namespace tensorrt } // namespace tensorrt
...@@ -50,3 +84,5 @@ class ReluOpConverter : public OpConverter { ...@@ -50,3 +84,5 @@ class ReluOpConverter : public OpConverter {
} // namespace paddle } // namespace paddle
REGISTER_TRT_OP_CONVERTER(relu, ReluOpConverter); REGISTER_TRT_OP_CONVERTER(relu, ReluOpConverter);
REGISTER_TRT_OP_CONVERTER(sigmoid, SigmoidOpConverter);
REGISTER_TRT_OP_CONVERTER(tanh, TanhOpConverter);
/* 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 {
VLOG(4) << "convert a fluid dropout op to tensorrt dropout layer";
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"));
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());
engine_->weight_map[op_desc.Output("Out").front() + "_dropout"] =
std::move(weight_tensor);
auto output_name = op_desc.Output("Out")[0];
layer->setName(("dropout (Output: " + output_name + ")").c_str());
engine_->SetITensor(output_name, layer->getOutput(0));
if (test_mode) {
engine_->DeclareOutput(output_name);
}
}
};
} // namespace tensorrt
} // namespace inference
} // namespace paddle
USE_OP(dropout);
REGISTER_TRT_OP_CONVERTER(dropout, DropoutOpConverter);
...@@ -20,18 +20,18 @@ namespace paddle { ...@@ -20,18 +20,18 @@ namespace paddle {
namespace inference { namespace inference {
namespace tensorrt { namespace tensorrt {
TEST(ReluOpConverter, main) { void test_activation(std::string act_type) {
framework::Scope scope; framework::Scope scope;
std::unordered_set<std::string> parameters; std::unordered_set<std::string> parameters;
TRTConvertValidation validator(10, parameters, scope, 1000); TRTConvertValidation validator(10, parameters, scope, 1000);
validator.DeclInputVar("relu-X", nvinfer1::Dims2(10, 6)); validator.DeclInputVar("act-X", nvinfer1::Dims2(10, 6));
validator.DeclOutputVar("relu-Out", nvinfer1::Dims2(10, 6)); validator.DeclOutputVar("act-Out", nvinfer1::Dims2(10, 6));
// Prepare Op description // Prepare Op description
framework::OpDesc desc; framework::OpDesc desc;
desc.SetType("relu"); desc.SetType(act_type);
desc.SetInput("X", {"relu-X"}); desc.SetInput("X", {"act-X"});
desc.SetOutput("Out", {"relu-Out"}); desc.SetOutput("Out", {"act-Out"});
LOG(INFO) << "set OP"; LOG(INFO) << "set OP";
validator.SetOp(*desc.Proto()); validator.SetOp(*desc.Proto());
...@@ -40,8 +40,16 @@ TEST(ReluOpConverter, main) { ...@@ -40,8 +40,16 @@ TEST(ReluOpConverter, main) {
validator.Execute(5); validator.Execute(5);
} }
TEST(ReluOpConverter, main) { test_activation("relu"); }
TEST(SigmoidOpConverter, main) { test_activation("sigmoid"); }
TEST(TanhOpConverter, main) { test_activation("tanh"); }
} // namespace tensorrt } // namespace tensorrt
} // namespace inference } // namespace inference
} // namespace paddle } // namespace paddle
USE_OP(relu); USE_OP(relu);
USE_OP(sigmoid);
USE_OP(tanh);
/* 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 <gtest/gtest.h>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace tensorrt {
TEST(DropoutOpConverter, main) {
framework::Scope scope;
std::unordered_set<std::string> parameters;
TRTConvertValidation validator(8, parameters, scope, 1000);
std::vector<int> tensor_shape{8, 10};
validator.DeclInputVar("dropout-X", tensor_shape,
nvinfer1::DimsCHW(10, 1, 1));
validator.DeclOutputVar("dropout-Out", nvinfer1::DimsCHW(10, 1, 1));
validator.DeclOutputVar("mask-Out", nvinfer1::DimsCHW(10, 1, 1));
// Prepare Op description
framework::OpDesc desc;
int is_test = 1;
float dropout_prob = 0.4;
desc.SetType("dropout");
desc.SetInput("X", {"dropout-X"});
desc.SetOutput("Mask", {"mask-Out"});
desc.SetOutput("Out", {"dropout-Out"});
desc.SetAttr("is_test", is_test);
desc.SetAttr("dropout_prob", dropout_prob);
LOG(INFO) << "set OP";
validator.SetOp(*desc.Proto());
LOG(INFO) << "execute";
std::unordered_set<std::string> neglected_output = {"mask-Out"};
validator.Execute(8, neglected_output);
}
} // namespace tensorrt
} // namespace inference
} // namespace paddle
USE_OP(dropout);
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册