Skip to content

S
Serving

PaddlePaddle / Serving
接近 2 年 前同步成功

通知 186
Star 833
Fork 253
S
Serving
提交 fcd66ef4 编写于 作者: W wangjiawei04
浏览文件
操作

Merge remote-tracking branch 'root/dygraph' into HEAD

上级 7ea59bda 695a7772
隐藏空白更改
内联 并排
Showing with 361 addition and 1274 deletion
cmake/external/boost.cmake
浏览文件 @ fcd66ef4
......@@ -22,8 +22,8 @@ set(BOOST_PROJECT "extern_boost")
# version of boost, say, 1.66.0, doesn't build on CentOS 6. We
# checked that the devtools package of CentOS 6 installs boost 1.41.0.
# So we use 1.41.0 here.
set(BOOST_VER "1.41.0")
set(BOOST_TAR "boost_1_41_0" CACHE STRING "" FORCE)
set(BOOST_VER "1.74.0")
set(BOOST_TAR "boost_1_74_0" CACHE STRING "" FORCE)
set(BOOST_URL "http://paddlepaddledeps.cdn.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
......
cmake/external/brpc.cmake
浏览文件 @ fcd66ef4
......@@ -13,6 +13,9 @@
# limitations under the License.
INCLUDE(ExternalProject)
set(BRPC_CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Wno-narrowing")
set(BRPC_CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing")
set(BRPC_CMAKE_CPP_FLAGS "${CMAKE_CPP_FLAGS} -Wno-narrowing")
find_package(OpenSSL REQUIRED)
......@@ -40,14 +43,15 @@ ExternalProject_Add(
extern_brpc
${EXTERNAL_PROJECT_LOG_ARGS}
# TODO(gongwb): change to de newst repo when they changed.
GIT_REPOSITORY "https://github.com/wangjiawei04/brpc"
GIT_TAG "6d79e0b17f25107c35b705ea58d888083f59ff47"
GIT_REPOSITORY "https://github.com/apache/incubator-brpc"
GIT_TAG "master"
PREFIX ${BRPC_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
-DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-DCMAKE_CXX_FLAGS=${BRPC_CMAKE_CXX_FLAGS}
-DCMAKE_C_FLAGS=${BRPC_CMAKE_C_FLAGS}
-DCMAKE_CPP_FLAGS=${BRPC_CMAKE_CPP_FLAGS}
-DCMAKE_INSTALL_PREFIX=${BRPC_INSTALL_DIR}
-DCMAKE_INSTALL_LIBDIR=${BRPC_INSTALL_DIR}/lib
-DCMAKE_POSITION_INDEPENDENT_CODE=ON
......
cmake/paddlepaddle.cmake
浏览文件 @ fcd66ef4
......@@ -31,11 +31,11 @@ message( "WITH_GPU = ${WITH_GPU}")
# Paddle Version should be one of:
# latest: latest develop build
# version number like 1.5.2
SET(PADDLE_VERSION "1.8.4")
SET(PADDLE_VERSION "2.0.0-rc1")
if (WITH_GPU)
if (WITH_TRT)
SET(PADDLE_LIB_VERSION "${PADDLE_VERSION}-gpu-cuda10.1-cudnn7.6-avx-mkl-trt6")
SET(PADDLE_LIB_VERSION "${PADDLE_VERSION}-gpu-cuda10.1-cudnn7-avx-mkl-trt6")
else()
SET(PADDLE_LIB_VERSION "${PADDLE_VERSION}-gpu-cuda10-cudnn7-avx-mkl")
endif()
......@@ -51,7 +51,7 @@ else()
endif()
endif()
SET(PADDLE_LIB_PATH "http://paddle-inference-lib.bj.bcebos.com/${PADDLE_LIB_VERSION}/fluid_inference.tgz")
SET(PADDLE_LIB_PATH "http://paddle-inference-lib.bj.bcebos.com/${PADDLE_LIB_VERSION}/paddle_inference.tgz")
MESSAGE(STATUS "PADDLE_LIB_PATH=${PADDLE_LIB_PATH}")
if (WITH_GPU OR WITH_MKLML)
if (WITH_TRT)
......
core/configure/CMakeLists.txt
浏览文件 @ fcd66ef4
......@@ -14,10 +14,6 @@ list(APPEND configure_srcs ${CMAKE_CURRENT_LIST_DIR}/src/configure_parser.cpp)
add_library(configure ${configure_srcs})
add_dependencies(configure brpc)
add_executable(test_configure
${CMAKE_CURRENT_LIST_DIR}/tests/test_configure.cpp)
target_link_libraries(test_configure configure protobuf)
install(TARGETS configure
ARCHIVE DESTINATION ${PADDLE_SERVING_INSTALL_DIR}/lib
)
......@@ -45,19 +41,19 @@ add_custom_target(sdk_configure_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E to
add_dependencies(sdk_configure_py_proto sdk_configure_py_proto_init)
add_custom_command(TARGET sdk_configure_py_proto POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_client/proto
COMMAND cp *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_client/proto
COMMAND cp -f *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_client/proto
COMMENT "Copy generated python proto into directory paddle_serving_client/proto."
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
add_custom_command(TARGET general_model_config_py_proto POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_client/proto
COMMAND cp *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_client/proto
COMMAND cp -f *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_client/proto
COMMENT "Copy generated general_model_config proto file into directory paddle_serving_client/proto."
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
add_custom_command(TARGET multi_lang_general_model_service_py_proto POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_client/proto
COMMAND cp *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_client/proto
COMMAND cp -f *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_client/proto
COMMENT "Copy generated multi_lang_general_model_service proto file into directory paddle_serving_client/proto."
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
endif()
......@@ -65,7 +61,7 @@ endif()
if (APP)
add_custom_command(TARGET general_model_config_py_proto POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_app/proto
COMMAND cp *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_app/proto
COMMAND cp -f *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_app/proto
COMMENT "Copy generated general_model_config proto file into directory paddle_serving_app/proto."
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
endif()
......@@ -77,26 +73,26 @@ add_dependencies(server_config_py_proto server_config_py_proto_init)
if (NOT WITH_GPU)
add_custom_command(TARGET server_config_py_proto POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server/proto
COMMAND cp *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server/proto
COMMAND cp -f *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server/proto
COMMENT "Copy generated python proto into directory paddle_serving_server/proto."
WORKING_DIRECTORY ${CMAKE_CURRENT_BINRARY_DIR})
add_custom_command(TARGET general_model_config_py_proto POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server/proto
COMMAND cp *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server/proto
COMMAND cp -f *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server/proto
COMMENT "Copy generated general_model_config proto file into directory paddle_serving_server/proto."
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
add_custom_command(TARGET multi_lang_general_model_service_py_proto POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server/proto
COMMAND cp *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server/proto
COMMAND cp -f *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server/proto
COMMENT "Copy generated multi_lang_general_model_service proto file into directory paddle_serving_server/proto."
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
else()
add_custom_command(TARGET server_config_py_proto POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory
${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server_gpu/proto
COMMAND cp *.py
COMMAND cp -f *.py
${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server_gpu/proto
COMMENT "Copy generated python proto into directory
paddle_serving_server_gpu/proto."
......@@ -105,7 +101,7 @@ add_custom_command(TARGET server_config_py_proto POST_BUILD
add_custom_command(TARGET general_model_config_py_proto POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory
${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server_gpu/proto
COMMAND cp *.py
COMMAND cp -f *.py
${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server_gpu/proto
COMMENT "Copy generated general_model_config proto file into directory
paddle_serving_server_gpu/proto."
......@@ -113,7 +109,7 @@ add_custom_command(TARGET general_model_config_py_proto POST_BUILD
add_custom_command(TARGET multi_lang_general_model_service_py_proto POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server_gpu/proto
COMMAND cp *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server_gpu/proto
COMMAND cp -f *.py ${PADDLE_SERVING_BINARY_DIR}/python/paddle_serving_server_gpu/proto
COMMENT "Copy generated multi_lang_general_model_service proto file into directory paddle_serving_server_gpu/proto."
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
endif()
......
