natas_hw
/
openvino
0
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

publish master branch snapshot, revision cdcab9d7ab48ffb0ee5629fabbfa06cb45debd9b

This commit is contained in:
Alexey Suhov 2020-04-15 19:01:57 +03:00
parent 95a57795dc
commit 127cbac5bc
971 changed files with 30678 additions and 5669 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

364
.gitignore vendored
View File

@ -1,342 +1,24 @@
## Ignore Visual Studio temporary files, build results, and
## files generated by popular Visual Studio add-ons.
# User-specific files
*.suo
*.user
*.userosscache
*.sln.docstates
# User-specific files (MonoDevelop/Xamarin Studio)
*.userprefs
# Build results
[Dd]ebug/
[Dd]ebugPublic/
[Rr]elease/
[Rr]eleases/
[Xx]64/
[Xx]86/
[Bb]uild/
bld/
[Bb]in/
[Oo]bj/
# PY.TEST
*.pyc
tests/integration/report.html
tests/integration/report.xml
tests/integration/assets/
tests/integration/__pycache__/
# Visual Studio 2015 cache/options directory
.vs/
# Uncomment if you have tasks that create the project's static files in wwwroot
#wwwroot/
# MSTest test Results
[Tt]est[Rr]esult*/
[Bb]uild[Ll]og.*
# NUNIT
*.VisualState.xml
TestResult.xml
# Build Results of an ATL Project
[Dd]ebugPS/
[Rr]eleasePS/
dlldata.c
# DNX
project.lock.json
artifacts/
*_i.c
*_p.c
*_i.h
*.ilk
*.meta
*.obj
*.pch
*.pdb
*.pgc
*.pgd
*.rsp
*.sbr
*.tlb
*.tli
*.tlh
*.tmp
*.tmp_proj
*.log
*.vspscc
*.vssscc
.builds
*.pidb
*.svclog
*.scc
# Chutzpah Test files
_Chutzpah*
# Visual C++ cache files
ipch/
*.aps
*.ncb
*.opendb
*.opensdf
*.sdf
*.cachefile
*.VC.db
# Visual Studio profiler
*.psess
*.vsp
*.vspx
*.sap
# TFS 2012 Local Workspace
$tf/
# Guidance Automation Toolkit
*.gpState
# ReSharper is a .NET coding add-in
_ReSharper*/
*.[Rr]e[Ss]harper
*.DotSettings.user
# JustCode is a .NET coding add-in
.JustCode
# TeamCity is a build add-in
_TeamCity*
# DotCover is a Code Coverage Tool
*.dotCover
# NCrunch
_NCrunch_*
.*crunch*.local.xml
nCrunchTemp_*
# MightyMoose
*.mm.*
AutoTest.Net/
# Web workbench (sass)
.sass-cache/
# Installshield output folder
[Ee]xpress/
# DocProject is a documentation generator add-in
DocProject/buildhelp/
DocProject/Help/*.HxT
DocProject/Help/*.HxC
DocProject/Help/*.hhc
DocProject/Help/*.hhk
DocProject/Help/*.hhp
DocProject/Help/Html2
DocProject/Help/html
# Click-Once directory
publish/
# Publish Web Output
*.[Pp]ublish.xml
*.azurePubxml
# TODO: Un-comment the next line if you do not want to checkin
# your web deploy settings because they may include unencrypted
# passwords
#*.pubxml
*.publishproj
# NuGet Packages
*.nupkg
# The packages folder can be ignored because of Package Restore
**/packages/*
# except build/, which is used as an MSBuild target.
!**/packages/build/
# Uncomment if necessary however generally it will be regenerated when needed
#!**/packages/repositories.config
# NuGet v3's project.json files produces more ignoreable files
*.nuget.props
*.nuget.targets
# Microsoft Azure Build Output
csx/
*.build.csdef
# Microsoft Azure Emulator
ecf/
rcf/
# Microsoft Azure ApplicationInsights config file
ApplicationInsights.config
# Windows Store app package directory
AppPackages/
BundleArtifacts/
# Visual Studio cache files
# files ending in .cache can be ignored
*.[Cc]ache
# but keep track of directories ending in .cache
!*.[Cc]ache/
# Others
ClientBin/
[Ss]tyle[Cc]op.*
~$*
*~
*.dbmdl
*.dbproj.schemaview
*.pfx
*.publishsettings
node_modules/
orleans.codegen.cs
# RIA/Silverlight projects
Generated_Code/
# Backup & report files from converting an old project file
# to a newer Visual Studio version. Backup files are not needed,
# because we have git ;-)
_UpgradeReport_Files/
Backup*/
UpgradeLog*.XML
UpgradeLog*.htm
# SQL Server files
*.mdf
*.ldf
# Business Intelligence projects
*.rdl.data
*.bim.layout
*.bim_*.settings
# Microsoft Fakes
FakesAssemblies/
# GhostDoc plugin setting file
*.GhostDoc.xml
# Target VS files:
vsx64
# Node.js Tools for Visual Studio
.ntvs_analysis.dat
# Visual Studio 6 build log
*.plg
# Visual Studio 6 workspace options file
*.opt
# Visual Studio LightSwitch build output
**/*.HTMLClient/GeneratedArtifacts
**/*.DesktopClient/GeneratedArtifacts
**/*.DesktopClient/ModelManifest.xml
**/*.Server/GeneratedArtifacts
**/*.Server/ModelManifest.xml
_Pvt_Extensions
# LightSwitch generated files
GeneratedArtifacts/
ModelManifest.xml
# Paket dependency manager
.paket/paket.exe
# FAKE - F# Make
.fake/
*.filters
/External
/Output
/InferenceEngineMain/models
/Test
/HTTPClient/*.a
/InferenceEngineMain/newModels
.DS_Store
# For IDEA
.idea/
VS/
Xcode/
temp/
report/
.kdev4/
*.kdev4
*.kate-swp
/lin-build
/win-build
/CMakeFiles
*.stamp
*.depend
*.vcxproj
*.sln
/CMakeCache.txt
.vimprj/
build_IA32/
.dir-locals.el
GTAGS
GPATH
GRTAGS
GSYMS
compile_commands.json
service/dot-net-service/Output
**/sublime_build
/.project
.vscode/
/vsx32
/service/dot-net-service/.klocwork/DotNetService
cmake-build-*/
/lin64
.gdb_history
.local_vimrc
.ycm_extra_conf.py
tags
# from Model Optimizer repo
# build/artifact dirs
_*
# but ensure we don't skip __init__.py
!__init__.py
# developer tools
.idea
.project
.cproject
.pydevproject
.settings
/bin/
/gen/
__pycache__
*.swp
/config.xml
# Python-specific
.env3
*.pyc
# Tests-specific
.coverage
htmlcov
pylint_report.txt
pylint_report_comments.txt
# Documentation-generated
docs/build
docs/source/_static
docs/source/_templates
docs/source/generated/
# Artifacts
/*.bin
/*.xml
/*.json
/*.so
/*.txt
/*.mapping
/*.dat
/*.svg
.vscode
cmake-build-debug
cmake-build-release
.DS_Store
**/tags
compile_commands.json
bin/
build/
.local_vimrc
.gdb_history
.vimspector.json
doc/
docs/build_documentation/work_dir/
inference-engine/plugins/
.repo/
docs/template_plugin/html/
CMakeLists.txt.user
docs/IE_PLUGIN_DG/html/

2
.gitmodules vendored
View File

@ -5,4 +5,4 @@
[submodule "ngraph"]
path = ngraph
url = https://github.com/NervanaSystems/ngraph.git
ignore = dirty
ignore = dirty

4
CMakeLists.txt
View File

@ -77,13 +77,13 @@ function(build_ngraph)
if (NOT ANDROID)
ngraph_set(NGRAPH_UNIT_TEST_ENABLE TRUE)
ngraph_set(NGRAPH_UNIT_TEST_OPENVINO_ENABLE TRUE)
ngraph_set(NGRAPH_IE_ENABLE TRUE)
# ngraph_set(NGRAPH_ONNX_IMPORT_ENABLE TRUE)
set(NGRAPH_ONNX_IMPORT_ENABLE TRUE CACHE BOOL "" FORCE)
else()
ngraph_set(NGRAPH_UNIT_TEST_ENABLE FALSE)
ngraph_set(NGRAPH_TEST_UTIL_ENABLE FALSE)
ngraph_set(NGRAPH_UNIT_TEST_OPENVINO_ENABLE FALSE)
ngraph_set(NGRAPH_IE_ENABLE FALSE)
ngraph_set(NGRAPH_ONNX_IMPORT_ENABLE FALSE)
endif()

3
Jenkinsfile vendored Normal file
View File

@ -0,0 +1,3 @@
#!groovy
dldtPipelineEntrypoint(this)

6
cmake/developer_package.cmake
View File

@ -37,8 +37,12 @@ function(ie_cpack_set_library_dir)
if(WIN32)
set(IE_CPACK_LIBRARY_PATH ${IE_CPACK_IE_DIR}/lib/${CMAKE_BUILD_TYPE}/${ARCH} PARENT_SCOPE)
set(IE_CPACK_RUNTIME_PATH ${IE_CPACK_IE_DIR}/bin/${CMAKE_BUILD_TYPE}/${ARCH} PARENT_SCOPE)
set(IE_CPACK_ARCHIVE_PATH ${IE_CPACK_IE_DIR}/lib/${CMAKE_BUILD_TYPE}/${ARCH} PARENT_SCOPE)
else()
set(IE_CPACK_LIBRARY_PATH ${IE_CPACK_IE_DIR}/lib/${ARCH} PARENT_SCOPE)
set(IE_CPACK_RUNTIME_PATH ${IE_CPACK_IE_DIR}/lib/${ARCH} PARENT_SCOPE)
set(IE_CPACK_ARCHIVE_PATH ${IE_CPACK_IE_DIR}/lib/${ARCH} PARENT_SCOPE)
endif()
endfunction()
@ -59,8 +63,10 @@ macro(ie_cpack)
set(CPACK_GENERATOR "TGZ")
if(WIN32)
set(CPACK_PACKAGE_NAME inference-engine_${CMAKE_BUILD_TYPE})
string(REPLACE "\\" "_" CPACK_PACKAGE_VERSION "${CI_BUILD_NUMBER}")
else()
set(CPACK_PACKAGE_NAME inference-engine)
string(REPLACE "/" "_" CPACK_PACKAGE_VERSION "${CI_BUILD_NUMBER}")
endif()
set(CPACK_INCLUDE_TOPLEVEL_DIRECTORY OFF)
set(CPACK_ARCHIVE_COMPONENT_INSTALL ON)

1
inference-engine/cmake/add_ie_target.cmake
View File

@ -118,7 +118,6 @@ function(addIeTarget)
if (ARG_ADD_CPPLINT)
# code style
add_cpplint_target(${ARG_NAME}_cpplint FOR_TARGETS ${ARG_NAME})
add_clang_format_target(${ARG_NAME}_clang_format FOR_TARGETS ${ARG_NAME})
endif()
if (ARG_DEVELOPER_PACKAGE)
# developer package

5
inference-engine/cmake/clang_format.cmake
View File

@ -35,10 +35,6 @@ function(add_clang_format_target TARGET_NAME)
set(multiValueArgs "FOR_TARGETS" "FOR_SOURCES" "EXCLUDE_PATTERNS")
cmake_parse_arguments(CLANG_FORMAT "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
if(CLANG_FORMAT_ALL)
set(all ALL)
endif()
foreach(target IN LISTS CLANG_FORMAT_FOR_TARGETS)
get_target_property(target_sources "${target}" SOURCES)
list(APPEND CLANG_FORMAT_FOR_SOURCES ${target_sources})
@ -95,7 +91,6 @@ function(add_clang_format_target TARGET_NAME)
"All clang-format output files")
add_custom_target(${TARGET_NAME}
${all}
DEPENDS ${all_output_files}
COMMENT "[clang-format] ${TARGET_NAME}")

19
inference-engine/cmake/dependencies.cmake
View File

@ -4,6 +4,8 @@
cmake_policy(SET CMP0054 NEW)
include(models)
#we have number of dependencies stored on ftp
include(dependency_solver)
@ -13,6 +15,23 @@ endif()
include(ExternalProject)
if (ENABLE_SAME_BRANCH_FOR_MODELS)
branchName(MODELS_BRANCH)
else()
set(MODELS_BRANCH "master")
endif()
if (ENABLE_DATA)
add_models_repo(${ENABLE_DATA} "data:inference-engine/open-source-data.git")
set(MODELS_PATH "${TEMP}/data/src/data")
set(DATA_PATH "${MODELS_PATH}")
endif()
message(STATUS "MODELS_PATH=" ${MODELS_PATH})
fetch_models_and_validation_set()
include(linux_name)
if(COMMAND get_linux_name)
get_linux_name(LINUX_OS_NAME)

6
inference-engine/cmake/developer_package_config.cmake.in
View File

@ -11,7 +11,11 @@ file(TO_CMAKE_PATH "${CMAKE_CURRENT_LIST_DIR}" cache_path)
set(ie_options "@IE_OPTIONS@;CMAKE_BUILD_TYPE;CMAKE_SKIP_RPATH")
load_cache("${cache_path}" READ_WITH_PREFIX "" ${ie_options})
foreach(option IN LISTS ie_options)
if(NOT DEFINED "${option}")
load_cache("${cache_path}" READ_WITH_PREFIX "" ${option})
endif()
endforeach()
message(STATUS "The following CMake options are exported from Inference Engine Developer package")
message("")

4
inference-engine/cmake/features_ie.cmake
View File

@ -78,7 +78,9 @@ ie_dependent_option (GAPI_TEST_PERF "if GAPI unit tests should examine performan
ie_dependent_option (ENABLE_MYRIAD_MVNC_TESTS "functional and behavior tests for mvnc api" OFF "ENABLE_TESTS;ENABLE_MYRIAD" OFF)
ie_dependent_option (ENABLE_SAMPLES "console samples are part of inference engine package" ON "NOT MINGW" OFF)
ie_dependent_option (ENABLE_DATA "fetch models from open-source-data repo" ON "ENABLE_FUNCTIONAL_TESTS;NOT ANDROID" OFF)
ie_dependent_option (ENABLE_SAME_BRANCH_FOR_MODELS "uses same branch for models and for inference engine, if not enabled models are taken from master" OFF "ENABLE_TESTS" OFF)
ie_dependent_option (ENABLE_BEH_TESTS "tests oriented to check inference engine API corecteness" ON "ENABLE_TESTS" OFF)

80
inference-engine/cmake/models.cmake Normal file
View File

@ -0,0 +1,80 @@
# Copyright (C) 2018-2020 Intel Corporation
# SPDX-License-Identifier: Apache-2.0
#
if(ENABLE_DOCKER)
cmake_minimum_required(VERSION 3.3 FATAL_ERROR)
else()
cmake_minimum_required(VERSION 3.8 FATAL_ERROR)
endif()
cmake_policy(SET CMP0054 NEW)
find_package(Git REQUIRED)
set(MODELS_LST "")
set(MODELS_LST_TO_FETCH "")
function (add_models_repo add_to_fetcher model_name)
list(LENGTH ARGV add_models_args)
if (add_models_args EQUAL 3)
list(GET ARGV 2 branch_name)
else()
set(branch_name ${MODELS_BRANCH})
endif()
if (add_to_fetcher)
set(model_name "${model_name}:${branch_name}")
list(APPEND MODELS_LST_TO_FETCH ${model_name})
endif()
list(APPEND MODELS_LST ${model_name})
set(MODELS_LST_TO_FETCH ${MODELS_LST_TO_FETCH} PARENT_SCOPE)
set(MODELS_LST ${MODELS_LST} PARENT_SCOPE)
endfunction()
function(add_lfs_repo name prefix url tag)
ExternalProject_Add(${name}
PREFIX ${prefix}
GIT_REPOSITORY ${url}
GIT_TAG ${tag}
GIT_CONFIG "http.sslverify=false"
GIT_PROGRESS 1
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
INSTALL_COMMAND ""
LOG_DOWNLOAD ON)
execute_process(
COMMAND ${GIT_EXECUTABLE} lfs install --local --force
WORKING_DIRECTORY ${prefix}/src/${name}
OUTPUT_VARIABLE lfs_output
RESULT_VARIABLE lfs_var)
if(lfs_var)
message(FATAL_ERROR [=[
Failed to setup Git LFS: ${lfs_output}
Git lfs must be installed in order to fetch models
Please install it from https://git-lfs.github.com/
]=])
endif()
endfunction()
function (fetch_models_and_validation_set)
foreach(loop_var ${MODELS_LST_TO_FETCH})
string(REPLACE ":" ";" MODEL_CONFIG_LST ${loop_var})
list(GET MODEL_CONFIG_LST 0 folder_name)
list(GET MODEL_CONFIG_LST 1 repo_name)
list(GET MODEL_CONFIG_LST 2 branch_name)
string(FIND ${folder_name} "model" IS_MODEL)
if(${folder_name} MATCHES "model*")
set(FOLDER_NAME "/models/src")
endif()
add_lfs_repo(
"${folder_name}"
${TEMP}${FOLDER_NAME}/${folder_name}
"git@gitlab-icv.inn.intel.com:${repo_name}"
"${branch_name}")
endforeach(loop_var)
endfunction()