core/general-server/op/general_dist_kv_infer_op.cpp
浏览文件 @ fcd66ef4
......@@ -39,142 +39,6 @@ using baidu::paddle_serving::predictor::InferManager;
using baidu::paddle_serving::predictor::PaddleGeneralModelConfig;
int GeneralDistKVInferOp::inference() {
VLOG(2) << "Going to run inference";
const std::vector<std::string> pre_node_names = pre_names();
if (pre_node_names.size() != 1) {
LOG(ERROR) << "This op(" << op_name()
<< ") can only have one predecessor op, but received "
<< pre_node_names.size();
return -1;
}
const std::string pre_name = pre_node_names[0];
const GeneralBlob *input_blob = get_depend_argument<GeneralBlob>(pre_name);
uint64_t log_id = input_blob->GetLogId();
VLOG(2) << "(logid=" << log_id << ") Get precedent op name: " << pre_name;
GeneralBlob *output_blob = mutable_data<GeneralBlob>();
if (!input_blob) {
LOG(ERROR) << "(logid=" << log_id
<< ") Failed mutable depended argument, op:" << pre_name;
return -1;
}
const TensorVector *in = &input_blob->tensor_vector;
TensorVector *out = &output_blob->tensor_vector;
int batch_size = input_blob->GetBatchSize();
VLOG(2) << "(logid=" << log_id << ") input batch size: " << batch_size;
std::vector<uint64_t> keys;
std::vector<rec::mcube::CubeValue> values;
int sparse_count = 0;
int dense_count = 0;
std::vector<std::pair<int64_t *, size_t>> dataptr_size_pairs;
size_t key_len = 0;
for (size_t i = 0; i < in->size(); ++i) {
if (in->at(i).dtype != paddle::PaddleDType::INT64) {
++dense_count;
continue;
}
++sparse_count;
size_t elem_num = 1;
for (size_t s = 0; s < in->at(i).shape.size(); ++s) {
elem_num *= in->at(i).shape[s];
}
key_len += elem_num;
int64_t *data_ptr = static_cast<int64_t *>(in->at(i).data.data());
dataptr_size_pairs.push_back(std::make_pair(data_ptr, elem_num));
}
keys.resize(key_len);
int key_idx = 0;
for (size_t i = 0; i < dataptr_size_pairs.size(); ++i) {
std::copy(dataptr_size_pairs[i].first,
dataptr_size_pairs[i].first + dataptr_size_pairs[i].second,
keys.begin() + key_idx);
key_idx += dataptr_size_pairs[i].second;
}
Timer timeline;
int64_t cube_start = timeline.TimeStampUS();
timeline.Start();
rec::mcube::CubeAPI *cube = rec::mcube::CubeAPI::instance();
std::vector<std::string> table_names = cube->get_table_names();
if (table_names.size() == 0) {
LOG(ERROR) << "(logid=" << log_id
<< ") cube init error or cube config not given.";
return -1;
}
int ret = cube->seek(table_names[0], keys, &values);
int64_t cube_end = timeline.TimeStampUS();
if (values.size() != keys.size() || values[0].buff.size() == 0) {
LOG(ERROR) << "(logid=" << log_id << ") cube value return null";
}
size_t EMBEDDING_SIZE = values[0].buff.size() / sizeof(float);
TensorVector sparse_out;
sparse_out.resize(sparse_count);
TensorVector dense_out;
dense_out.resize(dense_count);
int cube_val_idx = 0;
int sparse_idx = 0;
int dense_idx = 0;
std::unordered_map<int, int> in_out_map;
baidu::paddle_serving::predictor::Resource &resource =
baidu::paddle_serving::predictor::Resource::instance();
std::shared_ptr<PaddleGeneralModelConfig> model_config =
resource.get_general_model_config();
for (size_t i = 0; i < in->size(); ++i) {
if (in->at(i).dtype != paddle::PaddleDType::INT64) {
dense_out[dense_idx] = in->at(i);
++dense_idx;
continue;
}
sparse_out[sparse_idx].lod.resize(in->at(i).lod.size());
for (size_t x = 0; x < sparse_out[sparse_idx].lod.size(); ++x) {
sparse_out[sparse_idx].lod[x].resize(in->at(i).lod[x].size());
std::copy(in->at(i).lod[x].begin(),
in->at(i).lod[x].end(),
sparse_out[sparse_idx].lod[x].begin());
}
sparse_out[sparse_idx].dtype = paddle::PaddleDType::FLOAT32;
sparse_out[sparse_idx].shape.push_back(
sparse_out[sparse_idx].lod[0].back());
sparse_out[sparse_idx].shape.push_back(EMBEDDING_SIZE);
sparse_out[sparse_idx].name = model_config->_feed_name[i];
sparse_out[sparse_idx].data.Resize(sparse_out[sparse_idx].lod[0].back() *
EMBEDDING_SIZE * sizeof(float));
float *dst_ptr = static_cast<float *>(sparse_out[sparse_idx].data.data());
for (int x = 0; x < sparse_out[sparse_idx].lod[0].back(); ++x) {
float *data_ptr = dst_ptr + x * EMBEDDING_SIZE;
memcpy(data_ptr,
values[cube_val_idx].buff.data(),
values[cube_val_idx].buff.size());
cube_val_idx++;
}
++sparse_idx;
}
TensorVector infer_in;
infer_in.insert(infer_in.end(), dense_out.begin(), dense_out.end());
infer_in.insert(infer_in.end(), sparse_out.begin(), sparse_out.end());
output_blob->SetBatchSize(batch_size);
output_blob->SetLogId(log_id);
VLOG(2) << "(logid=" << log_id << ") infer batch size: " << batch_size;
int64_t start = timeline.TimeStampUS();
if (InferManager::instance().infer(
engine_name().c_str(), &infer_in, out, batch_size)) {
LOG(ERROR) << "(logid=" << log_id
<< ") Failed do infer in fluid model: " << engine_name();
return -1;
}
int64_t end = timeline.TimeStampUS();
CopyBlobInfo(input_blob, output_blob);
AddBlobInfo(output_blob, cube_start);
AddBlobInfo(output_blob, cube_end);
AddBlobInfo(output_blob, start);
AddBlobInfo(output_blob, end);
return 0;
}
DEFINE_OP(GeneralDistKVInferOp);
......
core/general-server/op/general_dist_kv_quant_infer_op.cpp
浏览文件 @ fcd66ef4
......@@ -188,21 +188,6 @@ int GeneralDistKVQuantInferOp::inference() {
VLOG(2) << "(logid=" << log_id << ") infer batch size: " << batch_size;
Timer timeline;
int64_t start = timeline.TimeStampUS();
timeline.Start();
if (InferManager::instance().infer(
engine_name().c_str(), &infer_in, out, batch_size)) {
LOG(ERROR) << "(logid=" << log_id
<< ") Failed do infer in fluid model: " << engine_name();
return -1;
}
int64_t end = timeline.TimeStampUS();
CopyBlobInfo(input_blob, output_blob);
AddBlobInfo(output_blob, start);
AddBlobInfo(output_blob, end);
return 0;
}
DEFINE_OP(GeneralDistKVQuantInferOp);
......
core/general-server/op/general_infer_op.cpp
浏览文件 @ fcd66ef4
......@@ -44,45 +44,11 @@ int GeneralInferOp::inference() {
<< pre_node_names.size();
return -1;
}
const std::string pre_name = pre_node_names[0];
const GeneralBlob *input_blob = get_depend_argument<GeneralBlob>(pre_name);
uint64_t log_id = input_blob->GetLogId();
VLOG(2) << "(logid=" << log_id << ") Get precedent op name: " << pre_name;
GeneralBlob *output_blob = mutable_data<GeneralBlob>();
output_blob->SetLogId(log_id);
if (!input_blob) {
LOG(ERROR) << "(logid=" << log_id
<< ") Failed mutable depended argument, op:" << pre_name;
return -1;
}
const TensorVector *in = &input_blob->tensor_vector;
TensorVector *out = &output_blob->tensor_vector;
int batch_size = input_blob->_batch_size;
VLOG(2) << "(logid=" << log_id << ") input batch size: " << batch_size;
output_blob->_batch_size = batch_size;
VLOG(2) << "(logid=" << log_id << ") infer batch size: " << batch_size;
Timer timeline;
int64_t start = timeline.TimeStampUS();
timeline.Start();
if (InferManager::instance().infer(
engine_name().c_str(), in, out, batch_size)) {
LOG(ERROR) << "(logid=" << log_id
<< ") Failed do infer in fluid model: " << engine_name().c_str();
engine_name().c_str())) {
return -1;
}
int64_t end = timeline.TimeStampUS();
CopyBlobInfo(input_blob, output_blob);
AddBlobInfo(output_blob, start);
AddBlobInfo(output_blob, end);
std::cout << "Infer Success" << std::endl;
return 0;
}
DEFINE_OP(GeneralInferOp);
......
core/general-server/op/general_reader_op.cpp
浏览文件 @ fcd66ef4
......@@ -20,6 +20,7 @@
#include "core/general-server/op/general_infer_helper.h"
#include "core/predictor/framework/infer.h"
#include "core/predictor/framework/memory.h"
#include "core/predictor/framework/resource.h"
#include "core/util/include/timer.h"
namespace baidu {
......@@ -32,6 +33,7 @@ using baidu::paddle_serving::predictor::general_model::Tensor;
using baidu::paddle_serving::predictor::general_model::Request;
using baidu::paddle_serving::predictor::general_model::FeedInst;
using baidu::paddle_serving::predictor::PaddleGeneralModelConfig;
using baidu::paddle_serving::predictor::InferManager;
int conf_check(const Request *req,
const std::shared_ptr<PaddleGeneralModelConfig> &model_config) {
......@@ -71,75 +73,34 @@ int conf_check(const Request *req,
int GeneralReaderOp::inference() {
// reade request from client
const Request *req = dynamic_cast<const Request *>(get_request_message());
// TODO: only support one engine here
std::string engine_name = "general_infer_0";
const Request *req = dynamic_cast<const Request*>(get_request_message());
uint64_t log_id = req->log_id();
int input_var_num = 0;
std::vector<int64_t> elem_type;
std::vector<int64_t> elem_size;
std::vector<int64_t> capacity;
GeneralBlob *res = mutable_data<GeneralBlob>();
TensorVector *out = &res->tensor_vector;
res->SetLogId(log_id);
if (!res) {
LOG(ERROR) << "(logid=" << log_id
<< ") Failed get op tls reader object output";
}
Timer timeline;
int64_t start = timeline.TimeStampUS();
int var_num = req->insts(0).tensor_array_size();
VLOG(2) << "(logid=" << log_id << ") var num: " << var_num;
VLOG(2) << "(logid=" << log_id
<< ") start to call load general model_conf op";
baidu::paddle_serving::predictor::Resource &resource =
baidu::paddle_serving::predictor::Resource::instance();
VLOG(2) << "(logid=" << log_id << ") get resource pointer done.";
std::shared_ptr<PaddleGeneralModelConfig> model_config =
resource.get_general_model_config();
VLOG(2) << "(logid=" << log_id << ") print general model config done.";
// TODO(guru4elephant): how to do conditional check?