4
inference-engine/cmake/plugins/plugins.cmake
View File

@ -90,8 +90,8 @@ function(ie_add_plugin)
ie_cpack_add_component(${install_component} REQUIRED DEPENDS core)
install(TARGETS ${IE_PLUGIN_NAME}
RUNTIME DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT ${install_component}
ARCHIVE DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT ${install_component}
RUNTIME DESTINATION ${IE_CPACK_RUNTIME_PATH} COMPONENT ${install_component}
ARCHIVE DESTINATION ${IE_CPACK_ARCHIVE_PATH} COMPONENT ${install_component}
LIBRARY DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT ${install_component})
endif()
endfunction()

2
inference-engine/cmake/vpu_dependencies.cmake
View File

@ -104,4 +104,4 @@ if(ANDROID)
set(LIBUSB_LIBRARY "${LIBUSB}/libs/${ANDROID_ABI}/libusb1.0.so")
log_rpath_from_dir(LIBUSB "${LIBUSB}/libs/${ANDROID_ABI}")
endif()
endif()

2
inference-engine/ie_bridges/c/samples/hello_nv12_input_classification/main.c
View File

@ -1,5 +1,5 @@
// Copyright (C) 2018-2020 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
// SPDX-License-Identifier : Apache-2.0
//
#include <stdlib.h>

2
inference-engine/ie_bridges/c/samples/object_detection_sample_ssd/main.c
View File

@ -1,5 +1,5 @@
// Copyright (C) 2018-2020 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
// SPDX-License-Identifier : Apache-2.0
//
#include <stdlib.h>

4
inference-engine/ie_bridges/c/src/CMakeLists.txt
View File

@ -28,8 +28,8 @@ export(TARGETS ${TARGET_NAME} NAMESPACE IE:: APPEND FILE "${CMAKE_BINARY_DIR}/ta
# install
install(TARGETS ${TARGET_NAME}
RUNTIME DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT core
ARCHIVE DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT core
RUNTIME DESTINATION ${IE_CPACK_RUNTIME_PATH} COMPONENT core
ARCHIVE DESTINATION ${IE_CPACK_ARCHIVE_PATH} COMPONENT core
LIBRARY DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT core)
install(DIRECTORY ${InferenceEngine_C_API_SOURCE_DIR}/include/

61
inference-engine/ie_bridges/python/sample/object_detection_sample_ssd/object_detection_sample_ssd.py
View File

@ -29,15 +29,18 @@ def build_argparser():
args = parser.add_argument_group("Options")
args.add_argument('-h', '--help', action='help', default=SUPPRESS, help='Show this help message and exit.')
args.add_argument("-m", "--model", help="Required. Path to an .xml file with a trained model.",
required=True, type=str)
required=True, type=str)
args.add_argument("-i", "--input", help="Required. Path to image file.",
required=True, type=str, nargs="+")
required=True, type=str, nargs="+")
args.add_argument("-l", "--cpu_extension",
help="Optional. Required for CPU custom layers. Absolute path to a shared library with the kernels implementations.",
type=str, default=None)
help="Optional. Required for CPU custom layers. "
"Absolute path to a shared library with the kernels implementations.",
type=str, default=None)
args.add_argument("-d", "--device",
help="Optional. Specify the target device to infer on; CPU, GPU, FPGA or MYRIAD is acceptable. Sample will look for a suitable plugin for device specified (CPU by default)",
default="CPU", type=str)
help="Optional. Specify the target device to infer on; "
"CPU, GPU, FPGA or MYRIAD is acceptable. "
"Sample will look for a suitable plugin for device specified (CPU by default)",
default="CPU", type=str)
args.add_argument("--labels", help="Optional. Labels mapping file", default=None, type=str)
args.add_argument("-nt", "--number_top", help="Optional. Number of top results", default=10, type=int)
@ -59,9 +62,10 @@ def main():
# ------------- 2. Load Plugin for inference engine and extensions library if specified --------------
log.info("Device info:")
versions = ie.get_versions(args.device)
print("{}{}".format(" "*8, args.device))
print("{}MKLDNNPlugin version ......... {}.{}".format(" "*8, versions[args.device].major, versions[args.device].minor))
print("{}Build ........... {}".format(" "*8, versions[args.device].build_number))
print("{}{}".format(" " * 8, args.device))
print("{}MKLDNNPlugin version ......... {}.{}".format(" " * 8, versions[args.device].major,
versions[args.device].minor))
print("{}Build ........... {}".format(" " * 8, versions[args.device].build_number))
if args.cpu_extension and "CPU" in args.device:
ie.add_extension(args.cpu_extension, "CPU")
@ -79,8 +83,15 @@ def main():
# -----------------------------------------------------------------------------------------------------
# --------------------------- 3. Read and preprocess input --------------------------------------------
input_blob = next(iter(net.inputs))
n, c, h, w = net.inputs[input_blob].shape
print("inputs number: " + str(len(net.inputs.keys())))
for input_key in net.inputs:
print("input shape: " + str(net.inputs[input_key].shape))
print("input key: " + input_key)
if len(net.inputs[input_key].layout) == 4:
n, c, h, w = net.inputs[input_key].shape
images = np.ndarray(shape=(n, c, h, w))
images_hw = []
for i in range(n):
@ -94,13 +105,14 @@ def main():
log.warning("Image {} is resized from {} to {}".format(args.input[i], image.shape[:-1], (h, w)))
image = image.transpose((2, 0, 1)) # Change data layout from HWC to CHW
images[i] = image
# -----------------------------------------------------------------------------------------------------
# --------------------------- 4. Configure input & output ---------------------------------------------
# --------------------------- Prepare input blobs -----------------------------------------------------
log.info("Preparing input blobs")
assert (len(net.inputs.keys()) == 1 or len(net.inputs.keys()) == 2), "Sample supports topologies only with 1 or 2 inputs"
input_blob = next(iter(net.inputs))
assert (len(net.inputs.keys()) == 1 or len(
net.inputs.keys()) == 2), "Sample supports topologies only with 1 or 2 inputs"
out_blob = next(iter(net.outputs))
input_name, input_info_name = "", ""
@ -112,9 +124,21 @@ def main():
elif len(net.inputs[input_key].layout) == 2:
input_info_name = input_key
net.inputs[input_key].precision = 'FP32'
if net.inputs[input_key].shape[1] != 3 and net.inputs[input_key].shape[1] != 6 or net.inputs[input_key].shape[0] != 1:
if net.inputs[input_key].shape[1] != 3 and net.inputs[input_key].shape[1] != 6 or \
net.inputs[input_key].shape[0] != 1:
log.error('Invalid input info. Should be 3 or 6 values length.')
data = {}
data[input_name] = images
if input_info_name != "":
infos = np.ndarray(shape=(n, c), dtype=float)
for i in range(n):
infos[i, 0] = h
infos[i, 1] = w
infos[i, 2] = 1.0
data[input_info_name] = infos
# --------------------------- Prepare output blobs ----------------------------------------------------
log.info('Preparing output blobs')
@ -141,7 +165,7 @@ def main():
log.info("Loading model to the device")
exec_net = ie.load_network(network=net, device_name=args.device)
log.info("Creating infer request and starting inference")
res = exec_net.infer(inputs={input_blob: images})
res = exec_net.infer(inputs=data)
# -----------------------------------------------------------------------------------------------------
# --------------------------- Read and postprocess output ---------------------------------------------
@ -159,8 +183,8 @@ def main():
ymin = np.int(ih * proposal[4])
xmax = np.int(iw * proposal[5])
ymax = np.int(ih * proposal[6])
print("[{},{}] element, prob = {:.6} ({},{})-({},{}) batch id : {}"\
.format(number, label, confidence, xmin, ymin, xmax, ymax, imid), end="")
print("[{},{}] element, prob = {:.6} ({},{})-({},{}) batch id : {}" \
.format(number, label, confidence, xmin, ymin, xmax, ymax, imid), end="")
if proposal[2] > 0.5:
print(" WILL BE PRINTED!")
if not imid in boxes.keys():
@ -181,7 +205,8 @@ def main():
# -----------------------------------------------------------------------------------------------------
log.info("Execution successful\n")
log.info("This sample is an API example, for any performance measurements please use the dedicated benchmark_app tool")
log.info(
"This sample is an API example, for any performance measurements please use the dedicated benchmark_app tool")
if __name__ == '__main__':

10
inference-engine/ie_bridges/python/src/openvino/inference_engine/CMakeLists.txt
View File

@ -39,6 +39,16 @@ add_custom_command(TARGET ${TARGET_NAME}
COMMAND ${CMAKE_COMMAND} -E copy ${PYTHON_BRIDGE_SRC_ROOT}/src/openvino/__init__.py ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/../__init__.py
)
# creates a folder in openvino directory and a symlink to benchmark
# inside bin directory for developers for running python benchmark_app
if(UNIX)
add_custom_command(TARGET ${TARGET_NAME}
POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/../tools
)
file(COPY ${OpenVINO_MAIN_SOURCE_DIR}/tools/benchmark DESTINATION ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/../tools/)
endif()
# install
install(TARGETS ${TARGET_NAME}

136
inference-engine/ie_bridges/python/src/openvino/inference_engine/ie_api.pyx
View File

@ -171,9 +171,9 @@ cdef class IECore:
#
# Usage example:\n
# ```python
# net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
# ie = IECore()
# exec_net = ie.load_network(network=net, device_name="CPU", num_requsts=2)
# net = ie.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# exec_net = ie.load_network(network=net, device_name="CPU", num_requests=2)
# ```
cpdef ExecutableNetwork load_network(self, IENetwork network, str device_name, config=None, int num_requests=1):
cdef ExecutableNetwork exec_net = ExecutableNetwork()
@ -197,8 +197,8 @@ cdef class IECore:
# @return An `ExecutableNetwork` object
# Usage example:\n
# ```python
# net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
# ie = IECore()
# net = ie.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# exec_net = ie.load_network(network=net, device_name="MYRIAD", num_requsts=2)
# # export executable network
# exec_net.export(path_to_file_to_save)
@ -226,8 +226,8 @@ cdef class IECore:
#
# Usage example:\n
# ```python
# net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
# ie = IECore()
# net = ie.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# layers_map = ie.query_network(network=net, device_name="HETERO:GPU,CPU")
# ```
def query_network(self, IENetwork network, str device_name, config=None):
@ -238,12 +238,19 @@ cdef class IECore:
return c_map_to_dict(res)
## Sets a configuration for a plugin
# NOTE: When specifying a key value of a config, the "KEY_" prefix is omitted.
#
# \note When specifying a key value of a config, the "KEY_" prefix is omitted.
#
# @param config: a dictionary of configuration parameters as keys and their values
# @param device_name: a device name of a target plugin
# @return None
#
# Usage examples: See the `set_affinity` method of the `IENetwork` class
# Usage examples:\n
# ```python
# ie = IECore()
# net = ie.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# ie.set_config({"DYN_BATCH_ENABLED": "YES"})
# ```
def set_config(self, config: dict, device_name: str):
cdef map[string, string] c_config = dict_to_c_map(config)
self.impl.setConfig(c_config, device_name.encode())
@ -316,7 +323,9 @@ cdef class IECore:
## Gets a configuration dedicated to device behavior. The method targets to extract information
# which can be set via set_config method.
# NOTE: When specifying a key value of a config, the "KEY_" prefix is omitted.
#
# \note When specifying a key value of a config, the "KEY_" prefix is omitted.
#
# @param device_name: A name of a device to get a config value.
# @param config_name: A config name to request.
# @return A config value corresponding to a config key.
@ -452,8 +461,8 @@ cdef class ExecutableNetwork:
#
# Usage example:\n
# ```python
# net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
# ie_core = IECore()
# net = ie_core.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# exec_net = ie_core.load_network(net, device, num_requests=2)
# res = exec_net.infer({'data': img})
# res
@ -531,8 +540,8 @@ cdef class ExecutableNetwork:
#
# Usage example:\n
# ```python
# net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
# ie_core = IECore()
# net = ie_core.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# exec_net = ie_core.load_network(net, device, num_requsts=2)
# exec_graph = exec_net.get_exec_graph_info()
# ```
@ -549,7 +558,7 @@ cdef class ExecutableNetwork:
# Usage example:\n
# ```python
# ie = IECore()
# net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
# net = ie.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# exec_net = ie.load_network(net, "CPU")
# exec_net.get_metric("NETWORK_NAME")
# ```
@ -564,7 +573,7 @@ cdef class ExecutableNetwork:
# Usage example:\n
# ```python
# ie = IECore()
# net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
# net = ie.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# exec_net = ie.load_network(net, "CPU")
# exec_net.get_metric("DEVICE_ID")
# ```
@ -576,8 +585,8 @@ cdef class ExecutableNetwork:
# @return None
#
# ```python
# net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
# ie = IECore()
# net = ie.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# exec_net = ie.load_network(network=net, device_name="MYRIAD", num_requsts=2)
# exec_net.export(path_to_file_to_save)
# ```
@ -632,8 +641,8 @@ cdef class InferRequest:
# Usage example:\n
# ```python
# callback = lambda status, py_data: print("Request with id {} finished with status {}".format(py_data, status))
# net = IENetwork("./model.xml", "./model.bin")
# ie = IECore()
# net = ie.read_network(model="./model.xml", weights="./model.bin")
# exec_net = ie.load_network(net, "CPU", num_requests=4)
# for id, req in enumerate(exec_net.requests):
# req.set_completion_callback(py_callback=callback, py_data=id)
@ -662,7 +671,7 @@ cdef class InferRequest:
#
# Usage example:\n
# ```python
# exec_net = plugin.load(network=net, num_requests=2)
# exec_net = ie_core.load_network(network=net, num_requests=2)
# exec_net.requests[0].infer({input_blob: image})
# res = exec_net.requests[0].outputs['prob']
# np.flip(np.sort(np.squeeze(res)),0)
@ -683,7 +692,7 @@ cdef class InferRequest:
#
# Usage example:\n
# ```python
# exec_net = plugin.load(network=net, num_requests=2)
# exec_net = ie_core.load_network(network=net, num_requests=2)
# exec_net.requests[0].async_infer({input_blob: image})
# request_status = exec_net.requests[0].wait()
# res = exec_net.requests[0].outputs['prob']
@ -697,7 +706,8 @@ cdef class InferRequest:
## Waits for the result to become available. Blocks until specified timeout elapses or the result
# becomes available, whichever comes first.
# NOTE: There are special values of the timeout parameter:
#
# \note There are special values of the timeout parameter:
# * 0 - Immediately returns the inference status. It does not block or interrupt execution.
# To find statuses meaning, please refer to InferenceEngine::StatusCode in Inference Engine C++ documentation
# * -1 - Waits until inference result becomes available (default value)
@ -724,12 +734,14 @@ cdef class InferRequest:
return deref(self.impl).wait(<int64_t> timeout)
## Queries performance measures per layer to get feedback of what is the most time consuming layer.
# NOTE: Performance counters data and format depends on the plugin
#
# \note Performance counters data and format depends on the plugin
#
# @return Dictionary containing per-layer execution information.
#
# Usage example:
# ```python
# exec_net = plugin.load(network=net, num_requests=2)
# exec_net = ie_core.load_network(network=net, num_requests=2)
# exec_net.requests[0].infer({input_blob: image})
# exec_net.requests[0].get_perf_counts()
# {'Conv2D': {'exec_type': 'jit_avx2_1x1',
@ -780,18 +792,20 @@ cdef class InferRequest:
## Sets new batch size for certain infer request when dynamic batching is enabled in executable network
# that created this request.
# NOTE: Support of dynamic batch size depends on the target plugin.
#
# \note Support of dynamic batch size depends on the target plugin.
#
# @param size: New batch size to be used by all the following inference calls for this request
# @return None
#
# Usage example:\n
# ```python
# net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
# ie = IECore()
# net = ie.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# # Set max batch size
# net.batch = 10
# plugin.set_config({"DYN_BATCH_ENABLED": "YES"})
# exec_net = plugin.load(network=net)
# ie.set_config({"DYN_BATCH_ENABLED": "YES"})
# exec_net = ie.load_network(network=net)
# # Set batch size for certain network.
# # NOTE: Input data shape will not be changed, but will be used partially in inference which increases performance
# exec_net.requests[0].set_batch(2)
@ -855,7 +869,11 @@ cdef class IENetLayer:
def type(self):
return deref(self._ptr).type.decode()
## Layer base operating precision. Provides getter and setter interfaces.
## \note This property is deprecated.
# Please, use out_data property to access DataPtr objects for all output ports, which contains full
# information about layer's output data including precision.
#
# Layer base operating precision. Provides getter and setter interfaces.
@property
def precision(self):
warnings.filterwarnings("always", category=DeprecationWarning)
@ -874,8 +892,8 @@ cdef class IENetLayer:
# The affinity attribute provides getter and setter interfaces, so the layer affinity can be modified directly.
# For example:\n
# ```python
# net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
# ie = IECore()
# net = ie.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# layers_map = ie.query_network(network=net, device_name="HETERO:GPU,CPU")
# layers = net.layers
# for layer, device in layers_map.items():
@ -922,8 +940,10 @@ cdef class IENetLayer:
input_to_list.append(deref(layer.second).name.decode())
return input_to_list
## Deprecated: use out_data property to access DataPtr objects for all output ports, which contains full
## \note This property is deprecated.
# Please, use out_data property to access DataPtr objects for all output ports, which contains full
# information about layer's output data including layout
#
# Returns the layout of the layer output data on 1st port
@property
def layout(self):
@ -936,8 +956,10 @@ cdef class IENetLayer:
cdef C.DataPtr c_input = deref(self._ptr).outData[0]
return layout_int_to_str_map[deref(c_input).getLayout()]
## Deprecated: use out_data property to access DataPtr objects for all output ports, which contains full
## \note This property is deprecated.
# Please, use out_data property to access DataPtr objects for all output ports, which contains full
# information about layer's output data including shape
#
# Return the list of dimension of the layer output data on 1st port
@property
def shape(self):
@ -988,7 +1010,10 @@ cdef class IENetLayer:
weights_buffer.reset(blob.second)
blobs_map[blob.first.decode()] = weights_buffer.to_numpy()
return blobs_map
## Dictionary with layer weights, biases or custom blobs if any
## \note This property is deprecated.
# Please use blobs property instead.
#
# Dictionary with layer weights, biases or custom blobs if any
@property
def weights(self):
warnings.filterwarnings("always", category=DeprecationWarning)
@ -1003,6 +1028,9 @@ cdef class IENetLayer:
cdef class IENetwork:
## Class constructor
#
# \note Reading networks using IENetwork constructor is deprecated.
# Please, use IECore.read_network() method instead.
#
# @param model: A `.xml` file of the IR or PyCapsule containing smart pointer to nGraph function.
# In case of passing a `.xml` file attribute value can be a string path or bytes with file content
# depending on `init_from_buffer` attribute value
@ -1100,8 +1128,9 @@ cdef class IENetwork:
## Batch size of the network. Provides getter and setter interfaces to get and modify the
# network batch size. For example:\n
# ```python
# net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
# print(et.batch_size)
# ie = IECore()
# net = ie.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# print(net.batch_size)
# net.batch_size = 4
# print(net.batch_size)
# print(net.inputs['data'].shape)
@ -1109,7 +1138,9 @@ cdef class IENetwork:
@property
def batch_size(self):
return self.impl.getBatch()
## Deprecated: network precision does not make sence, use precision on egdes.
## \note This property is deprecated:
# network precision does not make sense, use precision on edges.
#
# Precision of the network
@property
def precision(self):
@ -1139,13 +1170,16 @@ cdef class IENetwork:
layers[deref(l).name.decode()] = net_l
return layers
## Deprecated: new Calibration Tool doesn't generate statistics
## \note This property is deprecated.
# New Calibration Tool doesn't generate statistics
#
# Returns `LayersStatsMap` object containing dictionary that maps network layer names to calibration statistics
# represented by `LayerStats` objects.
#
# Usage example:\n
# ```python
# net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
# ie = IECore()
# net = ie.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# net.stats.update({"conv1_2d" : LayserStats(min=(-25, -1, 0), max=(63, 124, 70)),
# "conv2_2d" : LayserStats(min=(-5, -1, 0, 1, -7, 2), max=(63, 124, 70, 174, 99, 106))
# })
@ -1163,26 +1197,6 @@ cdef class IENetwork:
max=tuple(it.second["max".encode()]))
return py_stats_map
## NOTE: The function is deprecated. Please use the `IENetwork()` class constructor
# to create valid instance of `IENetwork`.
#
# Reads the model from the `.xml` and `.bin` files of the IR.
#
# @param model: Path to `.xml` file of the IR
# @param weights: Path to `.bin` file of the IR
# @return An instance of the `IENetwork` class
@classmethod
def from_ir(cls, model: str, weights: str):
warnings.filterwarnings("always", category=DeprecationWarning)
warnings.warn("from_ir() method of IENetwork is deprecated. "
"Please use IENetwork class constructor to create valid IENetwork instance",
DeprecationWarning)
if not os.path.isfile(model):
raise Exception("Path to the model {} doesn't exists or it's a directory".format(model))
if not os.path.isfile(weights):
raise Exception("Path to the weights {} doesn't exists or it's a directory".format(weights))
cdef IENetwork net = IENetwork(model, weights)
return net
## Marks any intermediate layer as output layer to retrieve the inference results from the specified layers.
# @param outputs: List of layers to be set as model outputs. The list can contain strings with layer names to be set
@ -1192,7 +1206,8 @@ cdef class IENetwork:
#
# Usage example:\n
# ```python
# net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
# ie = IECore()
# net = ie.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# net.add_outputs(["conv5_1', conv2_1', (split_2, 1)])]
# ```
def add_outputs(self, outputs):
@ -1216,14 +1231,16 @@ cdef class IENetwork:
#
# Usage example:
# ```python
# net = IENetwork(model=path_to_model, weights=path_to_weights)
# ie = IECore()
# net = ie.read_network(model=path_to_xml, weights=path_to_bin)
# net.serialize(path_to_xml, path_to_bin)
# ```
def serialize(self, path_to_xml, path_to_bin: str = ""):
self.impl.serialize(path_to_xml.encode(), path_to_bin.encode())
## Reshapes the network to change spatial dimensions, batch size, or any dimension.
# NOTE: Before using this method, make sure that the target shape is applicable for the network.
#
# \note Before using this method, make sure that the target shape is applicable for the network.
# Changing the network shape to an arbitrary value may lead to unpredictable behaviour.
#
# @param input_shapes: A dictionary that maps input layer names to tuples with the target shape
@ -1231,7 +1248,8 @@ cdef class IENetwork:
#
# Usage example:\n
# ```python
# net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
# ie = IECore()
# net = ie.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# input_layer = next(iter(net.inputs))
# n, c, h, w = net.inputs[input_layer]
# net.reshape({input_layer: (n, c, h*2, w*2)}]
@ -1255,9 +1273,11 @@ cdef class IENetwork:
# return self.impl.getFunction()
## This class is the main plugin interface and serves to initialize and configure the plugin.
#
#\note This class is deprecated: Use IECore instead
#
cdef class IEPlugin:
## Deprecated: Use IECore instead
# Class constructor
## Class constructor
#
# @param device: Target device name. Supported devices: CPU, GPU, FPGA, MYRIAD, HETERO, MULTI
# @param plugin_dirs: List of paths to plugin directories

4
inference-engine/include/cpp/ie_cnn_net_reader.h
View File

@ -40,7 +40,7 @@ public:
/**
* @brief A default constructor
*/
CNNNetReader(): actual(shared_from_irelease(InferenceEngine::CreateCNNNetReader())) {
CNNNetReader(): actual(InferenceEngine::CreateCNNNetReaderPtr()) {
if (actual == nullptr) {
THROW_IE_EXCEPTION << "CNNNetReader was not initialized.";
}
@ -182,7 +182,7 @@ public:
}
private:
std::shared_ptr<ICNNNetReader> actual;
CNNNetReaderPtr actual;
std::shared_ptr<CNNNetwork> network;
};
IE_SUPPRESS_DEPRECATED_END

27
inference-engine/include/cpp/ie_cnn_network.h
View File

@ -66,8 +66,11 @@ public:
* @param reader Pointer to the ICNNNetReader object
*/
IE_SUPPRESS_DEPRECATED_START
explicit CNNNetwork(std::shared_ptr<ICNNNetReader> reader): reader(reader), actual(reader->getNetwork(nullptr)) {
if (actual == nullptr) {
explicit CNNNetwork(CNNNetReaderPtr reader_): reader(reader_) {
if (reader == nullptr) {
THROW_IE_EXCEPTION << "ICNNNetReader was not initialized.";
}
if ((actual = reader->getNetwork(nullptr)) == nullptr) {
THROW_IE_EXCEPTION << "CNNNetwork was not initialized.";
}
}
@ -160,6 +163,15 @@ public:
return actual->getBatchSize();
}
/**
* @brief An overloaded operator cast to get pointer on current network
*
* @return A shared pointer of the current network
*/
operator std::shared_ptr<ICNNNetwork>() {
return network;
}
/**
* @brief An overloaded operator & to get current network
*
@ -178,6 +190,15 @@ public:
return *actual;
}
/**
* @brief Returns constant nGraph function
*
* @return constant nGraph function
*/
std::shared_ptr<ngraph::Function> getFunction() noexcept {
return actual->getFunction();
}
/**
* @brief Returns constant nGraph function
*
@ -297,7 +318,7 @@ protected:
* @brief Reader extra reference, might be nullptr
*/
IE_SUPPRESS_DEPRECATED_START
std::shared_ptr<ICNNNetReader> reader;
CNNNetReaderPtr reader;
IE_SUPPRESS_DEPRECATED_END
/**
* @brief Network extra interface, might be nullptr

11
inference-engine/include/details/ie_so_pointer.hpp
View File

@ -58,6 +58,7 @@ public:
IE_SUPPRESS_DEPRECATED_END
}
private:
/**
* @brief Loads function from the library and returns a pointer to it
* @param functionName Name of function to load
@ -126,6 +127,15 @@ public:
}
}
/**
* @brief Constructs an object with existing loader
* @param so_loader Existing pointer to a library loader
*/
explicit SOPointer(std::shared_ptr<Loader> so_loader)
: _so_loader(so_loader),
_pointedObj(details::shared_from_irelease(
SymbolLoader<Loader>(_so_loader).template instantiateSymbol<T>(SOCreatorTrait<T>::name))) {}
/**
* @brief The copy-like constructor, can create So Pointer that dereferenced into child type if T is derived of U
* @param that copied SOPointer object
@ -183,6 +193,7 @@ protected:
* @brief Gets a smart pointer to the DLL
*/
std::shared_ptr<Loader> _so_loader;
/**
* @brief Gets a smart pointer to the custom object
*/

39
inference-engine/include/ie_icnn_net_reader.h
View File

@ -11,8 +11,10 @@
#include <map>
#include <string>
#include <vector>
#include "details/ie_no_copy.hpp"
#include "details/ie_so_pointer.hpp"
#include "ie_api.h"
#include "ie_blob.h"
#include "ie_common.h"
@ -118,14 +120,45 @@ public:
* @return IR version number: 1 or 2
*/
virtual int getVersion(ResponseDesc* resp) noexcept = 0;
virtual void addExtensions(const std::vector<InferenceEngine::IExtensionPtr>& ext) = 0;
/**
* @brief A virtual destructor.
*/
~ICNNNetReader() override = default;
};
IE_SUPPRESS_DEPRECATED_START
namespace details {
/**
* @brief This class defines the name of the fabric for creating an IHeteroInferencePlugin object in DLL
*/
template<>
class SOCreatorTrait<ICNNNetReader> {
public:
/**
* @brief A name of the fabric for creating IInferencePlugin object in DLL
*/
static constexpr auto name = "CreateICNNNetReader";
};
} // namespace details
/**
* @brief A C++ helper to work with objects created by the IR readers plugin.
* Implements different interfaces.
*/
using CNNNetReaderPtr = InferenceEngine::details::SOPointer<ICNNNetReader, InferenceEngine::details::SharedObjectLoader>;
/**
* @brief Creates a CNNNetReader instance
*
* @return An object that implements the ICNNNetReader interface
*/
IE_SUPPRESS_DEPRECATED_START
INFERENCE_ENGINE_API(ICNNNetReader*) CreateCNNNetReader() noexcept;
INFERENCE_ENGINE_API_CPP(CNNNetReaderPtr) CreateCNNNetReaderPtr() noexcept;
IE_SUPPRESS_DEPRECATED_END
} // namespace InferenceEngine

6
inference-engine/include/ie_icnn_network.hpp
View File

@ -47,6 +47,12 @@ public:
*/
using Ptr = std::shared_ptr<ICNNNetwork>;
/**
* @brief Returns nGraph function
* @return nGraph function
*/
virtual std::shared_ptr<ngraph::Function> getFunction() noexcept = 0;
/**
* @brief Returns constant nGraph function
* @return constant nGraph function

30
inference-engine/include/ie_layers.h
View File

@ -2290,23 +2290,20 @@ public:
/**
* @deprecated Migrate to IR v10 and work with ngraph::Function directly. The method will be removed in 2020.3
* @brief This class represents a standard Scatter layer
* @brief This class represents a standard ScatterUpdate layer
*/
class INFERENCE_ENGINE_INTERNAL_CNNLAYER_CLASS(ScatterLayer): public CNNLayer {
class INFERENCE_ENGINE_INTERNAL_CNNLAYER_CLASS(ScatterUpdateLayer): public CNNLayer {
public:
/**
* @brief The axis in Dictionary to scatter Indexes from
*/
int axis = 0;
/**
* @brief Creates a new ScatterLayer instance.
* @brief Creates a new ScatterUpdateLayer instance.
*/
using CNNLayer::CNNLayer;
~ScatterLayer() override;
~ScatterUpdateLayer() override;
};
/**
* @deprecated Migrate to IR v10 and work with ngraph::Function directly. The method will be removed in 2020.3
* @brief This class represents an onnx ExperimentalDetectronPriorGridGenerator Layer
*/
class INFERENCE_ENGINE_INTERNAL_CNNLAYER_CLASS(ExperimentalDetectronPriorGridGeneratorLayer): public CNNLayer {
@ -2340,6 +2337,23 @@ public:
virtual ~ExperimentalDetectronPriorGridGeneratorLayer();
};
/**
* @brief This class represents a standard ExperimentalDetectronTopKROIs layer
*/
class INFERENCE_ENGINE_INTERNAL_CNNLAYER_CLASS(ExperimentalDetectronTopKROIs): public CNNLayer {
public:
/**
* @brief The maximum number of output rois
*/
int max_rois = 0;
/**
* @brief Creates a new ExperimentalDetectronTopKROIs instance.
*/
using CNNLayer::CNNLayer;
virtual ~ExperimentalDetectronTopKROIs();
};
/**
* @brief This class represents an onnx ExperimentalDetectronGenerateProposalsSingleImage Layer
*/

2
inference-engine/include/ie_parameter.hpp
View File

@ -21,6 +21,7 @@
#include <vector>
#include "ie_api.h"
#include "ie_blob.h"
namespace ngraph {
@ -325,6 +326,7 @@ private:
};
#ifdef __clang__
extern template struct INFERENCE_ENGINE_API_CLASS(InferenceEngine::Parameter::RealData<InferenceEngine::Blob::Ptr>);
extern template struct INFERENCE_ENGINE_API_CLASS(InferenceEngine::Parameter::RealData<int>);
extern template struct INFERENCE_ENGINE_API_CLASS(InferenceEngine::Parameter::RealData<bool>);
extern template struct INFERENCE_ENGINE_API_CLASS(InferenceEngine::Parameter::RealData<float>);

12
inference-engine/include/ie_plugin_config.hpp
View File

@ -348,5 +348,17 @@ DECLARE_CONFIG_KEY(EXCLUSIVE_ASYNC_REQUESTS);
*/
DECLARE_CONFIG_KEY(DUMP_EXEC_GRAPH_AS_DOT);
/**
* @brief The name for setting to execute in bfloat16 precision whenever it is possible
*
* This option let plugin know to downscale the precision where it see performance benefits from
* bfloat16 execution
* Such option do not guarantee accuracy of the network, the accuracy in this mode should be
* verified separately by the user and basing on performance and accuracy results it should be
* user's decision to use this option or not to use
*/
DECLARE_CONFIG_KEY(ENFORCE_BF16);
} // namespace PluginConfigParams
} // namespace InferenceEngine

13
inference-engine/include/ie_precision.hpp
View File

@ -26,7 +26,8 @@ public:
UNSPECIFIED = 255, /**< Unspecified value. Used by default */
MIXED = 0, /**< Mixed value. Can be received from network. No applicable for tensors */
FP32 = 10, /**< 32bit floating point value */
FP16 = 11, /**< 16bit floating point value */
FP16 = 11, /**< 16bit floating point value, 5 bit for exponent, 10 bit for mantisa */
BF16 = 12, /**< 16bit floating point value, 8 bit for exponent, 7 bit for mantisa*/
Q78 = 20, /**< 16bit specific signed fixed point precision */
I16 = 30, /**< 16bit signed integer value */
U8 = 40, /**< 8bit unsigned integer value */
@ -106,6 +107,7 @@ public:
switch (precisionInfo.value) {
CASE(FP32, float);
CASE2(FP16, int16_t, uint16_t);
CASE2(BF16, int16_t, uint16_t);
CASE(I16, int16_t);
CASE(I32, int32_t);
CASE(I64, int64_t);
@ -181,9 +183,9 @@ public:
static std::unordered_map<std::string, ePrecision> names = {
#define PRECISION_NAME(s) {#s, s}
PRECISION_NAME(Q78), PRECISION_NAME(U8), PRECISION_NAME(I8), PRECISION_NAME(I16),
PRECISION_NAME(I32), PRECISION_NAME(I64), PRECISION_NAME(U64), PRECISION_NAME(U16),
PRECISION_NAME(I32), PRECISION_NAME(I64), PRECISION_NAME(U64), PRECISION_NAME(U16),
PRECISION_NAME(FP32), PRECISION_NAME(FP16), PRECISION_NAME(MIXED), PRECISION_NAME(BIN),
PRECISION_NAME(BOOL),
PRECISION_NAME(BOOL), PRECISION_NAME(BF16),
#undef PRECISION_NAME
};
auto i = names.find(str);
@ -260,6 +262,7 @@ protected:
switch (v) {
CASE(FP32);
CASE(FP16);
CASE(BF16);
CASE(I16);
CASE(I32);
CASE(I64);
@ -295,6 +298,10 @@ struct PrecisionTrait<Precision::FP16> {
using value_type = int16_t;
};
template <>
struct PrecisionTrait<Precision::BF16> {
using value_type = int16_t;
};
template<>
struct PrecisionTrait<Precision::Q78> {
using value_type = uint16_t;
};