/*
int ret = conf_check(req, model_config);
if (ret != 0) {
LOG(ERROR) << "model conf of server:";
resource.print_general_model_config(model_config);
return 0;
}
*/
// package tensor
elem_type.resize(var_num);
elem_size.resize(var_num);
capacity.resize(var_num);
// prepare basic information for input
for (int i = 0; i < var_num; ++i) {
paddle::PaddleTensor lod_tensor;
elem_type[i] = req->insts(0).tensor_array(i).elem_type();
VLOG(2) << "var[" << i << "] has elem type: " << elem_type[i];
if (elem_type[i] == 0) { // int64
elem_size[i] = sizeof(int64_t);
lod_tensor.dtype = paddle::PaddleDType::INT64;
} else if (elem_type[i] == 1) {
elem_size[i] = sizeof(float);
lod_tensor.dtype = paddle::PaddleDType::FLOAT32;
} else if (elem_type[i] == 2) {
elem_size[i] = sizeof(int32_t);
lod_tensor.dtype = paddle::PaddleDType::INT32;
}
// implement lod tensor here
std::string tensor_name = model_config->_feed_name[i];
std::cout << "START Tensor Name: " <<tensor_name << std::endl;
auto lod_tensor = InferManager::instance().GetInputHandle(engine_name.c_str(), tensor_name.c_str());
std::cout << "PICK lod tensor. " << std::endl;
std::vector<std::vector<size_t>> lod;
std::vector<int> shape;
// get lod info here
if (req->insts(0).tensor_array(i).lod_size() > 0) {
VLOG(2) << "(logid=" << log_id << ") var[" << i << "] is lod_tensor";
lod_tensor.lod.resize(1);
lod.resize(1);
for (int k = 0; k < req->insts(0).tensor_array(i).lod_size(); ++k) {
lod_tensor.lod[0].push_back(req->insts(0).tensor_array(i).lod(k));
lod[0].push_back(req->insts(0).tensor_array(i).lod(k));
}
capacity[i] = 1;
for (int k = 0; k < req->insts(0).tensor_array(i).shape_size(); ++k) {
......@@ -147,103 +108,55 @@ int GeneralReaderOp::inference() {
VLOG(2) << "(logid=" << log_id << ") shape for var[" << i
<< "]: " << dim;
capacity[i] *= dim;
lod_tensor.shape.push_back(dim);
shape.push_back(dim);
}
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] is tensor, capacity: " << capacity[i];
} else {
}
else {
capacity[i] = 1;
for (int k = 0; k < req->insts(0).tensor_array(i).shape_size(); ++k) {
int dim = req->insts(0).tensor_array(i).shape(k);
VLOG(2) << "(logid=" << log_id << ") shape for var[" << i
<< "]: " << dim;
capacity[i] *= dim;
lod_tensor.shape.push_back(dim);
shape.push_back(dim);
}
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] is tensor, capacity: " << capacity[i];
}
lod_tensor.name = model_config->_feed_name[i];
out->push_back(lod_tensor);
}
// specify the memory needed for output tensor_vector
for (int i = 0; i < var_num; ++i) {
if (out->at(i).lod.size() == 1) {
int tensor_size = 0;
const Tensor &tensor = req->insts(0).tensor_array(i);
int data_len = 0;
if (tensor.int64_data_size() > 0) {
data_len = tensor.int64_data_size();
} else if (tensor.float_data_size() > 0) {
data_len = tensor.float_data_size();
} else if (tensor.int_data_size() > 0) {
data_len = tensor.int_data_size();
}
VLOG(2) << "(logid=" << log_id << ") tensor size for var[" << i
<< "]: " << data_len;
tensor_size += data_len;
int cur_len = out->at(i).lod[0].back();
VLOG(2) << "(logid=" << log_id << ") current len: " << cur_len;
int sample_len = 0;
if (tensor.shape_size() == 1) {
sample_len = data_len;
} else {
sample_len = tensor.shape(0);
}
VLOG(2) << "(logid=" << log_id << ") new len: " << cur_len + sample_len;
out->at(i).data.Resize(tensor_size * elem_size[i]);
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] is lod_tensor and len=" << out->at(i).lod[0].back();
} else {
out->at(i).data.Resize(capacity[i] * elem_size[i]);
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] is tensor and capacity=" << capacity[i];
}
}
// fill the data into output general_blob
for (int i = 0; i < var_num; ++i) {
if (elem_type[i] == 0) {
int64_t *dst_ptr = static_cast<int64_t *>(out->at(i).data.data());
VLOG(2) << "(logid=" << log_id << ") first element data in var[" << i
<< "] is " << req->insts(0).tensor_array(i).int64_data(0);
int offset = 0;
lod_tensor->SetLoD(lod);
lod_tensor->Reshape(shape);
std::cout << "FINI Set Lod and Reshape, and elem type: "<< elem_type[i] << std::endl;
// insert data here
if (req->insts(0).tensor_array(i).elem_type() == 0) {
// TODO: Copy twice here, can optimize
int elem_num = req->insts(0).tensor_array(i).int64_data_size();
std::vector<int64_t> data(elem_num);
int64_t* dst_ptr = data.data();
for (int k = 0; k < elem_num; ++k) {
dst_ptr[offset + k] = req->insts(0).tensor_array(i).int64_data(k);
dst_ptr[k] = req->insts(0).tensor_array(i).int64_data(k);
}
} else if (elem_type[i] == 1) {
float *dst_ptr = static_cast<float *>(out->at(i).data.data());
VLOG(2) << "(logid=" << log_id << ") first element data in var[" << i
<< "] is " << req->insts(0).tensor_array(i).float_data(0);
int offset = 0;
lod_tensor->CopyFromCpu(dst_ptr);
} else if(req->insts(0).tensor_array(i).elem_type() == 1) {
int elem_num = req->insts(0).tensor_array(i).float_data_size();
std::vector<float> data(elem_num);
float* dst_ptr = data.data();
for (int k = 0; k < elem_num; ++k) {
dst_ptr[offset + k] = req->insts(0).tensor_array(i).float_data(k);
dst_ptr[k] = req->insts(0).tensor_array(i).float_data(k);
}
} else if (elem_type[i] == 2) {
int32_t *dst_ptr = static_cast<int32_t *>(out->at(i).data.data());
VLOG(2) << "(logid=" << log_id << ") first element data in var[" << i
<< "] is " << req->insts(0).tensor_array(i).int_data(0);
int offset = 0;
lod_tensor->CopyFromCpu(dst_ptr);
} else if(req->insts(0).tensor_array(i).elem_type()== 2) {
int elem_num = req->insts(0).tensor_array(i).int_data_size();
std::vector<int32_t> data(elem_num);
int32_t* dst_ptr = data.data();
for (int k = 0; k < elem_num; ++k) {
dst_ptr[offset + k] = req->insts(0).tensor_array(i).int_data(k);
dst_ptr[k] = req->insts(0).tensor_array(i).int_data(k);
}
lod_tensor->CopyFromCpu(dst_ptr);
}
std::cout << "FINISH Tensor Name: " <<tensor_name << std::endl;
}
VLOG(2) << "(logid=" << log_id << ") output size: " << out->size();
timeline.Pause();
int64_t end = timeline.TimeStampUS();
res->p_size = 0;
res->_batch_size = 1;
AddBlobInfo(res, start);
AddBlobInfo(res, end);
VLOG(2) << "(logid=" << log_id << ") read data from client success";
return 0;
}
DEFINE_OP(GeneralReaderOp);
......
core/general-server/op/general_response_op.cpp
浏览文件 @ fcd66ef4
......@@ -40,160 +40,55 @@ using baidu::paddle_serving::predictor::InferManager;
using baidu::paddle_serving::predictor::PaddleGeneralModelConfig;
int GeneralResponseOp::inference() {
const std::vector<std::string> pre_node_names = pre_names();
VLOG(2) << "pre node names size: " << pre_node_names.size();
const GeneralBlob *input_blob;
uint64_t log_id =
get_depend_argument<GeneralBlob>(pre_node_names[0])->GetLogId();
const Request *req = dynamic_cast<const Request *>(get_request_message());
// response inst with only fetch_var_names
Response *res = mutable_data<Response>();
Timer timeline;
// double response_time = 0.0;
// timeline.Start();
int64_t start = timeline.TimeStampUS();
VLOG(2) << "(logid=" << log_id
<< ") start to call load general model_conf op";
baidu::paddle_serving::predictor::Resource &resource =
baidu::paddle_serving::predictor::Resource::instance();
VLOG(2) << "(logid=" << log_id << ") get resource pointer done.";
std::shared_ptr<PaddleGeneralModelConfig> model_config =
resource.get_general_model_config();
VLOG(2) << "(logid=" << log_id
<< ") max body size : " << brpc::fLU64::FLAGS_max_body_size;
std::vector<int> fetch_index;
fetch_index.resize(req->fetch_var_names_size());
std::vector<int> capacity(req->fetch_var_names_size(), 1);
std::string engine_name = "general_infer_0";
ModelOutput *output = res->add_outputs();
FetchInst *fetch_inst = output->add_insts();
FetchInst *fetch_p = output->mutable_insts(0);
std::vector<std::string> outs = InferManager::instance().GetOutputNames(engine_name.c_str());
for (int i = 0; i < req->fetch_var_names_size(); ++i) {
fetch_index[i] =
model_config->_fetch_alias_name_to_index[req->fetch_var_names(i)];
}
for (uint32_t pi = 0; pi < pre_node_names.size(); ++pi) {
const std::string &pre_name = pre_node_names[pi];
VLOG(2) << "(logid=" << log_id << ") pre names[" << pi << "]: " << pre_name
<< " (" << pre_node_names.size() << ")";
input_blob = get_depend_argument<GeneralBlob>(pre_name);
// fprintf(stderr, "input(%s) blob address %x\n", pre_names.c_str(),
// input_blob);
if (!input_blob) {
LOG(ERROR) << "(logid=" << log_id
<< ") Failed mutable depended argument, op: " << pre_name;
return -1;
Tensor *tensor = fetch_inst->add_tensor_array();
std::string tensor_name = outs[i];
auto lod_tensor = InferManager::instance().GetOutputHandle(engine_name.c_str(), tensor_name.c_str());
std::vector<int> shape = lod_tensor->shape();
for (int k = 0; k < shape.size(); ++k) {
capacity[i] *= shape[k];
tensor->add_shape(shape[k]);
}
const TensorVector *in = &input_blob->tensor_vector;
ModelOutput *output = res->add_outputs();
// To get the order of model return values
output->set_engine_name(pre_name);
FetchInst *fetch_inst = output->add_insts();
for (auto &idx : fetch_index) {
Tensor *tensor = fetch_inst->add_tensor_array();
if (model_config->_is_lod_fetch[idx]) {
VLOG(2) << "(logid=" << log_id << ") out[" << idx << "] "
<< model_config->_fetch_name[idx] << " is lod_tensor";
for (int k = 0; k < in->at(idx).shape.size(); ++k) {
VLOG(2) << "(logid=" << log_id << ") shape[" << k
<< "]: " << in->at(idx).shape[k];
tensor->add_shape(in->at(idx).shape[k]);
}
} else {
VLOG(2) << "(logid=" << log_id << ") out[" << idx << "] "
<< model_config->_fetch_name[idx] << " is tensor";
for (int k = 0; k < in->at(idx).shape.size(); ++k) {
VLOG(2) << "(logid=" << log_id << ") shape[" << k
<< "]: " << in->at(idx).shape[k];
tensor->add_shape(in->at(idx).shape[k]);
}
}
auto dtype = lod_tensor->type();
if (dtype == paddle::PaddleDType::INT64) {
std::vector<int64_t> datas(capacity[i]);
int64_t* data_ptr = datas.data();
lod_tensor->CopyToCpu(data_ptr);
google::protobuf::RepeatedField<int64_t> tmp_data(data_ptr, data_ptr + capacity[i]);
tensor->mutable_int64_data()->Swap(&tmp_data);
} else if (dtype == paddle::PaddleDType::FLOAT32) {
std::vector<float> datas(capacity[i]);
float* data_ptr = datas.data();
lod_tensor->CopyToCpu(data_ptr);
google::protobuf::RepeatedField<float> tmp_data(data_ptr, data_ptr + capacity[i]);
tensor->mutable_float_data()->Swap(&tmp_data);
} else if (dtype == paddle::PaddleDType::INT32) {
std::vector<int32_t> datas(capacity[i]);
int32_t* data_ptr = datas.data();
lod_tensor->CopyToCpu(data_ptr);
google::protobuf::RepeatedField<int32_t> tmp_data(data_ptr, data_ptr + capacity[i]);
tensor->mutable_int_data()->Swap(&tmp_data);
}
int var_idx = 0;
for (auto &idx : fetch_index) {
int cap = 1;
for (int j = 0; j < in->at(idx).shape.size(); ++j) {
cap *= in->at(idx).shape[j];
std::vector<std::vector<size_t>> lod = lod_tensor->lod();
if (lod.size() > 0) {
for (int j = 0; j < lod[0].size(); ++j) {
tensor->add_lod(lod[0][j]);
}
FetchInst *fetch_p = output->mutable_insts(0);
auto dtype = in->at(idx).dtype;
if (dtype == paddle::PaddleDType::INT64) {
VLOG(2) << "(logid=" << log_id << ") Prepare int64 var ["
<< model_config->_fetch_name[idx] << "].";
int64_t *data_ptr = static_cast<int64_t *>(in->at(idx).data.data());
// from
// https://stackoverflow.com/questions/15499641/copy-a-stdvector-to-a-repeated-field-from-protobuf-with-memcpy
// `Swap` method is faster than `{}` method.