7
inference-engine/samples/benchmark_app/README.md
View File

@ -92,14 +92,17 @@ Options:
Please note that although the automatic selection usually provides a reasonable performance,
it still may be non-optimal for some cases, especially for very small networks.
-nthreads "<integer>" Optional. Number of threads to use for inference on the CPU (including HETERO and MULTI cases).
-pin "YES"/"NUMA"/"NO" Optional. Enable threads->cores ("YES", default), threads->(NUMA)nodes ("NUMA") or completely disable ("NO")
CPU threads pinning for CPU-involved inference.
-enforcebf16 Optional. Enforcing of floating point operations execution in bfloat16 precision where it is acceptable.
-pin "YES"/"NO"/"NUMA" Optional. Enable threads->cores ("YES", default), threads->(NUMA)nodes ("NUMA") or completely disable ("NO") CPU threads pinning for CPU-involved inference.
Statistics dumping options:
-report_type "<type>" Optional. Enable collecting statistics report. "no_counters" report contains configuration options specified, resulting FPS and latency. "average_counters" report extends "no_counters" report and additionally includes average PM counters values for each layer from the network. "detailed_counters" report extends "average_counters" report and additionally includes per-layer PM counters and latency for each executed infer request.
-report_folder Optional. Path to a folder where statistics report is stored.
-exec_graph_path Optional. Path to a file where to store executable graph information serialized.
-pc Optional. Report performance counters.
-dump_config Optional. Path to XML/YAML/JSON file to dump IE parameters, which were set by application.
-load_config Optional. Path to XML/YAML/JSON file to load custom IE parameters. Please note, command line parameters have higher priority then parameters from configuration file.
```
Running the application with the empty list of options yields the usage message given above and an error message.

30
inference-engine/samples/benchmark_app/benchmark_app.hpp
View File

@ -48,6 +48,9 @@ static const char infer_num_streams_message[] = "Optional. Number of streams to
"usually provides a reasonable performance, it still may be non - optimal for some cases, especially for "
"very small networks. See sample's README for more details.";
/// @brief message for enforcing of BF16 execution where it is possible
static const char enforce_bf16_message[] = "Optional. Enforcing of floating point operations execution in bfloat16 precision where it is acceptable.";
/// @brief message for user library argument
static const char custom_cpu_library_message[] = "Required for CPU custom layers. Absolute path to a shared library with the kernels implementations.";
@ -85,6 +88,15 @@ static const char progress_message[] = "Optional. Show progress bar (can affect
// @brief message for performance counters option
static const char pc_message[] = "Optional. Report performance counters.";
#ifdef USE_OPENCV
// @brief message for load config option
static const char load_config_message[] = "Optional. Path to XML/YAML/JSON file to load custom IE parameters."
" Please note, command line parameters have higher priority then parameters from configuration file.";
// @brief message for dump config option
static const char dump_config_message[] = "Optional. Path to XML/YAML/JSON file to dump IE parameters, which were set by application.";
#endif
/// @brief Define flag for showing help message <br>
DEFINE_bool(h, false, help_message);
@ -130,6 +142,9 @@ DEFINE_uint32(nthreads, 0, infer_num_threads_message);
/// @brief Number of streams to use for inference on the CPU (also affects Hetero cases)
DEFINE_string(nstreams, "", infer_num_streams_message);
/// @brief Enforces bf16 execution with bfloat16 precision on systems having this capability
DEFINE_bool(enforcebf16, false, enforce_bf16_message);
/// @brief Define parameter for batch size <br>
/// Default is 0 (that means don't specify)
DEFINE_uint32(b, 0, batch_size_message);
@ -155,6 +170,14 @@ DEFINE_bool(progress, false, progress_message);
/// @brief Define flag for showing performance counters <br>
DEFINE_bool(pc, false, pc_message);
#ifdef USE_OPENCV
/// @brief Define flag for loading configuration file <br>
DEFINE_string(load_config, "", load_config_message);
/// @brief Define flag for dumping configuration file <br>
DEFINE_string(dump_config, "", dump_config_message);
#endif
/**
* @brief This function show a help message
*/
@ -180,10 +203,15 @@ static void showUsage() {
std::cout << std::endl << " device-specific performance options:" << std::endl;
std::cout << " -nstreams \"<integer>\" " << infer_num_streams_message << std::endl;
std::cout << " -nthreads \"<integer>\" " << infer_num_threads_message << std::endl;
std::cout << " -pin \"YES\"/\"NO\" " << infer_threads_pinning_message << std::endl;
std::cout << " -enforcebf16 " << enforce_bf16_message << std::endl;
std::cout << " -pin \"YES\"/\"NO\"/\"NUMA\" " << infer_threads_pinning_message << std::endl;
std::cout << std::endl << " Statistics dumping options:" << std::endl;
std::cout << " -report_type \"<type>\" " << report_type_message << std::endl;
std::cout << " -report_folder " << report_folder_message << std::endl;
std::cout << " -exec_graph_path " << exec_graph_path_message << std::endl;
std::cout << " -pc " << pc_message << std::endl;
#ifdef USE_OPENCV
std::cout << " -dump_config " << dump_config_message << std::endl;
std::cout << " -load_config " << load_config_message << std::endl;
#endif
}

263
inference-engine/samples/benchmark_app/main.cpp
View File

@ -55,9 +55,9 @@ bool ParseAndCheckCommandLine(int argc, char *argv[]) {
}
if (!FLAGS_report_type.empty() &&
FLAGS_report_type != noCntReport && FLAGS_report_type != averageCntReport && FLAGS_report_type != detailedCntReport) {
FLAGS_report_type != noCntReport && FLAGS_report_type != averageCntReport && FLAGS_report_type != detailedCntReport) {
std::string err = "only " + std::string(noCntReport) + "/" + std::string(averageCntReport) + "/" + std::string(detailedCntReport) +
" report types are supported (invalid -report_type option value)";
" report types are supported (invalid -report_type option value)";
throw std::logic_error(err);
}
@ -71,17 +71,17 @@ bool ParseAndCheckCommandLine(int argc, char *argv[]) {
static void next_step(const std::string additional_info = "") {
static size_t step_id = 0;
static const std::map<size_t, std::string> step_names = {
{ 1, "Parsing and validating input arguments" },
{ 2, "Loading Inference Engine" },
{ 3, "Setting device configuration" },
{ 4, "Reading the Intermediate Representation network" },
{ 5, "Resizing network to match image sizes and given batch" },
{ 6, "Configuring input of the model" },
{ 7, "Loading the model to the device" },
{ 8, "Setting optimal runtime parameters" },
{ 9, "Creating infer requests and filling input blobs with images" },
{ 10, "Measuring performance" },
{ 11, "Dumping statistics report" }
{ 1, "Parsing and validating input arguments" },
{ 2, "Loading Inference Engine" },
{ 3, "Setting device configuration" },
{ 4, "Reading the Intermediate Representation network" },
{ 5, "Resizing network to match image sizes and given batch" },
{ 6, "Configuring input of the model" },
{ 7, "Loading the model to the device" },
{ 8, "Setting optimal runtime parameters" },
{ 9, "Creating infer requests and filling input blobs with images" },
{ 10, "Measuring performance" },
{ 11, "Dumping statistics report" }
};
step_id++;
@ -121,38 +121,46 @@ int main(int argc, char *argv[]) {
slog::info << "Network is compiled" << slog::endl;
}
if (!FLAGS_report_type.empty()) {
std::vector<gflags::CommandLineFlagInfo> flags;
StatisticsReport::Parameters command_line_arguments;
gflags::GetAllFlags(&flags);
for (auto &flag : flags) {
if (!flag.is_default) {
command_line_arguments.push_back({ flag.name, flag.current_value });
}
std::vector<gflags::CommandLineFlagInfo> flags;
StatisticsReport::Parameters command_line_arguments;
gflags::GetAllFlags(&flags);
for (auto &flag : flags) {
if (!flag.is_default) {
command_line_arguments.push_back({ flag.name, flag.current_value });
}
}
if (!FLAGS_report_type.empty()) {
statistics = std::make_shared<StatisticsReport>(StatisticsReport::Config{FLAGS_report_type, FLAGS_report_folder});
statistics->addParameters(StatisticsReport::Category::COMMAND_LINE_PARAMETERS, command_line_arguments);
}
auto isFlagSetInCommandLine = [&command_line_arguments] (const std::string& name) {
return (std::find_if(command_line_arguments.begin(), command_line_arguments.end(),
[ name ] (const std::pair<std::string, std::string>& p) { return p.first == name;}) != command_line_arguments.end());
};
std::string device_name = FLAGS_d;
// Parse devices
auto devices = parseDevices(device_name);
// Parse nstreams per device
std::map<std::string, std::string> device_nstreams = parseNStreamsValuePerDevice(devices, FLAGS_nstreams);
// Load device config file if specified
std::map<std::string, std::map<std::string, std::string>> config;
#ifdef USE_OPENCV
if (!FLAGS_load_config.empty()) {
load_config(FLAGS_load_config, config);
}
#endif
/** This vector stores paths to the processed images **/
std::vector<std::string> inputFiles;
parseInputFilesArguments(inputFiles);
if (FLAGS_nstreams.empty()) {
slog::warn << "-nstreams default value is determined automatically for a device. "
"Although the automatic selection usually provides a reasonable performance,"
"but it still may be non-optimal for some cases, for more information look at README." << slog::endl<< slog::endl;
}
// ----------------- 2. Loading the Inference Engine -----------------------------------------------------------
next_step();
// Get optimal runtime parameters for device
std::string device_name = FLAGS_d;
Core ie;
if (FLAGS_d.find("CPU") != std::string::npos && !FLAGS_l.empty()) {
// CPU (MKLDNN) extensions is loaded as a shared library and passed as a pointer to base extension
const auto extension_ptr = InferenceEngine::make_so_pointer<InferenceEngine::IExtension>(FLAGS_l);
@ -160,10 +168,17 @@ int main(int argc, char *argv[]) {
slog::info << "CPU (MKLDNN) extensions is loaded " << FLAGS_l << slog::endl;
}
// Load clDNN Extensions
if ((FLAGS_d.find("GPU") != std::string::npos) && !FLAGS_c.empty()) {
// Load clDNN Extensions
ie.SetConfig({ {CONFIG_KEY(CONFIG_FILE), FLAGS_c} });
slog::info << "GPU extensions is loaded " << FLAGS_c << slog::endl;
// Override config if command line parameter is specified
if (!config.count("GPU"))
config["GPU"] = {};
config["GPU"][CONFIG_KEY(CONFIG_FILE)] = FLAGS_c;
}
if (config.count("GPU") && config.at("GPU").count(CONFIG_KEY(CONFIG_FILE))) {
auto ext = config.at("GPU").at(CONFIG_KEY(CONFIG_FILE));
ie.SetConfig({{ CONFIG_KEY(CONFIG_FILE), ext }}, "GPU");
slog::info << "GPU extensions is loaded " << ext << slog::endl;
}
slog::info << "InferenceEngine: " << GetInferenceEngineVersion() << slog::endl;
@ -173,70 +188,108 @@ int main(int argc, char *argv[]) {
// ----------------- 3. Setting device configuration -----------------------------------------------------------
next_step();
bool perf_counts = (FLAGS_report_type == detailedCntReport ||
FLAGS_report_type == averageCntReport ||
FLAGS_pc ||
!FLAGS_exec_graph_path.empty());
auto devices = parseDevices(device_name);
std::map<std::string, uint32_t> device_nstreams = parseNStreamsValuePerDevice(devices, FLAGS_nstreams);
for (auto& pair : device_nstreams) {
auto key = std::string(pair.first + "_THROUGHPUT_STREAMS");
std::vector<std::string> supported_config_keys = ie.GetMetric(pair.first, METRIC_KEY(SUPPORTED_CONFIG_KEYS));
if (std::find(supported_config_keys.begin(), supported_config_keys.end(), key) == supported_config_keys.end()) {
throw std::logic_error("Device " + pair.first + " doesn't support config key '" + key + "'! " +
"Please specify -nstreams for correct devices in format <dev1>:<nstreams1>,<dev2>:<nstreams2>");
}
}
bool perf_counts = false;
// Update config per device according to command line parameters
for (auto& device : devices) {
if (!config.count(device)) config[device] = {};
std::map<std::string, std::string>& device_config = config.at(device);
// Set performance counter
if (isFlagSetInCommandLine("pc")) {
// set to user defined value
device_config[CONFIG_KEY(PERF_COUNT)] = FLAGS_pc ? CONFIG_VALUE(YES) : CONFIG_VALUE(NO);
} else if (device_config.count(CONFIG_KEY(PERF_COUNT)) &&
(device_config.at(CONFIG_KEY(PERF_COUNT)) == "YES")) {
slog::warn << "Performance counters for " << device <<
" device is turned on. To print results use -pc option." << slog::endl;
} else if (FLAGS_report_type == detailedCntReport || FLAGS_report_type == averageCntReport) {
slog::warn << "Turn on performance counters for " << device <<
" device since report type is " << FLAGS_report_type << "." << slog::endl;
device_config[CONFIG_KEY(PERF_COUNT)] = CONFIG_VALUE(YES);
} else if (!FLAGS_exec_graph_path.empty()) {
slog::warn << "Turn on performance counters for " << device <<
" device due to execution graph dumping." << slog::endl;
device_config[CONFIG_KEY(PERF_COUNT)] = CONFIG_VALUE(YES);
} else {
// set to default value
device_config[CONFIG_KEY(PERF_COUNT)] = FLAGS_pc ? CONFIG_VALUE(YES) : CONFIG_VALUE(NO);
}
perf_counts = (device_config.at(CONFIG_KEY(PERF_COUNT)) == CONFIG_VALUE(YES)) ? true : perf_counts;
auto setThroughputStreams = [&] () {
const std::string key = device + "_THROUGHPUT_STREAMS";
if (device_nstreams.count(device)) {
// set to user defined value
std::vector<std::string> supported_config_keys = ie.GetMetric(device, METRIC_KEY(SUPPORTED_CONFIG_KEYS));
if (std::find(supported_config_keys.begin(), supported_config_keys.end(), key) == supported_config_keys.end()) {
throw std::logic_error("Device " + device + " doesn't support config key '" + key + "'! " +
"Please specify -nstreams for correct devices in format <dev1>:<nstreams1>,<dev2>:<nstreams2>" +
" or via configuration file.");
}
device_config[key] = device_nstreams.at(device);
} else if (!device_config.count(key) && (FLAGS_api == "async")) {
slog::warn << "-nstreams default value is determined automatically for " << device << " device. "
"Although the automatic selection usually provides a reasonable performance,"
"but it still may be non-optimal for some cases, for more information look at README." << slog::endl;
device_config[key] = std::string(device + "_THROUGHPUT_AUTO");
}
if (device_config.count(key))
device_nstreams[device] = device_config.at(key);
};
if (device == "CPU") { // CPU supports few special performance-oriented keys
// limit threading for CPU portion of inference
if (FLAGS_nthreads != 0)
ie.SetConfig({{ CONFIG_KEY(CPU_THREADS_NUM), std::to_string(FLAGS_nthreads) }}, device);
if (isFlagSetInCommandLine("nthreads"))