google::protobuf::RepeatedField<int64_t> tmp_data(data_ptr,
data_ptr + cap);
fetch_p->mutable_tensor_array(var_idx)->mutable_int64_data()->Swap(
&tmp_data);
} else if (dtype == paddle::PaddleDType::FLOAT32) {
VLOG(2) << "(logid=" << log_id << ") Prepare float var ["
<< model_config->_fetch_name[idx] << "].";
float *data_ptr = static_cast<float *>(in->at(idx).data.data());
google::protobuf::RepeatedField<float> tmp_data(data_ptr,
data_ptr + cap);
fetch_p->mutable_tensor_array(var_idx)->mutable_float_data()->Swap(
&tmp_data);
} else if (dtype == paddle::PaddleDType::INT32) {
VLOG(2) << "(logid=" << log_id << ")Prepare int32 var ["
<< model_config->_fetch_name[idx] << "].";
int32_t *data_ptr = static_cast<int32_t *>(in->at(idx).data.data());
google::protobuf::RepeatedField<int32_t> tmp_data(data_ptr,
data_ptr + cap);
fetch_p->mutable_tensor_array(var_idx)->mutable_int_data()->Swap(
&tmp_data);
}
if (model_config->_is_lod_fetch[idx]) {
if (in->at(idx).lod.size() > 0) {
for (int j = 0; j < in->at(idx).lod[0].size(); ++j) {
fetch_p->mutable_tensor_array(var_idx)->add_lod(
in->at(idx).lod[0][j]);
}
}
}
VLOG(2) << "(logid=" << log_id << ") fetch var ["
<< model_config->_fetch_name[idx] << "] ready";
var_idx++;
}
}
if (req->profile_server()) {
int64_t end = timeline.TimeStampUS();
// TODO(barriery): multi-model profile_time.
// At present, only the response_op is multi-input, so here we get
// the profile_time by hard coding. It needs to be replaced with
// a more elegant way.
for (uint32_t pi = 0; pi < pre_node_names.size(); ++pi) {
input_blob = get_depend_argument<GeneralBlob>(pre_node_names[pi]);
VLOG(2) << "(logid=" << log_id
<< ") p size for input blob: " << input_blob->p_size;
int profile_time_idx = -1;
if (pi == 0) {
profile_time_idx = 0;
} else {
profile_time_idx = input_blob->p_size - 2;
}
for (; profile_time_idx < input_blob->p_size; ++profile_time_idx) {
res->add_profile_time(input_blob->time_stamp[profile_time_idx]);
}
}
// TODO(guru4elephant): find more elegant way to do this
res->add_profile_time(start);
res->add_profile_time(end);
}
return 0;
}
......
core/predictor/CMakeLists.txt
浏览文件 @ fcd66ef4
......@@ -12,13 +12,12 @@ set_source_files_properties(
${pdserving_srcs}
PROPERTIES
COMPILE_FLAGS "-Wno-strict-aliasing -Wno-unused-variable -Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
add_dependencies(pdserving protobuf boost brpc leveldb pdcodegen configure)
add_dependencies(pdserving protobuf boost brpc leveldb pdcodegen configure extern_paddle paddle_fluid)
if (WITH_TRT)
add_definitions(-DWITH_TRT)
endif()
target_link_libraries(pdserving
brpc protobuf boost leveldb configure -lpthread -lcrypto -lm -lrt -lssl -ldl -lz)
brpc protobuf boost leveldb configure -lpthread -lcrypto -lm -lrt -lssl -ldl -lz paddle_fluid ${paddle_depend_libs})
# install
install(TARGETS pdserving
RUNTIME DESTINATION ${PADDLE_SERVING_INSTALL_DIR}/bin
......
core/predictor/framework/infer.h
浏览文件 @ fcd66ef4
......@@ -20,10 +20,9 @@
#include <utility>
#include <vector>
#include "core/predictor/common/inner_common.h"
#include "core/predictor/framework/bsf.h"
#include "core/predictor/framework/factory.h"
#include "core/predictor/framework/infer_data.h"
#include "paddle_inference_api.h" // NOLINT
namespace baidu {
namespace paddle_serving {
namespace predictor {
......@@ -105,8 +104,8 @@ class InferEngine {
virtual int thrd_initialize() { return thrd_initialize_impl(); }
virtual int thrd_clear() { return thrd_clear_impl(); }
virtual int thrd_finalize() { return thrd_finalize_impl(); }
virtual int infer(const void* in, void* out, uint32_t batch_size = -1) {
return infer_impl1(in, out, batch_size);
virtual int infer() {
return infer_impl();
}
virtual int reload() = 0;
......@@ -120,11 +119,11 @@ class InferEngine {
virtual int thrd_finalize_impl() = 0;
virtual int thrd_clear_impl() = 0;
virtual int proc_finalize_impl() = 0;
virtual int infer_impl1(const void* in,
void* out,
uint32_t batch_size = -1) = 0;
virtual int infer_impl2(const BatchTensor& in,
BatchTensor& out) = 0; // NOLINT
virtual std::vector<std::string> GetInputNames() = 0;
virtual std::vector<std::string> GetOutputNames() = 0;
virtual std::unique_ptr<paddle_infer::Tensor> GetInputHandle(const std::string& name) = 0;
virtual std::unique_ptr<paddle_infer::Tensor> GetOutputHandle(const std::string& name) = 0;
virtual int infer_impl() = 0;
// end: framework inner call
};
......@@ -138,8 +137,6 @@ class ReloadableInferEngine : public InferEngine {
uint64_t last_revision;
};
typedef im::bsf::Task<Tensor, Tensor> TaskT;
virtual int load(const InferEngineCreationParams& params) = 0;
int proc_initialize_impl(const configure::EngineDesc& conf, bool version) {
......@@ -201,44 +198,11 @@ class ReloadableInferEngine : public InferEngine {
LOG(ERROR) << "Failed proc initialize impl";
return -1;
}
// init bsf framework
if (_infer_thread_num <= 0) {
return 0;
}
im::bsf::TaskExecutor<TaskT>::instance()->set_thread_init_fn(
boost::bind(&InferEngine::thrd_initialize_impl, this));
im::bsf::TaskExecutor<TaskT>::instance()->set_thread_reset_fn(
boost::bind(&InferEngine::thrd_clear_impl, this));
im::bsf::TaskExecutor<TaskT>::instance()->set_thread_callback_fn(
boost::bind(&InferEngine::infer_impl2, this, _1, _2));
im::bsf::TaskExecutor<TaskT>::instance()->set_batch_size(_infer_batch_size);
im::bsf::TaskExecutor<TaskT>::instance()->set_batch_align(
_infer_batch_align);
if (im::bsf::TaskExecutor<TaskT>::instance()->start(_infer_thread_num) !=
0) {
LOG(ERROR) << "Failed start bsf executor, threads:" << _infer_thread_num;
return -1;
}
LOG(WARNING) << "Enable batch schedule framework, thread_num:"
<< _infer_thread_num << ", batch_size:" << _infer_batch_size
<< ", enable_batch_align:" << _infer_batch_align;
return 0;
}
int infer(const void* in, void* out, uint32_t batch_size = -1) {
if (_infer_thread_num <= 0) {
return infer_impl1(in, out, batch_size);
}
im::bsf::TaskManager<Tensor, Tensor> task_manager;
task_manager.schedule(*(reinterpret_cast<const BatchTensor*>(in)),
*(reinterpret_cast<BatchTensor*>(out)));
task_manager.wait();
return 0;
int infer() {
return infer_impl();
}
int thrd_initialize() {
......@@ -263,10 +227,6 @@ class ReloadableInferEngine : public InferEngine {
return -1;
}
if (_infer_thread_num > 0) {
im::bsf::TaskExecutor<TaskT>::instance()->stop();
}
return 0;
}
......@@ -417,10 +377,6 @@ class DBReloadableInferEngine : public ReloadableInferEngine {
virtual int thrd_initialize_impl() {
// memory pool to be inited in non-serving-threads
if (MempoolWrapper::instance().thread_initialize() != 0) {
LOG(ERROR) << "Failed thread initialize mempool";
return -1;
}
ModelData<EngineCore>* md = new (std::nothrow) ModelData<EngineCore>;
if (!md || load_data(md, _infer_engine_params) != 0) {
......@@ -430,17 +386,12 @@ class DBReloadableInferEngine : public ReloadableInferEngine {
}
THREAD_SETSPECIFIC(_skey, md);
im::bsf::AutoMutex lock(_mutex);
_reload_vec.push_back(md);
return 0;
}
int thrd_clear_impl() {
// for non-serving-threads
if (MempoolWrapper::instance().thread_clear() != 0) {
LOG(ERROR) << "Failed thread clear mempool";
return -1;
}
return 0;
}
......@@ -538,11 +489,6 @@ class CloneDBReloadableInferEngine
}
virtual int thrd_initialize_impl() {
// memory pool to be inited in non-serving-threads
if (MempoolWrapper::instance().thread_initialize() != 0) {
LOG(ERROR) << "Failed thread initialize mempool";
return -1;
}
ModelData<EngineCore>* md = new (std::nothrow) ModelData<EngineCore>;
if (!md || load_data(md, _pd->cores[_pd->current_idx]) != 0) {
......@@ -552,7 +498,6 @@ class CloneDBReloadableInferEngine
}
THREAD_SETSPECIFIC(DBReloadableInferEngine<EngineCore>::_skey, md);
im::bsf::AutoMutex lock(DBReloadableInferEngine<EngineCore>::_mutex);
DBReloadableInferEngine<EngineCore>::_reload_vec.push_back(md);
return 0;
}
......@@ -571,25 +516,51 @@ class FluidInferEngine : public CloneDBReloadableInferEngine<FluidFamilyCore> {
public: // NOLINT
FluidInferEngine() {}
~FluidInferEngine() {}
std::vector<std::string> GetInputNames() {
FluidFamilyCore* core = DBReloadableInferEngine<FluidFamilyCore>::get_core();
if (!core || !core->get()) {
LOG(ERROR) << "Failed get fluid core in GetInputHandle()";
}
return core->GetInputNames();
}
std::vector<std::string> GetOutputNames() {
FluidFamilyCore* core = DBReloadableInferEngine<FluidFamilyCore>::get_core();
if (!core || !core->get()) {
LOG(ERROR) << "Failed get fluid core in GetInputHandle()";
}
return core->GetOutputNames();
}
std::unique_ptr<paddle_infer::Tensor> GetInputHandle(const std::string& name) {
FluidFamilyCore* core = DBReloadableInferEngine<FluidFamilyCore>::get_core();
if (!core || !core->get()) {
LOG(ERROR) << "Failed get fluid core in GetInputHandle()";
}
return core->GetInputHandle(name);
}
std::unique_ptr<paddle_infer::Tensor> GetOutputHandle(const std::string& name) {
FluidFamilyCore* core = DBReloadableInferEngine<FluidFamilyCore>::get_core();
if (!core || !core->get()) {
LOG(ERROR) << "Failed get fluid core in GetOutputHandle()";
}
return core->GetOutputHandle(name);
}
int infer_impl1(const void* in, void* out, uint32_t batch_size = -1) {
FluidFamilyCore* core =
DBReloadableInferEngine<FluidFamilyCore>::get_core();
int infer_impl() {
FluidFamilyCore* core = DBReloadableInferEngine<FluidFamilyCore>::get_core();
if (!core || !core->get()) {
LOG(ERROR) << "Failed get fluid core in infer_impl()";
return -1;
}
if (!core->Run(in, out)) {
if (!core->Run()) {
LOG(ERROR) << "Failed run fluid family core";
return -1;
}
return 0;
}
int infer_impl2(const BatchTensor& in, BatchTensor& out) { // NOLINT
return infer_impl1(&in, &out);
}
};
typedef FactoryPool<InferEngine> StaticInferFactory;
......@@ -715,13 +686,43 @@ class VersionedInferEngine : public InferEngine {
return _versions.begin()->second;
}
int infer(const void* in, void* out, uint32_t batch_size) {
int infer() {
InferEngine* engine = default_engine();
if (!engine) {
LOG(WARNING) << "fail to get default engine";
return -1;
}
return engine->infer(in, out, batch_size);
return engine->infer();
}
std::vector<std::string> GetInputNames() {
InferEngine* engine = default_engine();
if (!engine) {
LOG(WARNING) << "fail to get default engine";
}
return engine->GetInputNames();
}
std::vector<std::string> GetOutputNames() {
InferEngine* engine = default_engine();
if (!engine) {
LOG(WARNING) << "fail to get default engine";
}
return engine->GetOutputNames();
}
std::unique_ptr<paddle_infer::Tensor> GetInputHandle(const std::string& name) {
InferEngine* engine = default_engine();
if (!engine) {
LOG(WARNING) << "fail to get default engine";
}
return engine->GetInputHandle(name);
}
std::unique_ptr<paddle_infer::Tensor> GetOutputHandle(const std::string& name) {
InferEngine* engine = default_engine();
if (!engine) {
LOG(WARNING) << "fail to get default engine";
}
return engine->GetOutputHandle(name);
}
template <typename T>
......@@ -740,14 +741,45 @@ class VersionedInferEngine : public InferEngine {
}
// versioned inference interface
int infer(const void* in, void* out, uint32_t batch_size, uint64_t version) {
int infer(uint64_t version) {
auto iter = _versions.find(version);
if (iter == _versions.end()) {
LOG(ERROR) << "Not found version engine: " << version;
return -1;
}
return iter->second->infer(in, out, batch_size);
return iter->second->infer();
}
std::vector<std::string> GetInputNames(uint64_t version) {
auto iter = _versions.find(version);
if (iter == _versions.end()) {
LOG(ERROR) << "Not found version engine: " << version;
}
return iter->second->GetInputNames();
}
std::vector<std::string> GetOutputNames(uint64_t version) {
auto iter = _versions.find(version);
if (iter == _versions.end()) {
LOG(ERROR) << "Not found version engine: " << version;
}
return iter->second->GetOutputNames();
}
std::unique_ptr<paddle_infer::Tensor> GetInputHandle(uint64_t version, const std::string& name) {
auto iter = _versions.find(version);
if (iter == _versions.end()) {
LOG(ERROR) << "Not found version engine: " << version;
}
return iter->second->GetInputHandle(name);
}
std::unique_ptr<paddle_infer::Tensor> GetOutputHandle(uint64_t version, const std::string& name) {
auto iter = _versions.find(version);
if (iter == _versions.end()) {
LOG(ERROR) << "Not found version engine: " << version;
}
return iter->second->GetOutputHandle(name);
}
template <typename T>
......@@ -774,12 +806,9 @@ class VersionedInferEngine : public InferEngine {
int thrd_finalize_impl() { return -1; }
int thrd_clear_impl() { return -1; }
int proc_finalize_impl() { return -1; }
int infer_impl1(const void* in, void* out, uint32_t batch_size = -1) {
int infer_impl() {
return -1;
}
int infer_impl2(const BatchTensor& in, BatchTensor& out) { // NOLINT
return -1;
} // NOLINT
private:
boost::unordered_map<uint64_t, InferEngine*> _versions;
......@@ -877,16 +906,42 @@ class InferManager {
}
// Inference interface
int infer(const char* model_name,
const void* in,
void* out,
uint32_t batch_size = -1) {
int infer(const char* model_name) {
auto it = _map.find(model_name);
if (it == _map.end()) {
LOG(WARNING) << "Cannot find engine in map, model name:" << model_name;
return -1;
}
return it->second->infer(in, out, batch_size);
return it->second->infer();
}
std::vector<std::string> GetInputNames(const char* model_name) {
auto it = _map.find(model_name);
if (it == _map.end()) {
LOG(WARNING) << "Cannot find engine in map, model name:" << model_name;
}
return it->second->GetInputNames();
}
std::vector<std::string> GetOutputNames(const char* model_name) {
auto it = _map.find(model_name);
if (it == _map.end()) {
LOG(WARNING) << "Cannot find engine in map, model name:" << model_name;
}
return it->second->GetOutputNames();
}
std::unique_ptr<paddle_infer::Tensor> GetInputHandle(const char* model_name, const std::string& name) {
auto it = _map.find(model_name);
if (it == _map.end()) {
LOG(WARNING) << "Cannot find engine in map, model name:" << model_name;
}
return it->second->GetInputHandle(name);
}
std::unique_ptr<paddle_infer::Tensor> GetOutputHandle(const char* model_name, const std::string& name) {
auto it = _map.find(model_name);
if (it == _map.end()) {
LOG(WARNING) << "Cannot find engine in map, model name:" << model_name;
}
return it->second->GetOutputHandle(name);
}
template <typename T>
......@@ -907,18 +962,44 @@ class InferManager {
// Versioned inference interface
int infer(const char* model_name,
const void* in,
void* out,
uint32_t batch_size,
uint64_t version) {
auto it = _map.find(model_name);
if (it == _map.end()) {
LOG(WARNING) << "Cannot find engine in map, model name:" << model_name;
return -1;
}
return it->second->infer(in, out, batch_size, version);
return it->second->infer(version);
}
std::vector<std::string> GetInputNames(const char* model_name, uint64_t version) {
auto it = _map.find(model_name);
if (it == _map.end()) {
LOG(WARNING) << "Cannot find engine in map, model name:" << model_name;
}
return it->second->GetInputNames(version);
}
std::vector<std::string> GetOutputNames(const char* model_name, uint64_t version) {
auto it = _map.find(model_name);
if (it == _map.end()) {
LOG(WARNING) << "Cannot find engine in map, model name:" << model_name;
}
return it->second->GetOutputNames(version);
}
std::unique_ptr<paddle_infer::Tensor> GetInputHandle(const char* model_name, uint64_t version, const std::string& name) {
auto it = _map.find(model_name);
if (it == _map.end()) {
LOG(WARNING) << "Cannot find engine in map, model name:" << model_name;
}
return it->second->GetInputHandle(version, name);
}
std::unique_ptr<paddle_infer::Tensor> GetOutputHandle(const char* model_name, uint64_t version, const std::string& name) {
auto it = _map.find(model_name);
if (it == _map.end()) {
LOG(WARNING) << "Cannot find engine in map, model name:" << model_name;
}
return it->second->GetOutputHandle(version, name);
}
template <typename T>
T* get_core(const char* model_name, uint64_t version) {
auto it = _map.find(model_name);
......