device_config[CONFIG_KEY(CPU_THREADS_NUM)] = std::to_string(FLAGS_nthreads);
if ((device_name.find("MULTI") != std::string::npos) &&
(device_name.find("GPU") != std::string::npos)) {
ie.SetConfig({{ CONFIG_KEY(CPU_BIND_THREAD), CONFIG_VALUE(NO) }}, device);
} else {
// pin threads for CPU portion of inference
ie.SetConfig({{ CONFIG_KEY(CPU_BIND_THREAD), FLAGS_pin }}, device);
if (isFlagSetInCommandLine("enforcebf16"))
device_config[CONFIG_KEY(ENFORCE_BF16)] = FLAGS_enforcebf16 ? CONFIG_VALUE(YES) : CONFIG_VALUE(NO);
if (isFlagSetInCommandLine("pin")) {
// set to user defined value
device_config[CONFIG_KEY(CPU_BIND_THREAD)] = FLAGS_pin;
} else if (!device_config.count(CONFIG_KEY(CPU_BIND_THREAD))) {
if ((device_name.find("MULTI") != std::string::npos) &&
(device_name.find("GPU") != std::string::npos)) {
slog::warn << "Turn off threads pinning for " << device <<
" device since multi-scenario with GPU device is used." << slog::endl;
device_config[CONFIG_KEY(CPU_BIND_THREAD)] = CONFIG_VALUE(NO);
} else {
// set to default value
device_config[CONFIG_KEY(CPU_BIND_THREAD)] = FLAGS_pin;
}
}
// for CPU execution, more throughput-oriented execution via streams
if (FLAGS_api == "async")
ie.SetConfig({{ CONFIG_KEY(CPU_THROUGHPUT_STREAMS),
(device_nstreams.count(device) > 0 ? std::to_string(device_nstreams.at(device)) :
"CPU_THROUGHPUT_AUTO") }}, device);
device_nstreams[device] = std::stoi(ie.GetConfig(device, CONFIG_KEY(CPU_THROUGHPUT_STREAMS)).as<std::string>());
setThroughputStreams();
} else if (device == ("GPU")) {
if (FLAGS_api == "async")
ie.SetConfig({{ CONFIG_KEY(GPU_THROUGHPUT_STREAMS),
(device_nstreams.count(device) > 0 ? std::to_string(device_nstreams.at(device)) :
"GPU_THROUGHPUT_AUTO") }}, device);
device_nstreams[device] = std::stoi(ie.GetConfig(device, CONFIG_KEY(GPU_THROUGHPUT_STREAMS)).as<std::string>());
// for GPU execution, more throughput-oriented execution via streams
setThroughputStreams();
if ((device_name.find("MULTI") != std::string::npos) &&
(device_name.find("CPU") != std::string::npos)) {
// multi-device execution with the CPU + GPU performs best with GPU trottling hint,
// which releases another CPU thread (that is otherwise used by the GPU driver for active polling)
ie.SetConfig({{ CLDNN_CONFIG_KEY(PLUGIN_THROTTLE), "1" }}, "GPU");
slog::warn << "Turn on GPU trottling. Multi-device execution with the CPU + GPU performs best with GPU trottling hint," <<
"which releases another CPU thread (that is otherwise used by the GPU driver for active polling)"<< slog::endl;
device_config[CLDNN_CONFIG_KEY(PLUGIN_THROTTLE)] = "1";
}
} else if (device == "MYRIAD") {
ie.SetConfig({{ CONFIG_KEY(LOG_LEVEL), CONFIG_VALUE(LOG_WARNING) }}, device);
device_config[CONFIG_KEY(LOG_LEVEL)] = CONFIG_VALUE(LOG_WARNING);
}
}
for (auto&& item : config) {
ie.SetConfig(item.second, item.first);
}
auto double_to_string = [] (const double number) {
std::stringstream ss;
ss << std::fixed << std::setprecision(2) << number;
return ss.str();
};
std::stringstream ss;
ss << std::fixed << std::setprecision(2) << number;
return ss.str();
};
auto get_total_ms_time = [] (Time::time_point& startTime) {
return std::chrono::duration_cast<ns>(Time::now() - startTime).count() * 0.000001;
};
size_t batchSize = FLAGS_b;
Precision precision = Precision::UNSPECIFIED;
std::string topology_name = "";
@ -253,7 +306,7 @@ int main(int argc, char *argv[]) {
if (statistics)
statistics->addParameters(StatisticsReport::Category::EXECUTION_RESULTS,
{
{"read network time (ms)", duration_ms}
{"read network time (ms)", duration_ms}
});
const InputsDataMap inputInfo(cnnNetwork.getInputsInfo());
@ -305,17 +358,14 @@ int main(int argc, char *argv[]) {
}
// ----------------- 7. Loading the model to the device --------------------------------------------------------
next_step();
std::map<std::string, std::string> config = {{ CONFIG_KEY(PERF_COUNT), perf_counts ? CONFIG_VALUE(YES) :
CONFIG_VALUE(NO) }};
startTime = Time::now();
exeNetwork = ie.LoadNetwork(cnnNetwork, device_name, config);
exeNetwork = ie.LoadNetwork(cnnNetwork, device_name);
duration_ms = double_to_string(get_total_ms_time(startTime));
slog::info << "Load network took " << duration_ms << " ms" << slog::endl;
if (statistics)
statistics->addParameters(StatisticsReport::Category::EXECUTION_RESULTS,
{
{"load network time (ms)", duration_ms}
{"load network time (ms)", duration_ms}
});
} else {
next_step();
@ -333,7 +383,7 @@ int main(int argc, char *argv[]) {
if (statistics)
statistics->addParameters(StatisticsReport::Category::EXECUTION_RESULTS,
{
{"import network time (ms)", duration_ms}
{"import network time (ms)", duration_ms}
});
if (batchSize == 0) {
batchSize = 1;
@ -342,6 +392,12 @@ int main(int argc, char *argv[]) {
// ----------------- 8. Setting optimal runtime parameters -----------------------------------------------------
next_step();
// Update number of streams
for (auto&& ds : device_nstreams) {
const std::string key = ds.first + "_THROUGHPUT_STREAMS";
device_nstreams[ds.first] = ie.GetConfig(ds.first, key).as<std::string>();
}
// Number of requests
uint32_t nireq = FLAGS_nireq;
if (nireq == 0) {
@ -384,21 +440,21 @@ int main(int argc, char *argv[]) {
if (statistics) {
statistics->addParameters(StatisticsReport::Category::RUNTIME_CONFIG,
{
{"topology", topology_name},
{"target device", device_name},
{"API", FLAGS_api},
{"precision", std::string(precision.name())},
{"batch size", std::to_string(batchSize)},
{"number of iterations", std::to_string(niter)},
{"number of parallel infer requests", std::to_string(nireq)},
{"duration (ms)", std::to_string(getDurationInMilliseconds(duration_seconds))},
{"topology", topology_name},
{"target device", device_name},
{"API", FLAGS_api},
{"precision", std::string(precision.name())},
{"batch size", std::to_string(batchSize)},
{"number of iterations", std::to_string(niter)},
{"number of parallel infer requests", std::to_string(nireq)},
{"duration (ms)", std::to_string(getDurationInMilliseconds(duration_seconds))},
});
for (auto& nstreams : device_nstreams) {
std::stringstream ss;
ss << "number of " << nstreams.first << " streams";
statistics->addParameters(StatisticsReport::Category::RUNTIME_CONFIG,
{
{ss.str(), std::to_string(nstreams.second)},
{ss.str(), nstreams.second},
});
}
}
@ -511,23 +567,23 @@ int main(int argc, char *argv[]) {
double latency = getMedianValue<double>(inferRequestsQueue.getLatencies());
double totalDuration = inferRequestsQueue.getDurationInMilliseconds();
double fps = (FLAGS_api == "sync") ? batchSize * 1000.0 / latency :
batchSize * 1000.0 * iteration / totalDuration;
batchSize * 1000.0 * iteration / totalDuration;
if (statistics) {
statistics->addParameters(StatisticsReport::Category::EXECUTION_RESULTS,
{
{"total execution time (ms)", double_to_string(totalDuration)},
{"total number of iterations", std::to_string(iteration)},
{"total execution time (ms)", double_to_string(totalDuration)},
{"total number of iterations", std::to_string(iteration)},
});
if (device_name.find("MULTI") == std::string::npos) {
statistics->addParameters(StatisticsReport::Category::EXECUTION_RESULTS,
{
{"latency (ms)", double_to_string(latency)},
{"latency (ms)", double_to_string(latency)},
});
}
statistics->addParameters(StatisticsReport::Category::EXECUTION_RESULTS,
{
{"throughput", double_to_string(fps)}
{"throughput", double_to_string(fps)}
});
}
@ -536,6 +592,13 @@ int main(int argc, char *argv[]) {
// ----------------- 11. Dumping statistics report -------------------------------------------------------------
next_step();
#ifdef USE_OPENCV
if (!FLAGS_dump_config.empty()) {
dump_config(FLAGS_dump_config, config);
slog::info << "Inference Engine configuration settings were dumped to " << FLAGS_dump_config << slog::endl;
}
#endif
if (!FLAGS_exec_graph_path.empty()) {
try {
CNNNetwork execGraphInfo = exeNetwork.GetExecGraphInfo();
@ -575,7 +638,7 @@ int main(int argc, char *argv[]) {
if (statistics) {
statistics->addParameters(StatisticsReport::Category::EXECUTION_RESULTS,
{
{"error", ex.what()},
{"error", ex.what()},
});
statistics->dump();
}

55
inference-engine/samples/benchmark_app/utils.cpp
View File

@ -13,6 +13,10 @@
#include "utils.hpp"
#ifdef USE_OPENCV
#include <opencv2/core.hpp>
#endif
uint32_t deviceDefaultDeviceDurationInSeconds(const std::string& device) {
static const std::map<std::string, uint32_t> deviceDefaultDurationInSeconds {
{ "CPU", 60 },
@ -60,32 +64,33 @@ std::vector<std::string> parseDevices(const std::string& device_string) {
if (comma_separated_devices.find(":") != std::string::npos) {
comma_separated_devices = comma_separated_devices.substr(comma_separated_devices.find(":") + 1);
}
if ((comma_separated_devices == "MULTI") || (comma_separated_devices == "HETERO"))
return std::vector<std::string>();
auto devices = split(comma_separated_devices, ',');
for (auto& device : devices)
device = device.substr(0, device.find_first_of(".("));
return devices;
}
std::map<std::string, uint32_t> parseNStreamsValuePerDevice(const std::vector<std::string>& devices,
const std::string& values_string) {
std::map<std::string, std::string> parseNStreamsValuePerDevice(const std::vector<std::string>& devices,
const std::string& values_string) {
// Format: <device1>:<value1>,<device2>:<value2> or just <value>
auto values_string_upper = values_string;
std::map<std::string, uint32_t> result;
auto device_value_strings = split(values_string_upper, ',');
std::map<std::string, std::string> result;
auto device_value_strings = split(values_string, ',');
for (auto& device_value_string : device_value_strings) {
auto device_value_vec = split(device_value_string, ':');
auto device_value_vec = split(device_value_string, ':');
if (device_value_vec.size() == 2) {
auto device_name = device_value_vec.at(0);
auto nstreams = device_value_vec.at(1);
auto it = std::find(devices.begin(), devices.end(), device_name);
if (it != devices.end()) {
result[device_name] = std::stoi(nstreams);
result[device_name] = nstreams;
} else {
throw std::logic_error("Can't set nstreams value " + std::string(nstreams) +
" for device '" + device_name + "'! Incorrect device name!");
}
} else if (device_value_vec.size() == 1) {
uint32_t value = std::stoi(device_value_vec.at(0));
auto value = device_value_vec.at(0);
for (auto& device : devices) {
result[device] = value;
}
@ -95,3 +100,37 @@ std::map<std::string, uint32_t> parseNStreamsValuePerDevice(const std::vector<st
}
return result;
}
#ifdef USE_OPENCV
void dump_config(const std::string& filename,
const std::map<std::string, std::map<std::string, std::string>>& config) {
cv::FileStorage fs(filename, cv::FileStorage::WRITE);
if (!fs.isOpened())
throw std::runtime_error("Error: Can't open config file : " + filename);
for (auto device_it = config.begin(); device_it != config.end(); ++device_it) {
fs << device_it->first << "{:";
for (auto param_it = device_it->second.begin(); param_it != device_it->second.end(); ++param_it)
fs << param_it->first << param_it->second;
fs << "}";
}
fs.release();
}
void load_config(const std::string& filename,
std::map<std::string, std::map<std::string, std::string>>& config) {
cv::FileStorage fs(filename, cv::FileStorage::READ);
if (!fs.isOpened())
throw std::runtime_error("Error: Can't load config file : " + filename);
cv::FileNode root = fs.root();
for (auto it = root.begin(); it != root.end(); ++it) {
auto device = *it;
if (!device.isMap()) {
throw std::runtime_error("Error: Can't parse config file : " + filename);
}
for (auto iit = device.begin(); iit != device.end(); ++iit) {
auto item = *iit;
config[device.name()][item.name()] = item.string();
}
}
}
#endif

10
inference-engine/samples/benchmark_app/utils.hpp
View File

@ -10,5 +10,11 @@
std::vector<std::string> parseDevices(const std::string& device_string);
uint32_t deviceDefaultDeviceDurationInSeconds(const std::string& device);
std::map<std::string, uint32_t> parseNStreamsValuePerDevice(const std::vector<std::string>& devices,
const std::string& values_string);
std::map<std::string, std::string> parseNStreamsValuePerDevice(const std::vector<std::string>& devices,
const std::string& values_string);
#ifdef USE_OPENCV
void dump_config(const std::string& filename,
const std::map<std::string, std::map<std::string, std::string>>& config);
void load_config(const std::string& filename,
std::map<std::string, std::map<std::string, std::string>>& config);
#endif

30
inference-engine/samples/hello_query_device/main.cpp
View File

@ -85,30 +85,28 @@ int main(int argc, char *argv[]) {
std::vector<std::string> availableDevices = ie.GetAvailableDevices();
// --------------------------- 3. Query and print supported metrics and config keys--------------------
std::set<std::string> printedDevices;
std::cout << "Available devices: " << std::endl;
for (auto && device : availableDevices) {
std::string deviceFamilyName = device.substr(0, device.find_first_of('.'));
if (printedDevices.find(deviceFamilyName) == printedDevices.end())
printedDevices.insert(deviceFamilyName);
else
continue;
std::cout << "\tDevice: " << deviceFamilyName << std::endl;
std::cout << "\tDevice: " << device << std::endl;
std::cout << "\tMetrics: " << std::endl;
std::vector<std::string> supportedMetrics = ie.GetMetric(deviceFamilyName, METRIC_KEY(SUPPORTED_METRICS));
std::vector<std::string> supportedMetrics = ie.GetMetric(device, METRIC_KEY(SUPPORTED_METRICS));
for (auto && metricName : supportedMetrics) {
std::cout << "\t\t" << metricName << " : " << std::flush;
printParameterValue(ie.GetMetric(device, metricName));
if (metricName != METRIC_KEY(AVAILABLE_DEVICES)) {
std::cout << "\t\t" << metricName << " : " << std::flush;
printParameterValue(ie.GetMetric(device, metricName));
}
}
std::cout << "\tDefault values for device configuration keys: " << std::endl;
std::vector<std::string> supportedConfigKeys = ie.GetMetric(deviceFamilyName, METRIC_KEY(SUPPORTED_CONFIG_KEYS));
for (auto && configKey : supportedConfigKeys) {
std::cout << "\t\t" << configKey << " : " << std::flush;
printParameterValue(ie.GetConfig(deviceFamilyName, configKey));
if (std::find(supportedMetrics.begin(), supportedMetrics.end(),
METRIC_KEY(SUPPORTED_CONFIG_KEYS)) != supportedMetrics.end()) {
std::cout << "\tDefault values for device configuration keys: " << std::endl;
std::vector<std::string> supportedConfigKeys = ie.GetMetric(device, METRIC_KEY(SUPPORTED_CONFIG_KEYS));
for (auto && configKey : supportedConfigKeys) {
std::cout << "\t\t" << configKey << " : " << std::flush;
printParameterValue(ie.GetConfig(device, configKey));
}
}
std::cout << std::endl;

5
inference-engine/samples/speech_sample/main.cpp
View File

@ -643,12 +643,13 @@ int main(int argc, char *argv[]) {
auto t0 = Time::now();
ExecutableNetwork executableNet;
ie.SetConfig(genericPluginConfig, deviceStr);