paddle_inference/inferencer-fluid-cpu/include/fluid_cpu_engine.h
浏览文件 @ fcd66ef4
......@@ -28,8 +28,6 @@ namespace baidu {
namespace paddle_serving {
namespace fluid_cpu {
using configure::SigmoidConf;
class AutoLock {
public:
explicit AutoLock(pthread_mutex_t& mutex) : _mut(mutex) {
......@@ -57,31 +55,36 @@ class GlobalPaddleCreateMutex {
pthread_mutex_t _mut;
};
class GlobalSigmoidCreateMutex {
public:
pthread_mutex_t& mutex() { return _mut; }
static pthread_mutex_t& instance() {
static GlobalSigmoidCreateMutex gmutex;
return gmutex.mutex();
}
private:
GlobalSigmoidCreateMutex() { pthread_mutex_init(&_mut, NULL); }
pthread_mutex_t _mut;
};
using paddle_infer::Config;
using paddle_infer::Predictor;
using paddle_infer::Tensor;
using paddle_infer::CreatePredictor;
// data interface
class FluidFamilyCore {
public:
virtual ~FluidFamilyCore() {}
virtual bool Run(const void* in_data, void* out_data) {
if (!_core->Run(*(std::vector<paddle::PaddleTensor>*)in_data,
(std::vector<paddle::PaddleTensor>*)out_data)) {
virtual std::vector<std::string> GetInputNames() {
return _core->GetInputNames();
}
virtual std::unique_ptr<Tensor> GetInputHandle(const std::string& name) {
return _core->GetInputHandle(name);
}
virtual std::vector<std::string> GetOutputNames() {
return _core->GetOutputNames();
}
virtual std::unique_ptr<Tensor> GetOutputHandle(const std::string& name) {
return _core->GetOutputHandle(name);
}
virtual bool Run() {
if (!_core->Run()) {
LOG(ERROR) << "Failed call Run with paddle predictor";
return false;
}
return true;
}
......@@ -92,8 +95,8 @@ class FluidFamilyCore {
LOG(ERROR) << "origin paddle Predictor is null.";
return -1;
}
paddle::PaddlePredictor* p_predictor =
(paddle::PaddlePredictor*)origin_core;
Predictor* p_predictor =
(Predictor*)origin_core;
_core = p_predictor->Clone();
if (_core.get() == NULL) {
LOG(ERROR) << "fail to clone paddle predictor: " << origin_core;
......@@ -105,7 +108,7 @@ class FluidFamilyCore {
virtual void* get() { return _core.get(); }
protected:
std::unique_ptr<paddle::PaddlePredictor> _core;
std::shared_ptr<Predictor> _core;
};
// infer interface
......@@ -119,51 +122,19 @@ class FluidCpuAnalysisCore : public FluidFamilyCore {
return -1;
}
paddle::AnalysisConfig analysis_config;
analysis_config.SetParamsFile(data_path + "/__params__");
analysis_config.SetProgFile(data_path + "/__model__");
analysis_config.DisableGpu();
analysis_config.SetCpuMathLibraryNumThreads(1);
Config config;
config.SetParamsFile(data_path + "/__params__");
config.SetProgFile(data_path + "/__model__");
config.DisableGpu();
config.SetCpuMathLibraryNumThreads(1);
if (params.enable_memory_optimization()) {
analysis_config.EnableMemoryOptim();
config.EnableMemoryOptim();
}
analysis_config.SwitchSpecifyInputNames(true);
AutoLock lock(GlobalPaddleCreateMutex::instance());
_core =
paddle::CreatePaddlePredictor<paddle::AnalysisConfig>(analysis_config);
if (NULL == _core.get()) {
LOG(ERROR) << "create paddle predictor failed, path: " << data_path;
return -1;
}
VLOG(2) << "create paddle predictor sucess, path: " << data_path;
return 0;
}
};
class FluidCpuNativeCore : public FluidFamilyCore {
public:
int create(const predictor::InferEngineCreationParams& params) {
std::string data_path = params.get_path();
if (access(data_path.c_str(), F_OK) == -1) {
LOG(ERROR) << "create paddle predictor failed, path not exits: "
<< data_path;
return -1;
}
paddle::NativeConfig native_config;
native_config.param_file = data_path + "/__params__";
native_config.prog_file = data_path + "/__model__";
native_config.use_gpu = false;
native_config.device = 0;
native_config.fraction_of_gpu_memory = 0;
config.SwitchSpecifyInputNames(true);
AutoLock lock(GlobalPaddleCreateMutex::instance());
_core = paddle::CreatePaddlePredictor<paddle::NativeConfig,
paddle::PaddleEngineKind::kNative>(
native_config);
_core = CreatePredictor(config);
if (NULL == _core.get()) {
LOG(ERROR) << "create paddle predictor failed, path: " << data_path;
return -1;
......@@ -184,54 +155,24 @@ class FluidCpuAnalysisDirCore : public FluidFamilyCore {
return -1;
}
paddle::AnalysisConfig analysis_config;
analysis_config.SetModel(data_path);
analysis_config.DisableGpu();
analysis_config.SwitchSpecifyInputNames(true);
analysis_config.SetCpuMathLibraryNumThreads(1);
Config config;
config.SetModel(data_path);
config.DisableGpu();
config.SwitchSpecifyInputNames(true);
config.SetCpuMathLibraryNumThreads(1);
if (params.enable_memory_optimization()) {
analysis_config.EnableMemoryOptim();
config.EnableMemoryOptim();
}
if (params.enable_ir_optimization()) {
analysis_config.SwitchIrOptim(true);
config.SwitchIrOptim(true);
} else {
analysis_config.SwitchIrOptim(false);
config.SwitchIrOptim(false);
}
AutoLock lock(GlobalPaddleCreateMutex::instance());
_core =
paddle::CreatePaddlePredictor<paddle::AnalysisConfig>(analysis_config);
if (NULL == _core.get()) {
LOG(ERROR) << "create paddle predictor failed, path: " << data_path;
return -1;
}
VLOG(2) << "create paddle predictor sucess, path: " << data_path;
return 0;
}
};
class FluidCpuNativeDirCore : public FluidFamilyCore {
public:
int create(const predictor::InferEngineCreationParams& params) {
std::string data_path = params.get_path();
if (access(data_path.c_str(), F_OK) == -1) {
LOG(ERROR) << "create paddle predictor failed, path not exits: "
<< data_path;
return -1;
}
paddle::NativeConfig native_config;
native_config.model_dir = data_path;
native_config.use_gpu = false;
native_config.device = 0;
native_config.fraction_of_gpu_memory = 0;
AutoLock lock(GlobalPaddleCreateMutex::instance());
_core = paddle::CreatePaddlePredictor<paddle::NativeConfig,
paddle::PaddleEngineKind::kNative>(
native_config);
_core = CreatePredictor(config);
if (NULL == _core.get()) {
LOG(ERROR) << "create paddle predictor failed, path: " << data_path;
return -1;
......@@ -323,214 +264,6 @@ class Parameter {
float* _params;
};
class SigmoidModel {
public:
~SigmoidModel() {}
int load(const char* sigmoid_w_file,
const char* sigmoid_b_file,
float exp_max,
float exp_min) {
AutoLock lock(GlobalSigmoidCreateMutex::instance());
if (0 != _sigmoid_w.init(2, 1, sigmoid_w_file) || 0 != _sigmoid_w.load()) {
LOG(ERROR) << "load params sigmoid_w failed.";
return -1;
}
VLOG(2) << "load sigmoid_w [" << _sigmoid_w._params[0] << "] ["
<< _sigmoid_w._params[1] << "].";
if (0 != _sigmoid_b.init(2, 1, sigmoid_b_file) || 0 != _sigmoid_b.load()) {
LOG(ERROR) << "load params sigmoid_b failed.";
return -1;
}
VLOG(2) << "load sigmoid_b [" << _sigmoid_b._params[0] << "] ["
<< _sigmoid_b._params[1] << "].";
_exp_max_input = exp_max;
_exp_min_input = exp_min;
return 0;
}
int softmax(float x, double& o) { // NOLINT
float _y0 = x * _sigmoid_w._params[0] + _sigmoid_b._params[0];
float _y1 = x * _sigmoid_w._params[1] + _sigmoid_b._params[1];
_y0 = (_y0 > _exp_max_input)
? _exp_max_input
: ((_y0 < _exp_min_input) ? _exp_min_input : _y0);
_y1 = (_y1 > _exp_max_input)
? _exp_max_input
: ((_y1 < _exp_min_input) ? _exp_min_input : _y1);
o = 1.0f / (1.0f + exp(_y0 - _y1));
return 0;
}
public:
Parameter _sigmoid_w;
Parameter _sigmoid_b;
float _exp_max_input;
float _exp_min_input;
};
class SigmoidFluidModel {
public:
int softmax(float x, double& o) { // NOLINT
return _sigmoid_core->softmax(x, o);
} // NOLINT
std::unique_ptr<SigmoidFluidModel> Clone() {
std::unique_ptr<SigmoidFluidModel> clone_model;
clone_model.reset(new SigmoidFluidModel());
clone_model->_sigmoid_core = _sigmoid_core;
clone_model->_fluid_core = _fluid_core->Clone();
return std::move(clone_model); // NOLINT
}
public:
std::unique_ptr<paddle::PaddlePredictor> _fluid_core;
std::shared_ptr<SigmoidModel> _sigmoid_core;
};
class FluidCpuWithSigmoidCore : public FluidFamilyCore {
public:
virtual ~FluidCpuWithSigmoidCore() {}
public:
int create(const predictor::InferEngineCreationParams& params) {
std::string model_path = params.get_path();
size_t pos = model_path.find_last_of("/\\");
std::string conf_path = model_path.substr(0, pos);
std::string conf_file = model_path.substr(pos);
configure::SigmoidConf conf;
if (configure::read_proto_conf(conf_path, conf_file, &conf) != 0) {
LOG(ERROR) << "failed load model path: " << model_path;
return -1;
}
_core.reset(new SigmoidFluidModel);
std::string fluid_model_data_path = conf.dnn_model_path();
predictor::InferEngineCreationParams new_params(params);
new_params.set_path(fluid_model_data_path);
int ret = load_fluid_model(new_params);
if (ret < 0) {
LOG(ERROR) << "fail to load fluid model.";
return -1;
}
const char* sigmoid_w_file = conf.sigmoid_w_file().c_str();
const char* sigmoid_b_file = conf.sigmoid_b_file().c_str();
float exp_max = conf.exp_max_input();
float exp_min = conf.exp_min_input();
_core->_sigmoid_core.reset(new SigmoidModel);
VLOG(2) << "create sigmoid core[" << _core->_sigmoid_core.get()
<< "], use count[" << _core->_sigmoid_core.use_count() << "].";
ret = _core->_sigmoid_core->load(
sigmoid_w_file, sigmoid_b_file, exp_max, exp_min);
if (ret < 0) {
LOG(ERROR) << "fail to load sigmoid model.";
return -1;
}
return 0;
}
virtual bool Run(const void* in_data, void* out_data) {
if (!_core->_fluid_core->Run(
*(std::vector<paddle::PaddleTensor>*)in_data,
(std::vector<paddle::PaddleTensor>*)out_data)) {
LOG(ERROR) << "Failed call Run with paddle predictor";
return false;
}
return true;
}
virtual int clone(SigmoidFluidModel* origin_core) {
if (origin_core == NULL) {
LOG(ERROR) << "origin paddle Predictor is null.";
return -1;
}
_core = origin_core->Clone();
if (_core.get() == NULL) {
LOG(ERROR) << "fail to clone paddle predictor: " << origin_core;
return -1;
}
VLOG(2) << "clone sigmoid core[" << _core->_sigmoid_core.get()
<< "] use count[" << _core->_sigmoid_core.use_count() << "].";
return 0;
}
virtual SigmoidFluidModel* get() { return _core.get(); }
virtual int load_fluid_model(
const predictor::InferEngineCreationParams& params) = 0;
int softmax(float x, double& o) { // NOLINT
return _core->_sigmoid_core->softmax(x, o);
}
protected:
std::unique_ptr<SigmoidFluidModel> _core; // NOLINT
};
class FluidCpuNativeDirWithSigmoidCore : public FluidCpuWithSigmoidCore {
public:
int load_fluid_model(const predictor::InferEngineCreationParams& params) {
std::string data_path = params.get_path();
if (access(data_path.c_str(), F_OK) == -1) {
LOG(ERROR) << "create paddle predictor failed, path not exits: "
<< data_path;
return -1;
}
paddle::NativeConfig native_config;
native_config.model_dir = data_path;
native_config.use_gpu = false;
native_config.device = 0;
native_config.fraction_of_gpu_memory = 0;
AutoLock lock(GlobalPaddleCreateMutex::instance());
_core->_fluid_core =
paddle::CreatePaddlePredictor<paddle::NativeConfig,
paddle::PaddleEngineKind::kNative>(
native_config);
if (NULL == _core.get()) {
LOG(ERROR) << "create paddle predictor failed, path: " << data_path;
return -1;
}
VLOG(2) << "create paddle predictor sucess, path: " << data_path;
return 0;
}
};
class FluidCpuAnalysisDirWithSigmoidCore : public FluidCpuWithSigmoidCore {
public:
int load_fluid_model(const predictor::InferEngineCreationParams& params) {
std::string data_path = params.get_path();
if (access(data_path.c_str(), F_OK) == -1) {
LOG(ERROR) << "create paddle predictor failed, path not exits: "
<< data_path;
return -1;
}
paddle::AnalysisConfig analysis_config;
analysis_config.SetModel(data_path);
analysis_config.DisableGpu();
analysis_config.SwitchSpecifyInputNames(true);
analysis_config.SetCpuMathLibraryNumThreads(1);
if (params.enable_memory_optimization()) {
analysis_config.EnableMemoryOptim();
}
AutoLock lock(GlobalPaddleCreateMutex::instance());
_core->_fluid_core =
paddle::CreatePaddlePredictor<paddle::AnalysisConfig>(analysis_config);
if (NULL == _core.get()) {
LOG(ERROR) << "create paddle predictor failed, path: " << data_path;
return -1;
}
VLOG(2) << "create paddle predictor sucess, path: " << data_path;
return 0;
}
};
} // namespace fluid_cpu
} // namespace paddle_serving
} // namespace baidu
paddle_inference/inferencer-fluid-cpu/src/fluid_cpu_engine.cpp
浏览文件 @ fcd66ef4
......@@ -30,28 +30,6 @@ REGIST_FACTORY_OBJECT_IMPL_WITH_NAME(
::baidu::paddle_serving::predictor::InferEngine,
"FLUID_CPU_ANALYSIS_DIR");
REGIST_FACTORY_OBJECT_IMPL_WITH_NAME(
::baidu::paddle_serving::predictor::FluidInferEngine<
FluidCpuAnalysisDirWithSigmoidCore>,
::baidu::paddle_serving::predictor::InferEngine,
"FLUID_CPU_ANALYSIS_DIR_SIGMOID");
REGIST_FACTORY_OBJECT_IMPL_WITH_NAME(
::baidu::paddle_serving::predictor::FluidInferEngine<FluidCpuNativeCore>,
::baidu::paddle_serving::predictor::InferEngine,
"FLUID_CPU_NATIVE");
REGIST_FACTORY_OBJECT_IMPL_WITH_NAME(
::baidu::paddle_serving::predictor::FluidInferEngine<FluidCpuNativeDirCore>,
::baidu::paddle_serving::predictor::InferEngine,
"FLUID_CPU_NATIVE_DIR");
REGIST_FACTORY_OBJECT_IMPL_WITH_NAME(
::baidu::paddle_serving::predictor::FluidInferEngine<
FluidCpuNativeDirWithSigmoidCore>,
::baidu::paddle_serving::predictor::InferEngine,
"FLUID_CPU_NATIVE_DIR_SIGMOID");
} // namespace fluid_cpu
} // namespace paddle_serving
} // namespace baidu
paddle_inference/inferencer-fluid-gpu/include/fluid_gpu_engine.h
浏览文件 @ fcd66ef4
......@@ -61,31 +61,36 @@ class GlobalPaddleCreateMutex {
pthread_mutex_t _mut;
};
class GlobalSigmoidCreateMutex {
public:
pthread_mutex_t& mutex() { return _mut; }
static pthread_mutex_t& instance() {
static GlobalSigmoidCreateMutex gmutex;
return gmutex.mutex();
}
private:
GlobalSigmoidCreateMutex() { pthread_mutex_init(&_mut, NULL); }
pthread_mutex_t _mut;
};
using paddle_infer::Config;
using paddle_infer::Predictor;
using paddle_infer::Tensor;
using paddle_infer::CreatePredictor;
// data interface
class FluidFamilyCore {
public:
virtual ~FluidFamilyCore() {}
virtual bool Run(const void* in_data, void* out_data) {
if (!_core->Run(*(std::vector<paddle::PaddleTensor>*)in_data,
(std::vector<paddle::PaddleTensor>*)out_data)) {
virtual std::vector<std::string> GetInputNames() {
return _core->GetInputNames();
}
virtual std::unique_ptr<Tensor> GetInputHandle(const std::string& name) {
return _core->GetInputHandle(name);
}
virtual std::vector<std::string> GetOutputNames() {
return _core->GetOutputNames();
}
virtual std::unique_ptr<Tensor> GetOutputHandle(const std::string& name) {
return _core->GetOutputHandle(name);
}
virtual bool Run() {
if (!_core->Run()) {
LOG(ERROR) << "Failed call Run with paddle predictor";
return false;
}
return true;
}
......@@ -96,8 +101,8 @@ class FluidFamilyCore {
LOG(ERROR) << "origin paddle Predictor is null.";
return -1;
}
paddle::PaddlePredictor* p_predictor =
(paddle::PaddlePredictor*)origin_core;
Predictor* p_predictor =
(Predictor*)origin_core;
_core = p_predictor->Clone();
if (_core.get() == NULL) {
LOG(ERROR) << "fail to clone paddle predictor: " << origin_core;
......@@ -109,7 +114,7 @@ class FluidFamilyCore {
virtual void* get() { return _core.get(); }
protected:
std::unique_ptr<paddle::PaddlePredictor> _core;
std::shared_ptr<Predictor> _core;
};
// infer interface
......@@ -123,51 +128,19 @@ class FluidGpuAnalysisCore : public FluidFamilyCore {
return -1;
}
paddle::AnalysisConfig analysis_config;
analysis_config.SetParamsFile(data_path + "/__params__");
analysis_config.SetProgFile(data_path + "/__model__");
analysis_config.EnableUseGpu(100, FLAGS_gpuid);
analysis_config.SetCpuMathLibraryNumThreads(1);
Config config;
config.SetParamsFile(data_path + "/__params__");
config.SetProgFile(data_path + "/__model__");
config.EnableUseGpu(100, FLAGS_gpuid);
config.SetCpuMathLibraryNumThreads(1);
if (params.enable_memory_optimization()) {
analysis_config.EnableMemoryOptim();
config.EnableMemoryOptim();
}
analysis_config.SwitchSpecifyInputNames(true);
config.SwitchSpecifyInputNames(true);
AutoLock lock(GlobalPaddleCreateMutex::instance());
_core =
paddle::CreatePaddlePredictor<paddle::AnalysisConfig>(analysis_config);
if (NULL == _core.get()) {
LOG(ERROR) << "create paddle predictor failed, path: " << data_path;
return -1;
}
VLOG(2) << "create paddle predictor sucess, path: " << data_path;
return 0;
}
};
class FluidGpuNativeCore : public FluidFamilyCore {
public:
int create(const predictor::InferEngineCreationParams& params) {
std::string data_path = params.get_path();
if (access(data_path.c_str(), F_OK) == -1) {
LOG(ERROR) << "create paddle predictor failed, path not exits: "
<< data_path;
return -1;
}
paddle::NativeConfig native_config;
native_config.param_file = data_path + "/__params__";
native_config.prog_file = data_path + "/__model__";
native_config.use_gpu = true;
native_config.fraction_of_gpu_memory = 0.01;
native_config.device = FLAGS_gpuid;
AutoLock lock(GlobalPaddleCreateMutex::instance());
_core = paddle::CreatePaddlePredictor<paddle::NativeConfig,
paddle::PaddleEngineKind::kNative>(
native_config);
_core = CreatePredictor(config);
if (NULL == _core.get()) {
LOG(ERROR) << "create paddle predictor failed, path: " << data_path;
return -1;
......@@ -188,81 +161,39 @@ class FluidGpuAnalysisDirCore : public FluidFamilyCore {
return -1;
}
paddle::AnalysisConfig analysis_config;
analysis_config.SetModel(data_path);
analysis_config.EnableUseGpu(1500, FLAGS_gpuid);
analysis_config.SwitchSpecifyInputNames(true);
analysis_config.SetCpuMathLibraryNumThreads(1);
Config config;
config.SetModel(data_path);
config.EnableUseGpu(1500, FLAGS_gpuid);
config.SwitchSpecifyInputNames(true);
config.SetCpuMathLibraryNumThreads(1);
if (params.enable_memory_optimization()) {
analysis_config.EnableMemoryOptim();
config.EnableMemoryOptim();
}
#if 0 // todo: support flexible shape
int min_seq_len = 1;
int max_seq_len = 512;
int opt_seq_len = 128;
int head_number = 12;
int batch = 50;
std::vector<int> min_in_shape = {batch, min_seq_len, 1};
std::vector<int> max_in_shape = {batch, max_seq_len, 1};
std::vector<int> opt_in_shape = {batch, opt_seq_len, 1};
std::string input1_name = "src_text_a_ids";
std::string input2_name = "pos_text_a_ids";
std::string input3_name = "sent_text_a_ids";
std::string input4_name = "stack_0.tmp_0";
std::map<std::string, std::vector<int>> min_input_shape = {
{input1_name, min_in_shape},
{input2_name, min_in_shape},
{input3_name, min_in_shape},
{input4_name, {batch, head_number, min_seq_len, min_seq_len}},
};
std::map<std::string, std::vector<int>> max_input_shape = {
{input1_name, max_in_shape},
{input2_name, max_in_shape},
{input3_name, max_in_shape},
{input4_name, {batch, head_number, max_seq_len, max_seq_len}},
};
std::map<std::string, std::vector<int>> opt_input_shape = {
{input1_name, opt_in_shape},
{input2_name, opt_in_shape},
{input3_name, opt_in_shape},
{input4_name, {batch, head_number, opt_seq_len, opt_seq_len}},
};
analysis_config.SetTRTDynamicShapeInfo(
min_input_shape, max_input_shape, opt_input_shape);
#endif
int max_batch = 32;
int min_subgraph_size = 3;