if (!FLAGS_m.empty()) {
slog::info << "Loading model to the device" << slog::endl;
executableNet = ie.LoadNetwork(network, deviceStr, genericPluginConfig);
executableNet = ie.LoadNetwork(network, deviceStr);
} else {
slog::info << "Importing model to the device" << slog::endl;
executableNet = ie.ImportNetwork(FLAGS_rg.c_str(), deviceStr, genericPluginConfig);
executableNet = ie.ImportNetwork(FLAGS_rg.c_str(), deviceStr);
}
ms loadTime = std::chrono::duration_cast<ms>(Time::now() - t0);

83
inference-engine/scripts/run_code_checks.sh Normal file
View File

@ -0,0 +1,83 @@
#!/bin/bash
CURRENT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
command -v realpath >/dev/null 2>&1 || { echo >&2 "cpplint require realpath executable but it's not installed. Aborting."; exit 1; }
SOURCE_DIR=$(realpath ${CURRENT_DIR}/..)
REPORT_DIR="${SOURCE_DIR}/report"
CPPLINT_REPORT_DIR="${REPORT_DIR}/cpplint"
PROJECT_NAME="Inference Engine"
function run_cpplint() {
echo "-> CppLint started..."
if [ -d ${CPPLINT_REPORT_DIR} ]; then
rm -Rf ${CPPLINT_REPORT_DIR}
fi
mkdir -p ${CPPLINT_REPORT_DIR}
python ${CURRENT_DIR}/cpplint.py --linelength=160 --counting=detailed --quiet --filter="
-build/header_guard,
-build/include,
-build/include_order,
-build/include_subdir,
-build/include_what_you_use,
-build/namespaces,
-build/c++11,
-whitespace/indent,
-whitespace/comments,
-whitespace/ending_newline,
-runtime/references,
-runtime/int,
-runtime/explicit,
-readability/todo,
-readability/fn_size
" $(find ${SOURCE_DIR} -name '*.h' -or -name '*.cc' -or -name '*.c' -or -name '*.cpp' -or -name '*.hpp' |
grep -v 'inference-engine/bin\|inference-engine/build\|inference-engine/report\|inference-engine/scripts\|inference-engine/temp\|inference-engine/tests_deprecated/\|gtest\|inference-engine/ie_bridges\|pugixml\|inference-engine/tools/vpu_perfcheck\|thirdparty/gflags\|thirdparty/ade\|thirdparty/fluid\|thirdparty/mkl-dnn\|thirdparty/movidius\|thirdparty/ocv\|thirdparty/plugixml\|thirdparty/std_lib\|thirdparty/clDNN/common\|thirdparty/clDNN/tutorial\|thirdparty/clDNN/utils' |
grep 'include\|src\|inference-engine/samples\|thirdparty/clDNN/kernel_selector\|thirdparty/clDNN/api\|thirdparty/clDNN/api_extension\|inference-engine/tests_' ) 2>&1 |
sed 's/"/\&quot;/g' >&1| sed 's/</\&lt;/g' >&1| sed 's/>/\&gt;/g' >&1| sed "s/'/\&apos;/g" >&1|
sed 's/\&/\&amp;/g' >&1| python ${CURRENT_DIR}/cpplint_to_cppcheckxml.py &> ${CPPLINT_REPORT_DIR}/cpplint-cppcheck-result.xml
# Generate html from it
${CURRENT_DIR}/cppcheck-htmlreport.py --file=${CPPLINT_REPORT_DIR}/cpplint-cppcheck-result.xml --report-dir=${CPPLINT_REPORT_DIR} --source-dir=${SOURCE_DIR} --title=${PROJECT_NAME}
# Change Cppcheck things to cpplint
sed -i.bak 's/Cppcheck/cpplint/g' ${CPPLINT_REPORT_DIR}/index.html
sed -i.bak 's/a\ tool\ for\ static\ C\/C++\ code\ analysis/an\ open\ source\ lint\-like\ tool\ from\ Google/g' ${CPPLINT_REPORT_DIR}/index.html
sed -i.bak 's/http:\/\/cppcheck.sourceforge.net/http:\/\/google\-styleguide.googlecode.com\/svn\/trunk\/cpplint\/cpplint.py/g' ${CPPLINT_REPORT_DIR}/index.html
sed -i.bak 's/IRC: <a href=\"irc:\/\/irc.freenode.net\/cppcheck\">irc:\/\/irc.freenode.net\/cppcheck<\/a>/\ /g' ${CPPLINT_REPORT_DIR}/index.html
echo "-> CppLint finished..."
}
function run_cpp_check() {
echo "-> Cppcheck started..."
CPPCHECK_REPORT_DIR="${REPORT_DIR}/cppcheck"
if [ -d ${CPPCHECK_REPORT_DIR} ]; then
rm -Rf ${CPPCHECK_REPORT_DIR}
fi
mkdir -p ${CPPCHECK_REPORT_DIR}
# Generate cppcheck xml
cppcheck -v --enable=all --suppress=missingIncludeSystem --std=c++11 ${SOURCE_DIR} -i${SOURCE_DIR}/thirdparty -i${SOURCE_DIR}/tests/libs -i${SOURCE_DIR}/temp -i${SOURCE_DIR}/build \
-i${SOURCE_DIR}/bin -i${SOURCE_DIR}/report -I${SOURCE_DIR}/include -I${SOURCE_DIR}/src -I${SOURCE_DIR}/thirdparty/pugixml/src -I${SOURCE_DIR}/thirdparty/gflags/src -I${SOURCE_DIR}/samples/scoring_agent/HTTPClient -I${SOURCE_DIR}/src/inference_engine --xml-version=2 2> ${CPPCHECK_REPORT_DIR}/cppcheck-only-result.xml
# Generate html from it
python ${CURRENT_DIR}/cppcheck-htmlreport.py\
--file=${CPPCHECK_REPORT_DIR}/cppcheck-only-result.xml\
--report-dir=${CPPCHECK_REPORT_DIR}\
--source-dir=${SOURCE_DIR}\
--title=${PROJECT_NAME}
echo "-> Cppcheck finished..."
}
if [ ! -d ${REPORT_DIR} ]; then
mkdir -p ${REPORT_DIR}
fi
run_cpplint
out_cpp_lint=`cat ${CPPLINT_REPORT_DIR}/cpplint-cppcheck-result.xml`
if [[ ${out_cpp_lint} == *"error"* ]]; then
exit 1
fi
#run_cpp_check

2
inference-engine/src/CMakeLists.txt
View File

@ -4,6 +4,8 @@
add_subdirectory(preprocessing)
add_subdirectory(ir_readers)
add_subdirectory(legacy_api)
if(ENABLE_MKL_DNN)

1
inference-engine/src/cldnn_engine/cldnn_config.cpp
View File

@ -269,5 +269,6 @@ void Config::adjustKeyMapValues() {
key_config_map[PluginConfigParams::KEY_GPU_THROUGHPUT_STREAMS] = std::to_string(throughput_streams);
key_config_map[PluginConfigParams::KEY_DEVICE_ID] = device_id;
key_config_map[PluginConfigParams::KEY_CONFIG_FILE] = "";
}
} // namespace CLDNNPlugin

2
inference-engine/src/cldnn_engine/cldnn_program.cpp
View File

@ -2704,7 +2704,7 @@ void Program::CreatePoolingPrimitive(cldnn::topology& topology, InferenceEngine:
} else {
size = (cldnn::tensor) cldnn::spatial(TensorValue(poolLayer->_kernel[X_AXIS]), TensorValue(poolLayer->_kernel[Y_AXIS]));
stride = (cldnn::tensor) cldnn::spatial(TensorValue(poolLayer->_stride[X_AXIS]), TensorValue(poolLayer->_stride[Y_AXIS]));
input_offset = { 0, 0, -TensorValue(allPads.begin[X_AXIS]), -TensorValue(allPads.begin[Y_AXIS]) };
input_offset = { 0, 0, -TensorValue(allPads.begin[X_AXIS]), -TensorValue(allPads.begin[Y_AXIS]), 0 };
}
auto dt = DataTypeFromPrecision(poolLayer->outData[0]->getPrecision());

59
inference-engine/src/gna_plugin/backend/am_intel_dnn.cpp
View File

@ -8,6 +8,7 @@
#include <set>
#include <string>
#include <algorithm>
#include <map>
#if defined __INTEL_COMPILER || defined _MSC_VER
#include <malloc.h>
@ -28,6 +29,7 @@
#include "gna2_model_debug_log.hpp"
#else
#include <gna-api-types-xnn.h>
#include <map>
#endif
@ -373,6 +375,13 @@ float GNAPluginNS::backend::AMIntelDNN::OutputScaleFactor(intel_dnn_component_t
return comp.output_scale_factor;
}
struct InputEndPoint {
int idx = 0;
size_t size = 0;
size_t num_bytes_per_output = 1;
InputEndPoint() = default;
InputEndPoint(int nidx, size_t sz, size_t esize) : idx(nidx), size(sz), num_bytes_per_output(esize) {}
};
void GNAPluginNS::backend::AMIntelDNN::WriteGraphWizModel(const char *filename) {
auto & components = component;
@ -414,11 +423,21 @@ void GNAPluginNS::backend::AMIntelDNN::WriteGraphWizModel(const char *filename)
return ptra >= ptrb && ptra < reinterpret_cast<char*>(ptrb) + bsize;
};
auto startPtr = [](void* ptr, size_t size) {
return reinterpret_cast<int8_t*>(ptr);
};
auto endPtr = [](void* ptr, size_t size) {
return reinterpret_cast<int8_t*>(ptr) + size;
};
auto sizeofTensor = [](void* ptr, size_t size) {
return size;
};
std::fstream graph(filename, std::ios::out);
graph << "strict digraph {";
std::set<void*> weights;
std::set<void*> biases;
std::set<void*> outputs;
std::map<void*, InputEndPoint> outputs;
std::set<std::string> layersNames;
auto generate_layer_name = [&](int k) {
@ -565,11 +584,25 @@ void GNAPluginNS::backend::AMIntelDNN::WriteGraphWizModel(const char *filename)
}
}
if (!inputConnected) {
// drawing tmp connection
outputs.insert(components[k].ptr_inputs);
auto tidx = std::distance(outputs.begin(), outputs.find(components[k].ptr_inputs));
graph << tidx << " -> " << l
<< " [label=\"FROM_TMP\", fontcolor=darkgreen,color=orange, style=dashed];";
// searching for TMP connection
size_t tidx = -1;
for (auto && en : outputs) {
if (intersected(en.first, en.second.size, INPUTS(k))) {
tidx = en.second.idx;
auto updated_ptr = std::min(startPtr(en.first, en.second.size), startPtr(INPUTS(k)));
auto updated_size = std::max(endPtr(en.first, en.second.size), endPtr(INPUTS(k))) - updated_ptr;
outputs.erase(en.first);
outputs[updated_ptr] = InputEndPoint(tidx, updated_size, components[k].num_bytes_per_input);
break;
}
}
if (tidx == -1) {
outputs[components[k].ptr_inputs] = InputEndPoint(outputs.size(), sizeofTensor(INPUTS(k)), components[k].num_bytes_per_input);
}
tidx = outputs[components[k].ptr_inputs].idx;
graph << "parameter_" << tidx << " -> " << l
<< " [fontcolor=darkgreen,color=orange, style=dashed];";
}
}
@ -578,13 +611,25 @@ void GNAPluginNS::backend::AMIntelDNN::WriteGraphWizModel(const char *filename)
int tidx = 0;
for (auto tmpOutPtrs : outputs) {
if (components[k].ptr_outputs == tmpOutPtrs) {
if (components[k].ptr_outputs == tmpOutPtrs.first) {
graph << l << " -> " << tidx << " [label=\"TO_TMP\", fontcolor=darkgreen,color=orange, style=dashed];";
}
tidx++;
}
}
// writing inputs info
for (auto && en : outputs) {
std::string l = "parameter_" + std::to_string(en.second.idx);
graph << l << " [shape=box, style=filled, fillcolor=\"#85C1E9\"";
graph << ", label=<<TABLE BORDER=\"0\" CELLBORDER=\"1\" CELLSPACING=\"0\">\n"
" <TR><TD colspan=\"2\">" << l << "</TD></TR>\n";
graph << " <TR><TD> dims</TD><TD>" << 1 << "x" << en.second.size / en.second.num_bytes_per_output << "</TD></TR>\n";
graph << " <TR><TD> obit</TD><TD>" << en.second.num_bytes_per_output << "</TD></TR>\n";
graph << " <TR><TD> ptr</TD><TD>" << en.first << "</TD></TR>\n";
graph << "</TABLE>>];\n";
}
graph << "}";
}

29
inference-engine/src/gna_plugin/gna_executable_network.hpp
View File

@ -21,20 +21,28 @@ class GNAExecutableNetwork : public InferenceEngine::ExecutableNetworkThreadSafe
std::shared_ptr<GNAPlugin> plg;
public:
GNAExecutableNetwork(const std::string &aotFileName, const std::map<std::string, std::string> &config) :
plg(std::make_shared<GNAPlugin>(config)) {
GNAExecutableNetwork(const std::string &aotFileName, std::shared_ptr<GNAPlugin> plg)
: plg(plg) {
plg->ImportNetwork(aotFileName);
_networkInputs = plg->GetInputs();
_networkOutputs = plg->GetOutputs();
}
GNAExecutableNetwork(InferenceEngine::ICNNNetwork &network, const std::map<std::string, std::string> &config)
: plg(std::make_shared<GNAPlugin>(config)) {
GNAExecutableNetwork(InferenceEngine::ICNNNetwork &network, std::shared_ptr<GNAPlugin> plg)
: plg(plg) {
InferenceEngine::NetPass::ConvertPrecision(network, InferenceEngine::Precision::I64, InferenceEngine::Precision::I32);
InferenceEngine::NetPass::ConvertPrecision(network, InferenceEngine::Precision::U64, InferenceEngine::Precision::I32);
plg->LoadNetwork(network);
}
GNAExecutableNetwork(const std::string &aotFileName, const std::map<std::string, std::string> &config)
: GNAExecutableNetwork(aotFileName, std::make_shared<GNAPlugin>(config)) {
}
GNAExecutableNetwork(InferenceEngine::ICNNNetwork &network, const std::map<std::string, std::string> &config)
: GNAExecutableNetwork(network, std::make_shared<GNAPlugin>(config)) {
}
InferenceEngine::AsyncInferRequestInternal::Ptr
CreateAsyncInferRequestImpl(InferenceEngine::InputsDataMap networkInputs,
InferenceEngine::OutputsDataMap networkOutputs) override {
@ -58,5 +66,18 @@ class GNAExecutableNetwork : public InferenceEngine::ExecutableNetworkThreadSafe
void ExportImpl(std::ostream&) override {
THROW_IE_EXCEPTION << NOT_IMPLEMENTED_str;
}
void GetConfig(const std::string &name,
InferenceEngine::Parameter &result,
InferenceEngine::ResponseDesc* /*resp*/) const override {
result = plg->GetConfig(name, {});
}
void GetMetric(const std::string& name,
InferenceEngine::Parameter& result,
InferenceEngine::ResponseDesc* /* resp */) const override {
result = plg->GetMetric(name, {});
}
};
} // namespace GNAPluginNS

109
inference-engine/src/gna_plugin/gna_graph_compiler.cpp
View File

@ -34,6 +34,7 @@
#include "layers/gna_concat_layer.hpp"
#include "layers/gna_crop_layer.hpp"
#include "round_float_define.hpp"
#include "gna_plugin_policy.hpp"
using namespace InferenceEngine;
using namespace std;
@ -58,6 +59,10 @@ void GNAGraphCompiler::setGNAFlagsPtr(std::shared_ptr<GNAPluginNS::GNAFlags> gna
this->gnaFlags = std::move(gnaFlagsPtr);
}
void GNAGraphCompiler::setPolicy(GNAPluginNS::Policy policyToSet) {
this->policy = policyToSet;
}
intel_dnn_component_t * GNAGraphCompiler::find_first_unused_input(InferenceEngine::CNNLayerPtr current) {
if (current->insData.empty())
return nullptr;
@ -987,13 +992,57 @@ void GNAGraphCompiler::ConcatAlignFilterPrimitive(InferenceEngine::CNNLayerPtr l
auto outputs = *layer->outData.begin();
auto inputs = layer->insData.begin()->lock();
// auto offset = filterLayer->GetParamAsInt("output_offset");
uint32_t num_columns_in = FROM_IR_DIM(inputs, 2);
uint32_t num_rows_out = FROM_IR_DIM(outputs, 1);
uint32_t num_rows_in = filterLayer->_weights->size() / num_rows_out;
uint32_t num_padding = ALIGN(num_rows_in, 8) - num_rows_in;
auto numRowsPadded = filterLayer->GetParamAsInt("num_rows_padded");
// number of rows we handled by inserting copy layer
uint32_t num_rows_copied = 0;
// in case of left alignment succeed, but due to number of elements not multiple of 8 we need to insert align_filter
// we are improving it by inserting copy layer of size that covers most of elements - remained max of 32x31 affine filter
if (policy.ConcatAlignmentPolicy == Policy::ConcatAlignment::FAST && 0 == numRowsPadded && ALIGN(num_rows_in, 32) > 32) {
// can we use copy at all
num_rows_copied = ALIGN(num_rows_in, 32) - 32;
auto orientation = kDnnInterleavedOrientation;
auto& copyComponent = dnnComponents.addComponent(layer->name + "_synthetic_copy", "copy");
dnn->InitCopyComponent(copyComponent,
orientation,
num_rows_copied,
num_columns_in,
num_rows_copied,
num_columns_in,
inputs->getPrecision().size(),
inputs->getPrecision().size(),
quantized == nullptr ? 1 : quantized->_dst_quant.scale,
num_rows_copied,
num_columns_in,
ptr_inputs,
ptr_outputs);
size_t num_data_bytes_in = num_rows_copied * num_rows_copied * num_columns_in
* inputs->getPrecision().size();
// need to reserve full tensor so using original size with assumption of identity activation attached to filter lateron
size_t num_data_bytes_out = num_rows_out * num_columns_in * inputs->getPrecision().size();
connectInput(layer, ptr_inputs, num_data_bytes_in);
auto isNonFunctional = [](CNNLayerPtr l) {
return LayerInfo(l).isNonFunctional();
};
auto identity = CNNNetGetNextLayerSkipCertain(layer, 0, 0, isNonFunctional);
connectOutput(identity.first, ptr_outputs, num_data_bytes_out);
num_rows_in -= num_rows_copied;
num_rows_out -= num_rows_copied;
}
filterLayer->params["rows_copied_offset"] = std::to_string(num_rows_copied * inputs->getPrecision().size());
auto biasPrecision = filterLayer->_biases ? filterLayer->_biases->getTensorDesc().getPrecision() : outputs->getPrecision();
auto& currentComponent = dnnComponents.addComponent(layer->name, "affine");
@ -1013,35 +1062,36 @@ void GNAGraphCompiler::ConcatAlignFilterPrimitive(InferenceEngine::CNNLayerPtr l