if (params.use_trt()) {
analysis_config.EnableTensorRtEngine(
config.EnableTensorRtEngine(
1 << 20,
max_batch,
min_subgraph_size,
paddle::AnalysisConfig::Precision::kFloat32,
Config::Precision::kFloat32,
false,
false);
LOG(INFO) << "create TensorRT predictor";
} else {
if (params.enable_memory_optimization()) {
analysis_config.EnableMemoryOptim();
config.EnableMemoryOptim();
}
if (params.enable_ir_optimization()) {
analysis_config.SwitchIrOptim(true);
config.SwitchIrOptim(true);
} else {
analysis_config.SwitchIrOptim(false);
config.SwitchIrOptim(false);
}
}
AutoLock lock(GlobalPaddleCreateMutex::instance());
_core =
paddle::CreatePaddlePredictor<paddle::AnalysisConfig>(analysis_config);
_core = CreatePredictor(config);
if (NULL == _core.get()) {
LOG(ERROR) << "create paddle predictor failed, path: " << data_path;
return -1;
......@@ -273,34 +204,6 @@ class FluidGpuAnalysisDirCore : public FluidFamilyCore {
}
};
class FluidGpuNativeDirCore : public FluidFamilyCore {
public:
int create(const predictor::InferEngineCreationParams& params) {
std::string data_path = params.get_path();
if (access(data_path.c_str(), F_OK) == -1) {
LOG(ERROR) << "create paddle predictor failed, path not exits: "
<< data_path;
return -1;
}
paddle::NativeConfig native_config;
native_config.model_dir = data_path;
native_config.use_gpu = true;
native_config.fraction_of_gpu_memory = 0.01;
native_config.device = FLAGS_gpuid;
AutoLock lock(GlobalPaddleCreateMutex::instance());
_core = paddle::CreatePaddlePredictor<paddle::NativeConfig,
paddle::PaddleEngineKind::kNative>(
native_config);
if (NULL == _core.get()) {
LOG(ERROR) << "create paddle predictor failed, path: " << data_path;
return -1;
}
VLOG(2) << "create paddle predictor sucess, path: " << data_path;
return 0;
}
};
class Parameter {
public:
......@@ -383,214 +286,6 @@ class Parameter {
float* _params;
};
class SigmoidModel {
public:
~SigmoidModel() {}
int load(const char* sigmoid_w_file,
const char* sigmoid_b_file,
float exp_max,
float exp_min) {
AutoLock lock(GlobalSigmoidCreateMutex::instance());
if (0 != _sigmoid_w.init(2, 1, sigmoid_w_file) || 0 != _sigmoid_w.load()) {
LOG(ERROR) << "load params sigmoid_w failed.";
return -1;
}
VLOG(2) << "load sigmoid_w [" << _sigmoid_w._params[0] << "] ["
<< _sigmoid_w._params[1] << "].";
if (0 != _sigmoid_b.init(2, 1, sigmoid_b_file) || 0 != _sigmoid_b.load()) {
LOG(ERROR) << "load params sigmoid_b failed.";
return -1;
}
VLOG(2) << "load sigmoid_b [" << _sigmoid_b._params[0] << "] ["
<< _sigmoid_b._params[1] << "].";
_exp_max_input = exp_max;
_exp_min_input = exp_min;
return 0;
}
int softmax(float x, double& o) { // NOLINT
float _y0 = x * _sigmoid_w._params[0] + _sigmoid_b._params[0];
float _y1 = x * _sigmoid_w._params[1] + _sigmoid_b._params[1];
_y0 = (_y0 > _exp_max_input)
? _exp_max_input
: ((_y0 < _exp_min_input) ? _exp_min_input : _y0);
_y1 = (_y1 > _exp_max_input)
? _exp_max_input
: ((_y1 < _exp_min_input) ? _exp_min_input : _y1);
o = 1.0f / (1.0f + exp(_y0 - _y1));
return 0;
}
public:
Parameter _sigmoid_w;
Parameter _sigmoid_b;
float _exp_max_input;
float _exp_min_input;
};
class SigmoidFluidModel {
public:
int softmax(float x, double& o) { // NOLINT
return _sigmoid_core->softmax(x, o);
} // NOLINT
std::unique_ptr<SigmoidFluidModel> Clone() {
std::unique_ptr<SigmoidFluidModel> clone_model;
clone_model.reset(new SigmoidFluidModel());
clone_model->_sigmoid_core = _sigmoid_core;
clone_model->_fluid_core = _fluid_core->Clone();
return std::move(clone_model);
}
public:
std::unique_ptr<paddle::PaddlePredictor> _fluid_core;
std::shared_ptr<SigmoidModel> _sigmoid_core;
};
class FluidGpuWithSigmoidCore : public FluidFamilyCore {
public:
virtual ~FluidGpuWithSigmoidCore() {}
public:
int create(const predictor::InferEngineCreationParams& params) {
std::string model_path = params.get_path();
size_t pos = model_path.find_last_of("/\\");
std::string conf_path = model_path.substr(0, pos);
std::string conf_file = model_path.substr(pos);
configure::SigmoidConf conf;
if (configure::read_proto_conf(conf_path, conf_file, &conf) != 0) {
LOG(ERROR) << "failed load model path: " << model_path;
return -1;
}
_core.reset(new SigmoidFluidModel);
std::string fluid_model_data_path = conf.dnn_model_path();
predictor::InferEngineCreationParams new_params(params);
new_params.set_path(fluid_model_data_path);
int ret = load_fluid_model(new_params);
if (ret < 0) {
LOG(ERROR) << "fail to load fluid model.";
return -1;
}
const char* sigmoid_w_file = conf.sigmoid_w_file().c_str();
const char* sigmoid_b_file = conf.sigmoid_b_file().c_str();
float exp_max = conf.exp_max_input();
float exp_min = conf.exp_min_input();
_core->_sigmoid_core.reset(new SigmoidModel);
LOG(INFO) << "create sigmoid core[" << _core->_sigmoid_core.get()
<< "], use count[" << _core->_sigmoid_core.use_count() << "].";
ret = _core->_sigmoid_core->load(
sigmoid_w_file, sigmoid_b_file, exp_max, exp_min);
if (ret < 0) {
LOG(ERROR) << "fail to load sigmoid model.";
return -1;
}
return 0;
}
virtual bool Run(const void* in_data, void* out_data) {
if (!_core->_fluid_core->Run(
*(std::vector<paddle::PaddleTensor>*)in_data,
(std::vector<paddle::PaddleTensor>*)out_data)) {
LOG(ERROR) << "Failed call Run with paddle predictor";
return false;
}
return true;
}
virtual int clone(SigmoidFluidModel* origin_core) {
if (origin_core == NULL) {
LOG(ERROR) << "origin paddle Predictor is null.";
return -1;
}
_core = origin_core->Clone();
if (_core.get() == NULL) {
LOG(ERROR) << "fail to clone paddle predictor: " << origin_core;
return -1;
}
LOG(INFO) << "clone sigmoid core[" << _core->_sigmoid_core.get()
<< "] use count[" << _core->_sigmoid_core.use_count() << "].";
return 0;
}
virtual SigmoidFluidModel* get() { return _core.get(); }
virtual int load_fluid_model(
const predictor::InferEngineCreationParams& params) = 0;
int softmax(float x, double& o) { // NOLINT
return _core->_sigmoid_core->softmax(x, o);
}
protected:
std::unique_ptr<SigmoidFluidModel> _core;
};
class FluidGpuNativeDirWithSigmoidCore : public FluidGpuWithSigmoidCore {
public:
int load_fluid_model(const predictor::InferEngineCreationParams& params) {
std::string data_path = params.get_path();
if (access(data_path.c_str(), F_OK) == -1) {
LOG(ERROR) << "create paddle predictor failed, path not exits: "
<< data_path;
return -1;
}
paddle::NativeConfig native_config;
native_config.model_dir = data_path;
native_config.use_gpu = true;
native_config.fraction_of_gpu_memory = 0.01;
native_config.device = FLAGS_gpuid;
AutoLock lock(GlobalPaddleCreateMutex::instance());
_core->_fluid_core =
paddle::CreatePaddlePredictor<paddle::NativeConfig,
paddle::PaddleEngineKind::kNative>(
native_config);
if (NULL == _core.get()) {
LOG(ERROR) << "create paddle predictor failed, path: " << data_path;
return -1;
}
VLOG(2) << "create paddle predictor sucess, path: " << data_path;
return 0;
}
};
class FluidGpuAnalysisDirWithSigmoidCore : public FluidGpuWithSigmoidCore {
public:
int load_fluid_model(const predictor::InferEngineCreationParams& params) {
std::string data_path = params.get_path();
if (access(data_path.c_str(), F_OK) == -1) {
LOG(ERROR) << "create paddle predictor failed, path not exits: "
<< data_path;
return -1;
}
paddle::AnalysisConfig analysis_config;
analysis_config.SetModel(data_path);
analysis_config.EnableUseGpu(100, FLAGS_gpuid);
analysis_config.SwitchSpecifyInputNames(true);
analysis_config.SetCpuMathLibraryNumThreads(1);
if (params.enable_memory_optimization()) {
analysis_config.EnableMemoryOptim();
}
AutoLock lock(GlobalPaddleCreateMutex::instance());
_core->_fluid_core =
paddle::CreatePaddlePredictor<paddle::AnalysisConfig>(analysis_config);
if (NULL == _core.get()) {
LOG(ERROR) << "create paddle predictor failed, path: " << data_path;
return -1;
}
VLOG(2) << "create paddle predictor sucess, path: " << data_path;
return 0;
}
};
} // namespace fluid_gpu
} // namespace paddle_serving
} // namespace baidu
paddle_inference/inferencer-fluid-gpu/src/fluid_gpu_engine.cpp
浏览文件 @ fcd66ef4
......@@ -32,28 +32,6 @@ REGIST_FACTORY_OBJECT_IMPL_WITH_NAME(
::baidu::paddle_serving::predictor::InferEngine,
"FLUID_GPU_ANALYSIS_DIR");
REGIST_FACTORY_OBJECT_IMPL_WITH_NAME(
::baidu::paddle_serving::predictor::FluidInferEngine<
FluidGpuAnalysisDirWithSigmoidCore>,
::baidu::paddle_serving::predictor::InferEngine,
"FLUID_GPU_ANALYSIS_DIR_SIGMOID");
REGIST_FACTORY_OBJECT_IMPL_WITH_NAME(
::baidu::paddle_serving::predictor::FluidInferEngine<FluidGpuNativeCore>,
::baidu::paddle_serving::predictor::InferEngine,
"FLUID_GPU_NATIVE");
REGIST_FACTORY_OBJECT_IMPL_WITH_NAME(
::baidu::paddle_serving::predictor::FluidInferEngine<FluidGpuNativeDirCore>,
::baidu::paddle_serving::predictor::InferEngine,
"FLUID_GPU_NATIVE_DIR");
REGIST_FACTORY_OBJECT_IMPL_WITH_NAME(
::baidu::paddle_serving::predictor::FluidInferEngine<
FluidGpuNativeDirWithSigmoidCore>,
::baidu::paddle_serving::predictor::InferEngine,
"FLUID_GPU_NATIVE_DIR_SIGMOID");
} // namespace fluid_gpu
} // namespace paddle_serving
} // namespace baidu
python/examples/fit_a_line/local_train.py
浏览文件 @ fcd66ef4
......@@ -16,7 +16,7 @@
import sys
import paddle
import paddle.fluid as fluid
paddle.enable_static()
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.uci_housing.train(), buf_size=500),
......
Markdown is supported
0% 或 .
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 或