ptr_biases,
false);
size_t num_data_bytes_out =
InferenceEngine::details::product(
begin(outputs->getDims()), end(outputs->getDims())) * 4;
size_t num_data_bytes_out = num_rows_out * num_columns_in * outputs->getPrecision().size();
size_t num_data_bytes_in = num_columns_in *
ALIGN(num_rows_in, 8) * inputs->getPrecision().size();
connectInput(layer, ptr_inputs, num_data_bytes_in, 0, 0);
connectInput(layer, ptr_inputs, num_data_bytes_in, num_rows_copied * inputs->getPrecision().size(), 0);
connectOutput(layer, ptr_outputs, num_data_bytes_out);
if (num_padding == 0) {
gnamem->readonly().push_ptr(ptr_weights,
filterLayer->_weights->cbuffer().as<const void*>(),
filterLayer->_weights->byteSize(),
64);
} else {
{
auto weightsElementSize = filterLayer->_weights->getTensorDesc().getPrecision().size();
auto elementsIn = (num_rows_in + num_padding) * num_columns_in;
auto paddedWeights = elementsIn * num_rows_out;
auto paddedWeightsSize = paddedWeights * filterLayer->precision.size();
auto paddedWeightsSize = paddedWeights * weightsElementSize;
// TODO: this can be improved to not generate unneeded weights at all
size_t weights_stride = (num_rows_in + num_rows_copied) * weightsElementSize;
size_t weights_offset = weights_stride * num_rows_copied + num_rows_copied * weightsElementSize;
gnamem->readonly().push_initializer(ptr_weights, paddedWeightsSize, [=](void* data, size_t size) {
size_t offset = 0;
for (int i = 0; i < num_rows_out && size >= offset; i++) {
ie_memcpy(reinterpret_cast<uint8_t*>(data) + offset, size - offset,
filterLayer->_weights->cbuffer().as<const uint8_t*>() + num_rows_in * i * filterLayer->precision.size(),
num_rows_in* filterLayer->precision.size());
offset += (num_rows_in + num_padding) * filterLayer->precision.size();
size_t roffset = weights_offset;
size_t woffset = 0;
for (int i = 0; i < num_rows_out && size >= woffset; i++) {
ie_memcpy(reinterpret_cast<uint8_t*>(data) + woffset,
size - woffset,
filterLayer->_weights->cbuffer().as<const uint8_t*>() + roffset,
num_rows_in * weightsElementSize);
roffset += weights_stride;
woffset += elementsIn * weightsElementSize;
}
}, 64);
}, 64);
}
if (filterLayer->_biases) {
@ -1189,11 +1239,18 @@ void GNAGraphCompiler::PWLPrimitive(InferenceEngine::CNNLayerPtr layer) {
num_rows = FROM_IR_DIM(inputs, 1);
}
size_t num_data_bytes_out = InferenceEngine::details::product(begin(outputs->getDims()), end(outputs->getDims()))
* outputs->getPrecision().size();
// TODO: solve this by layer level transformations
auto concatAlignFilter = CNNNetPrevLayer(layer, 0);
if (LayerInfo(concatAlignFilter).isConcatAlignFilter()) {
auto rowsCopiedOffset = concatAlignFilter->GetParamAsInt("rows_copied_offset");
if (rowsCopiedOffset != 0) {
num_rows -= rowsCopiedOffset / outputs->getPrecision().size();
layer->params["output_offset"] = std::to_string(rowsCopiedOffset);
}
}
size_t num_data_bytes_in = InferenceEngine::details::product(begin(inputs->getDims()), end(inputs->getDims()))
* inputs->getPrecision().size();
size_t num_data_bytes_out = num_columns * num_rows * outputs->getPrecision().size();
size_t num_data_bytes_in = num_columns * num_rows * inputs->getPrecision().size();
static InferenceEngine::details::caseless_unordered_map<std::string, DnnActivationType> supportedActivations = {
{"sigmoid", kActSigmoid},
@ -1626,7 +1683,7 @@ GNAPluginNS::ConnectionDetails GNAGraphCompiler::connectInput(CNNLayerPtr layer,
if (it != splitLayerInfoItem.splitOutputLayers.end()) {
gnalog() << "Connecting " << splitName << " input \n";
auto res = connectInput(splittingLayer, ptr, splitLayerInfoItem.reserved_size, it->offset, 0);
auto res = connectInput(splittingLayer, ptr, splitLayerInfoItem.reserved_size, it->offset + offset, 0);
gnalog() << "Connected \n";
return res;
}

3
inference-engine/src/gna_plugin/gna_graph_compiler.hpp
View File

@ -26,6 +26,7 @@
#include "backend/am_intel_dnn.hpp"
#include "gna_device.hpp"
#include "gna_data_types.hpp"
#include "gna_plugin_policy.hpp"
namespace GNAPluginNS {
class GNAGraphCompiler {
@ -34,6 +35,7 @@ private:
std::shared_ptr<GNAPluginNS::gna_memory_type> gnamem;
std::shared_ptr<GNAPluginNS::InputDesc> inputDesc;
std::shared_ptr<GNAPluginNS::GNAFlags> gnaFlags;
Policy policy;
// layers with extra storage for connections and additional
// non trivial processing
@ -53,6 +55,7 @@ public:
void setDNNPtr(std::shared_ptr<GNAPluginNS::backend::AMIntelDNN> dnnPtr);
void setInputDescPtr(std::shared_ptr<GNAPluginNS::InputDesc> inputDescPtr);
void setGNAFlagsPtr(std::shared_ptr<GNAPluginNS::GNAFlags> gnaFlagsPtr);
void setPolicy(GNAPluginNS::Policy policy);
void fillMemoryConnections(std::unordered_map<std::string,
std::vector<InferenceEngine::CNNLayerPtr>> &memoryPairs);

1
inference-engine/src/gna_plugin/gna_model_serial.cpp
View File

@ -237,7 +237,6 @@ void GNAModelSerial::Export(void * basePointer, size_t gnaGraphSize, std::ostrea
auto convert_to_serial = [getOffsetFromBase](const GNAModelSerial::RuntimeEndPoint& ep) {
ModelHeader::EndPoint out;
out.elements_count = ep.elements_count;
out.element_size = ep.element_size;
out.descriptor_offset = offsetFromBase(ep.descriptor_ptr);
out.scaleFactor = ep.scaleFactor;
return out;

278
inference-engine/src/gna_plugin/gna_plugin.cpp
View File

@ -22,6 +22,7 @@
#include <graph_tools.hpp>
#include <debug.h>
#include <gna/gna_config.hpp>
#include "gna_plugin_config.hpp"
#include <ie_util_internal.hpp>
#include "gna_plugin.hpp"
#include "optimizer/gna_pass_manager.hpp"
@ -302,6 +303,7 @@ void GNAPlugin::ImportFrames(
GNAPlugin::GNAPlugin() {
Init();
UpdateFieldsFromConfig();
}
GNAPlugin::GNAPlugin(const std::map<std::string, std::string>& configMap) {
@ -321,13 +323,13 @@ void GNAPlugin::Init() {
void GNAPlugin::InitGNADevice() {
#if GNA_LIB_VER == 1
gnadevice = std::make_shared<GNADeviceHelper>(gna_proc_type,
gnadevice = std::make_shared<GNADeviceHelper>(config.gna_proc_type,
gnaFlags->gna_lib_async_threads_num,
gnaFlags->gna_openmp_multithreading,
gnaFlags->performance_counting);
#else
gnadevice = std::make_shared<GNADeviceHelper>(pluginGna2AccMode,
pluginGna2DeviceConsistent,
gnadevice = std::make_shared<GNADeviceHelper>(config.pluginGna2AccMode,
config.pluginGna2DeviceConsistent,
gnaFlags->gna_lib_async_threads_num,
gnaFlags->gna_openmp_multithreading,
gnaFlags->performance_counting);
@ -387,7 +389,7 @@ void GNAPlugin::LoadNetwork(ICNNNetwork &network) {
run_passes(newNet, true);
run_passes(newNet, false);
} else {
switch (gnaPrecision) {
switch (config.gnaPrecision) {
case Precision::I16:
ModelQuantizer<QuantI16> q16;
newNet = q16.quantize(network, run_passes, inputsDesc->inputScaleFactors);
@ -421,6 +423,9 @@ void GNAPlugin::LoadNetwork(ICNNNetwork &network) {
auto sortedNet = CNNNetSortTopologicallyEx(*newNet, make_fuzed_order);
// passing policy to compiler
graphCompiler.setPolicy(policy);
if (sortedNet.empty()) {
THROW_GNA_EXCEPTION << "Sorted network is empty";
}
@ -534,10 +539,33 @@ void GNAPlugin::LoadNetwork(ICNNNetwork &network) {
gnalog() << "[UFS] from : "<< outPort.first <<" reached: " << layer->name << "\n";
// probing gna_primitives
if (irLayerAvatar != graphCompiler.dnnComponents.components.end()) {
initOutput(portId, irLayerAvatar->second, layer);
stopSearching = true;
}
// probing concatInfo
if (!stopSearching && LayerInfo(layer).isConcat()) {
auto concatConnection = graphCompiler.concat_connection.find(layer->name);
if (concatConnection != graphCompiler.concat_connection.end()) {
//initOutput(portId, irLayerAvatar->second, layer);
auto &desc = outputsDesc[portId];
auto quantized = InferenceEngine::getInjectedData<QuantizedLayerParams>(layer);
desc.ptrs.resize(gnaFlags->gna_lib_async_threads_num);
// TODO: what is orientation for concat
desc.orientation = kDnnInterleavedOrientation;
desc.num_bytes_per_element = layer->outData.front()->getPrecision().size();
desc.scale_factor = quantized != nullptr ? quantized->_dst_quant.scale : 1.0f;
desc.num_elements = concatConnection->second.reserved_size / desc.num_bytes_per_element;
// binding ptr for first infer request - then others will be setup during relocation
gnamem->bind_ptr(&desc.ptrs.front(), &concatConnection->second.gna_ptr);
stopSearching = true;
}
}
}, true, [&stopSearching](InferenceEngine::CNNLayer* from) {
return make_upstream_order(!stopSearching ? from : nullptr);
});
@ -722,20 +750,20 @@ void GNAPlugin::createRequestConfigsForGnaModels() {
void GNAPlugin::DumpXNNToFile() const {
// TODO: output precision as well as pointer might be incorrect, LSTM for sure
// gna looks automatically set layer 0 as output and adjust it's pointer / precision/ size respectively
if (dumpXNNPath.empty()) {
if (config.dumpXNNPath.empty()) {
return;
}
if (dumpXNNGeneration != "GNA1" &&
dumpXNNGeneration != "GNA3" &&
!dumpXNNGeneration.empty()) {
THROW_GNA_EXCEPTION << "Wrong GNA generation for embedded model dump: " << dumpXNNGeneration;
if (config.dumpXNNGeneration != "GNA1" &&
config.dumpXNNGeneration != "GNA3" &&
!config.dumpXNNGeneration.empty()) {
THROW_GNA_EXCEPTION << "Wrong GNA generation for embedded model dump: " << config.dumpXNNGeneration;
}
if (!gnadevice) {
THROW_GNA_EXCEPTION << "Cannot generate XNNDump for float network";
}
std::ofstream dumpStream(dumpXNNPath, std::ios::out | std::ios::binary);
std::ofstream dumpStream(config.dumpXNNPath, std::ios::out | std::ios::binary);
#if GNA_LIB_VER == 1
auto dump = gnadevice->dumpXnn(&std::get<0>(nnets.front())->obj, ptr_active_indices, num_active_indices);
dump.header.rw_region_size = gnamem->getRWBytes();
@ -745,7 +773,7 @@ void GNAPlugin::DumpXNNToFile() const {
dumpStream.write(reinterpret_cast<char*>(dump.model.get()), dump.header.model_size);
#else
auto const modelId = gnadevice->createModel(std::get<0>(gnaModels.front())->obj);
if (dumpXNNGeneration != "GNA3") {
if (config.dumpXNNGeneration != "GNA3") {
auto dump = gnadevice->dumpXnn(modelId);
dump.header.RwRegionSize = gnamem->getRWBytes();
dump.header.InputScalingFactor = inputsDesc->inputScaleFactors.front();
@ -1204,228 +1232,14 @@ void GNAPlugin::GetPerformanceCounts(std::map<std::string, InferenceEngine::Infe
void GNAPlugin::AddExtension(InferenceEngine::IExtensionPtr extension) {}
void GNAPlugin::SetConfig(const std::map<std::string, std::string> &config) {
Init();
auto supportedConfigOptions = supportedConfigKeys();
void GNAPlugin::SetConfig(const std::map<std::string, std::string> &config_map) {
config.UpdateFromMap(config_map);
UpdateFieldsFromConfig();
}
for (auto& item : config) {
auto keys = std::find_if(supportedConfigOptions.begin(), supportedConfigOptions.end(), [&item](const std::string& supportedConfigOption) {
return item.first == supportedConfigOption ||
item.first.find(GNA_CONFIG_KEY(SCALE_FACTOR)) == 0;
});
if (keys == supportedConfigOptions.end()) {
THROW_GNA_EXCEPTION << as_status << NOT_FOUND << "Incorrect GNA Plugin config. Key " << item.first << " not supported";
}
}
// holds actual value of a found key
std::string key;
std::string value;
auto if_set = [&](const std::string& keyInput, const std::function<void()> & handler) {
auto keyInMap = config.find(keyInput);
if (keyInMap != config.end()) {
value = keyInMap->second;
handler();
}
};
auto if_start = [&](const std::string& keyInput, const std::function<void()> & handler) {
for (auto && c : config) {
if (c.first.find(keyInput) == 0) {
if (c.first.size() > keyInput.size() + 1) {
key = c.first.substr(keyInput.size() + 1);
value = c.second;
handler();
}
}
}
};
auto fp32eq = [](float p1, float p2) -> bool {
return (std::abs(p1 - p2) <= 0.00001f * std::min(std::abs(p1), std::abs(p2)));
};
auto & log = gnalog();
if_start(GNA_CONFIG_KEY(SCALE_FACTOR), [&, this] {
uint64_t scaleForInput = std::stoul(key, NULL, 10);
if (scaleForInput > 10) {
THROW_GNA_EXCEPTION << "input scale factor with index(" << key << ") unsupported";
}
auto scaleFactor = std::stod(value);
if (fp32eq(scaleFactor, 0.0f)) {
THROW_GNA_EXCEPTION << "input scale factor of 0.0f not supported";
}
// not appeared scale factors are to be 1.0f
if (inputsDesc->inputScaleFactors.size() <= scaleForInput) {
inputsDesc->inputScaleFactors.resize(scaleForInput + 1, 1.f);
}
inputsDesc->inputScaleFactors[scaleForInput] = InferenceEngine::CNNLayer::ie_parse_float(value);
});
if (inputsDesc->inputScaleFactors.empty()) {
if_set(GNA_CONFIG_KEY(SCALE_FACTOR), [&] {
auto scaleFactor = InferenceEngine::CNNLayer::ie_parse_float(value);
if (fp32eq(scaleFactor, 0.0f)) {
THROW_GNA_EXCEPTION << "input scale factor of 0.0f not supported";
}
inputsDesc->inputScaleFactors.push_back(scaleFactor);
});
}
if (inputsDesc->inputScaleFactors.empty()) {
inputsDesc->inputScaleFactors.push_back(1.f);
}
if_set(GNA_CONFIG_KEY(FIRMWARE_MODEL_IMAGE), [&] {
dumpXNNPath = value;
});
if_set(GNA_CONFIG_KEY(FIRMWARE_MODEL_IMAGE_GENERATION), [&] {
dumpXNNGeneration = value;
});
if_set(GNA_CONFIG_KEY(DEVICE_MODE), [&] {
#if GNA_LIB_VER == 1
static caseless_unordered_map <std::string, uint32_t> supported_values = {
{GNAConfigParams::GNA_AUTO, GNA_AUTO},
{GNAConfigParams::GNA_HW, GNA_HARDWARE},
{GNAConfigParams::GNA_SW, GNA_SOFTWARE},
{GNAConfigParams::GNA_SW_EXACT, GNA_SOFTWARE & GNA_HARDWARE}
};
static std::vector <std::string> supported_values_on_gna2 = {
GNAConfigParams::GNA_GEN,
GNAConfigParams::GNA_GEN_EXACT,
GNAConfigParams::GNA_SSE,
GNAConfigParams::GNA_SSE_EXACT,
GNAConfigParams::GNA_AVX1,
GNAConfigParams::GNA_AVX1_EXACT,
GNAConfigParams::GNA_AVX2,
GNAConfigParams::GNA_AVX2_EXACT
};
#else
static caseless_unordered_map <std::string, std::pair<Gna2AccelerationMode, Gna2DeviceVersion> > supported_values = {
{GNAConfigParams::GNA_AUTO, {Gna2AccelerationModeAuto, Gna2DeviceVersionSoftwareEmulation}},
{GNAConfigParams::GNA_HW, {Gna2AccelerationModeHardware, Gna2DeviceVersionSoftwareEmulation}},
{GNAConfigParams::GNA_SW, {Gna2AccelerationModeSoftware, Gna2DeviceVersionSoftwareEmulation}},
{GNAConfigParams::GNA_SW_EXACT, {Gna2AccelerationModeSoftware, Gna2DeviceVersion1_0}},
{GNAConfigParams::GNA_GEN, {Gna2AccelerationModeGeneric, Gna2DeviceVersionSoftwareEmulation}},
{GNAConfigParams::GNA_GEN_EXACT, {Gna2AccelerationModeGeneric, Gna2DeviceVersion1_0}},
{GNAConfigParams::GNA_SSE, {Gna2AccelerationModeSse4x2, Gna2DeviceVersionSoftwareEmulation}},
{GNAConfigParams::GNA_SSE_EXACT, {Gna2AccelerationModeSse4x2, Gna2DeviceVersion1_0}},
{GNAConfigParams::GNA_AVX1, {Gna2AccelerationModeAvx1, Gna2DeviceVersionSoftwareEmulation}},
{GNAConfigParams::GNA_AVX1_EXACT, {Gna2AccelerationModeAvx1, Gna2DeviceVersion1_0}},
{GNAConfigParams::GNA_AVX2, {Gna2AccelerationModeAvx2, Gna2DeviceVersionSoftwareEmulation}},
{GNAConfigParams::GNA_AVX2_EXACT, {Gna2AccelerationModeAvx2, Gna2DeviceVersion1_0}},
};
#endif
auto procType = supported_values.find(value);
if (procType == supported_values.end()) {
if (value == GNA_CONFIG_VALUE(SW_FP32)) {
gnaFlags->sw_fp32 = true;
} else {
#if GNA_LIB_VER == 1
auto is_gna2_mode = std::find(
supported_values_on_gna2.begin(),
supported_values_on_gna2.end(),
value);
if (is_gna2_mode != supported_values_on_gna2.end()) {
THROW_GNA_EXCEPTION << "This GNA device mode require GNA2 library: " << value;
}
#endif
THROW_GNA_EXCEPTION << "GNA device mode unsupported: " << value;
}
} else {
#if GNA_LIB_VER == 1
gna_proc_type = static_cast<intel_gna_proc_t>(procType->second);
#else
pluginGna2AccMode = procType->second.first;
pluginGna2DeviceConsistent = procType->second.second;
#endif
}
});
if_set(GNA_CONFIG_KEY(COMPACT_MODE), [&] {
if (value == PluginConfigParams::YES) {
gnaFlags->compact_mode = true;
} else if (value == PluginConfigParams::NO) {
gnaFlags->compact_mode = false;
} else {
log << "GNA compact mode should be YES/NO, but not" << value;
THROW_GNA_EXCEPTION << "GNA compact mode should be YES/NO, but not" << value;
}
});
if_set(CONFIG_KEY(EXCLUSIVE_ASYNC_REQUESTS), [&] {
if (value == PluginConfigParams::YES) {
gnaFlags->exclusive_async_requests = true;
} else if (value == PluginConfigParams::NO) {
gnaFlags->exclusive_async_requests = false;
} else {
log << "EXCLUSIVE_ASYNC_REQUESTS should be YES/NO, but not" << value;
THROW_GNA_EXCEPTION << "EXCLUSIVE_ASYNC_REQUESTS should be YES/NO, but not" << value;
}
});
if_set(GNA_CONFIG_KEY(PRECISION), [&] {
auto precision = Precision::FromStr(value);
if (precision != Precision::I8 && precision != Precision::I16) {
log << "Unsupported precision of GNA hardware, should be Int16 or Int8, but was: " << value;
THROW_GNA_EXCEPTION << "Unsupported precision of GNA hardware, should be Int16 or Int8, but was: " << value;
}
gnaPrecision = precision;
});
if_set(GNA_CONFIG_KEY(PWL_UNIFORM_DESIGN), [&] {
if (value == PluginConfigParams::YES) {
gnaFlags->uniformPwlDesign = true;
} else if (value == PluginConfigParams::NO) {
gnaFlags->uniformPwlDesign = false;
} else {
log << "GNA pwl uniform algorithm parameter "
<< "should be equal to YES/NO, but not" << value;
THROW_GNA_EXCEPTION << "GNA pwl uniform algorithm parameter "
<< "should be equal to YES/NO, but not" << value;
}
});
if_set(CONFIG_KEY(PERF_COUNT), [&] {
if (value == PluginConfigParams::YES) {
gnaFlags->performance_counting = true;
} else if (value == PluginConfigParams::NO) {
gnaFlags->performance_counting = false;
} else {
log << "GNA performance counter enabling parameter "
<< "should be equal to YES/NO, but not" << value;
THROW_GNA_EXCEPTION << "GNA performance counter enabling parameter "
<< "should be equal to YES/NO, but not" << value;
}
});
if_set(GNA_CONFIG_KEY(LIB_N_THREADS), [&] {
uint64_t lib_threads = std::stoul(value, NULL, 10);
if (lib_threads == 0 || lib_threads > std::numeric_limits<uint8_t>::max()/2-1) {
log << "Unsupported accelerator lib number of threads: " << value << ", should be greateer than 0 and less than 127";
THROW_GNA_EXCEPTION << "Unsupported accelerator lib number of threads: " << value
<< ", should be greateer than 0 and less than 127";
}
gnaFlags->gna_lib_async_threads_num = lib_threads;
});
if_set(CONFIG_KEY(SINGLE_THREAD), [&] {
if (value == PluginConfigParams::YES) {
gnaFlags->gna_openmp_multithreading = false;
} else if (value == PluginConfigParams::NO) {
gnaFlags->gna_openmp_multithreading = true;
} else {
log << "EXCLUSIVE_ASYNC_REQUESTS should be YES/NO, but not" << value;
THROW_GNA_EXCEPTION << "EXCLUSIVE_ASYNC_REQUESTS should be YES/NO, but not" << value;
}
});
if (gnaFlags->sw_fp32 && gnaFlags->gna_lib_async_threads_num > 1) {
THROW_GNA_EXCEPTION << "GNA plugin not support async mode on GNA_SW_FP32!";
}
void GNAPlugin::UpdateFieldsFromConfig() {
inputsDesc->inputScaleFactors = config.inputScaleFactors;
*gnaFlags = config.gnaFlags;
}
void GNAPlugin::QueryNetwork(const InferenceEngine::ICNNNetwork& network,

26
inference-engine/src/gna_plugin/gna_plugin.hpp
View File

@ -22,6 +22,7 @@
#include "gna_graph_compiler.hpp"
#include "gna_plugin_policy.hpp"
#include "gna_plugin_log.hpp"
#include "gna_plugin_config.hpp"
#if GNA_LIB_VER == 2
#include <gna2-model-api.h>
@ -32,6 +33,7 @@ class GNAPlugin : public InferenceEngine::IInferencePluginInternal, public std::
protected:
std::string _pluginName = "GNA";
Config config;
std::shared_ptr<GNAPluginNS::backend::AMIntelDNN> dnn;
std::shared_ptr<GNAPluginNS::GNAFlags> gnaFlags;
std::shared_ptr<GNAPluginNS::gna_memory_type> gnamem;
@ -63,20 +65,12 @@ class GNAPlugin : public InferenceEngine::IInferencePluginInternal, public std::
// index matches iterating order of cnnnetwork outputs info
std::vector<GNAPluginNS::OutputDesc> outputsDesc = std::vector<OutputDesc>();
// precision of GNA hardware model
InferenceEngine::Precision gnaPrecision = InferenceEngine::Precision::I16;
intel_dnn_number_type_t output_type = kDnnInt;
GNAPluginNS::Policy policy;
std::string dumpXNNPath;
std::string dumpXNNGeneration;
#if GNA_LIB_VER == 1
intel_gna_proc_t gna_proc_type = static_cast<intel_gna_proc_t>(GNA_SOFTWARE & GNA_HARDWARE);
#else
Gna2AccelerationMode pluginGna2AccMode = Gna2AccelerationModeSoftware;
Gna2DeviceVersion pluginGna2DeviceConsistent = Gna2DeviceVersion1_0;
void createRequestConfigsForGnaModels();
#if GNA_LIB_VER == 2
void createRequestConfigsForGnaModels();
#endif
std::shared_ptr<GNADeviceHelper> gnadevice;
@ -104,15 +98,12 @@ void createRequestConfigsForGnaModels();
void GetPerformanceCounts(std::map<std::string, InferenceEngine::InferenceEngineProfileInfo> &perfMap);
void AddExtension(InferenceEngine::IExtensionPtr extension) override;
std::vector<std::string> supportedConfigKeys() const;
std::map<std::string, std::string> supportedConfigKeysWithDefaults() const;
void SetConfig(const std::map<std::string, std::string> &config) override;
void LoadNetwork(InferenceEngine::IExecutableNetwork::Ptr &executableNetwork,
const InferenceEngine::ICNNNetwork &network,
const std::map<std::string, std::string> &config) override { THROW_GNA_EXCEPTION << "Not implemented"; }
const std::map<std::string, std::string> &config_map) override { THROW_GNA_EXCEPTION << "Not implemented"; }
InferenceEngine::ExecutableNetwork LoadNetwork(const InferenceEngine::ICNNNetwork &network,
const std::map<std::string, std::string> &config,
const std::map<std::string, std::string> &config_map,
InferenceEngine::RemoteContext::Ptr context) override { THROW_GNA_EXCEPTION << "Not implemented"; }
void Infer(const InferenceEngine::Blob &input, InferenceEngine::Blob &result);
void SetCore(InferenceEngine::ICore*) noexcept override {}
@ -221,5 +212,8 @@ void createRequestConfigsForGnaModels();
const GNASplitLayer& splitInfo,
size_t precision_size,
int idx = 0);
void UpdateFieldsFromConfig();
};
} // namespace GNAPluginNS

278
inference-engine/src/gna_plugin/gna_plugin_config.cpp Normal file
View File

@ -0,0 +1,278 @@
// Copyright (C) 2020 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
#include <gna/gna_config.hpp>
#include "gna_plugin.hpp"
#include "gna_plugin_config.hpp"
#include "ie_common.h"
#include <details/caseless.hpp>
#include <unordered_map>
using namespace InferenceEngine;
using namespace InferenceEngine::details;
namespace GNAPluginNS {
#if GNA_LIB_VER == 1
static caseless_unordered_map<std::string, uint32_t> supported_values = {
{GNAConfigParams::GNA_AUTO, GNA_AUTO},
{GNAConfigParams::GNA_HW, GNA_HARDWARE},
{GNAConfigParams::GNA_SW, GNA_SOFTWARE},
{GNAConfigParams::GNA_SW_EXACT, GNA_SOFTWARE & GNA_HARDWARE}
};
static std::vector<std::string> supported_values_on_gna2 = {
GNAConfigParams::GNA_GEN,
GNAConfigParams::GNA_GEN_EXACT,
GNAConfigParams::GNA_SSE,
GNAConfigParams::GNA_SSE_EXACT,
GNAConfigParams::GNA_AVX1,
GNAConfigParams::GNA_AVX1_EXACT,
GNAConfigParams::GNA_AVX2,
GNAConfigParams::GNA_AVX2_EXACT
};
#else
static caseless_unordered_map <std::string, std::pair<Gna2AccelerationMode, Gna2DeviceVersion>> supported_values = {
{GNAConfigParams::GNA_AUTO, {Gna2AccelerationModeAuto, Gna2DeviceVersionSoftwareEmulation}},
{GNAConfigParams::GNA_HW, {Gna2AccelerationModeHardware, Gna2DeviceVersionSoftwareEmulation}},
{GNAConfigParams::GNA_SW, {Gna2AccelerationModeSoftware, Gna2DeviceVersionSoftwareEmulation}},
{GNAConfigParams::GNA_SW_EXACT, {Gna2AccelerationModeSoftware, Gna2DeviceVersion1_0}},
{GNAConfigParams::GNA_GEN, {Gna2AccelerationModeGeneric, Gna2DeviceVersionSoftwareEmulation}},
{GNAConfigParams::GNA_GEN_EXACT, {Gna2AccelerationModeGeneric, Gna2DeviceVersion1_0}},
{GNAConfigParams::GNA_SSE, {Gna2AccelerationModeSse4x2, Gna2DeviceVersionSoftwareEmulation}},
{GNAConfigParams::GNA_SSE_EXACT, {Gna2AccelerationModeSse4x2, Gna2DeviceVersion1_0}},
{GNAConfigParams::GNA_AVX1, {Gna2AccelerationModeAvx1, Gna2DeviceVersionSoftwareEmulation}},
{GNAConfigParams::GNA_AVX1_EXACT, {Gna2AccelerationModeAvx1, Gna2DeviceVersion1_0}},
{GNAConfigParams::GNA_AVX2, {Gna2AccelerationModeAvx2, Gna2DeviceVersionSoftwareEmulation}},
{GNAConfigParams::GNA_AVX2_EXACT, {Gna2AccelerationModeAvx2, Gna2DeviceVersion1_0}},
};
#endif
void Config::UpdateFromMap(const std::map<std::string, std::string>& config) {
for (auto&& item : config) {
auto key = item.first;
auto value = item.second;
auto fp32eq = [](float p1, float p2) -> bool {
return (std::abs(p1 - p2) <= 0.00001f * std::min(std::abs(p1), std::abs(p2)));
};
auto &log = gnalog();
if (key.find(GNA_CONFIG_KEY(SCALE_FACTOR)) == 0) {
uint64_t input_index;
if (key == GNA_CONFIG_KEY(SCALE_FACTOR)) {
input_index = 0;
} else {
key.erase(0, strlen(GNA_CONFIG_KEY(SCALE_FACTOR)));
if (key[0] != '_') {
THROW_GNA_EXCEPTION << "Invalid format of scale factor configuration key";
}
key.erase(0, 1);
try {
input_index = std::stoi(key);
if (input_index < 0 | input_index > 9) {
throw std::out_of_range("");
}
} catch (std::invalid_argument&) {
THROW_GNA_EXCEPTION << "Invalid value of index of input scale factor";
} catch (std::out_of_range&) {
THROW_GNA_EXCEPTION << "Index of input scale factor must be in the range [0..9], " << value << " provided";
}
}
auto scale_factor = InferenceEngine::CNNLayer::ie_parse_float(value);
if (fp32eq(scale_factor, 0.0f)) {
THROW_GNA_EXCEPTION << "input scale factor of 0.0f not supported";
}
// missing scale factors are set to be 1.0f
if (inputScaleFactors.size() <= input_index) {
inputScaleFactors.resize(input_index + 1, 1.f);
}
inputScaleFactors[input_index] = InferenceEngine::CNNLayer::ie_parse_float(value);
} else if (key == GNA_CONFIG_KEY(FIRMWARE_MODEL_IMAGE)) {
dumpXNNPath = value;
} else if (key == GNA_CONFIG_KEY(FIRMWARE_MODEL_IMAGE_GENERATION)) {
dumpXNNGeneration = value;
} else if (key == GNA_CONFIG_KEY(DEVICE_MODE)) {
auto procType = supported_values.find(value);
if (procType == supported_values.end()) {
if (value == GNA_CONFIG_VALUE(SW_FP32)) {
gnaFlags.sw_fp32 = true;
} else {
#if GNA_LIB_VER == 1
auto is_gna2_mode = std::find(
supported_values_on_gna2.begin(),
supported_values_on_gna2.end(),
value);
if (is_gna2_mode != supported_values_on_gna2.end()) {
THROW_GNA_EXCEPTION << "This GNA device mode requires GNA2 library: " << value;
}
#endif
THROW_GNA_EXCEPTION << "GNA device mode unsupported: " << value;
}
} else {
#if GNA_LIB_VER == 1
gna_proc_type = static_cast<intel_gna_proc_t>(procType->second);
#else
pluginGna2AccMode = procType->second.first;
pluginGna2DeviceConsistent = procType->second.second;
#endif
}
} else if (key == GNA_CONFIG_KEY(COMPACT_MODE)) {
if (value == PluginConfigParams::YES) {
gnaFlags.compact_mode = true;
} else if (value == PluginConfigParams::NO) {
gnaFlags.compact_mode = false;
} else {
log << "GNA compact mode should be YES/NO, but not " << value;
THROW_GNA_EXCEPTION << "GNA compact mode should be YES/NO, but not " << value;
}
} else if (key == CONFIG_KEY(EXCLUSIVE_ASYNC_REQUESTS)) {
if (value == PluginConfigParams::YES) {
gnaFlags.exclusive_async_requests = true;
} else if (value == PluginConfigParams::NO) {
gnaFlags.exclusive_async_requests = false;
} else {
log << "EXCLUSIVE_ASYNC_REQUESTS should be YES/NO, but not" << value;
THROW_GNA_EXCEPTION << "EXCLUSIVE_ASYNC_REQUESTS should be YES/NO, but not" << value;
}
} else if (key == GNA_CONFIG_KEY(PRECISION)) {
auto precision = Precision::FromStr(value);
if (precision != Precision::I8 && precision != Precision::I16) {
log << "Unsupported precision of GNA hardware, should be Int16 or Int8, but was: " << value;
THROW_GNA_EXCEPTION << "Unsupported precision of GNA hardware, should be Int16 or Int8, but was: "
<< value;
}
gnaPrecision = precision;
} else if (key == GNA_CONFIG_KEY(PWL_UNIFORM_DESIGN)) {
if (value == PluginConfigParams::YES) {
gnaFlags.uniformPwlDesign = true;
} else if (value == PluginConfigParams::NO) {
gnaFlags.uniformPwlDesign = false;
} else {
log << "GNA pwl uniform algorithm parameter "
<< "should be equal to YES/NO, but not" << value;
THROW_GNA_EXCEPTION << "GNA pwl uniform algorithm parameter "
<< "should be equal to YES/NO, but not" << value;
}
} else if (key == CONFIG_KEY(PERF_COUNT)) {
if (value == PluginConfigParams::YES) {
gnaFlags.performance_counting = true;
} else if (value == PluginConfigParams::NO) {
gnaFlags.performance_counting = false;
} else {
log << "GNA performance counter enabling parameter "
<< "should be equal to YES/NO, but not" << value;
THROW_GNA_EXCEPTION << "GNA performance counter enabling parameter "
<< "should be equal to YES/NO, but not" << value;
}
} else if (key == GNA_CONFIG_KEY(LIB_N_THREADS)) {
uint64_t lib_threads;
try {
lib_threads = std::stoul(value);
if (lib_threads == 0 || lib_threads > (std::numeric_limits<uint8_t>::max()+1) / 2 - 1) {
throw std::out_of_range("");
}
} catch (std::invalid_argument&) {
THROW_GNA_EXCEPTION << "Invalid value of number of threads";
} catch (std::out_of_range&) {
log << "Unsupported accelerator lib number of threads: " << value
<< ", should be greater than 0 and less than 127";
THROW_GNA_EXCEPTION << "Unsupported accelerator lib number of threads: " << value
<< ", should be greater than 0 and less than 127";
}
gnaFlags.gna_lib_async_threads_num = lib_threads;
} else if (key == CONFIG_KEY(SINGLE_THREAD)) {
if (value == PluginConfigParams::YES) {
gnaFlags.gna_openmp_multithreading = false;
} else if (value == PluginConfigParams::NO) {
gnaFlags.gna_openmp_multithreading = true;
} else {
log << "EXCLUSIVE_ASYNC_REQUESTS should be YES/NO, but not" << value;
THROW_GNA_EXCEPTION << "EXCLUSIVE_ASYNC_REQUESTS should be YES/NO, but not" << value;
}
} else {
THROW_GNA_EXCEPTION << as_status << NOT_FOUND << "Incorrect GNA Plugin config. Key " << item.first
<< " not supported";
}
if (gnaFlags.sw_fp32 && gnaFlags.gna_lib_async_threads_num > 1) {
THROW_GNA_EXCEPTION << "GNA plugin does not support async mode on GNA_SW_FP32!";
}
}
if (inputScaleFactors.empty()) {
inputScaleFactors.push_back(1.0f);
}
AdjustKeyMapValues();
}
void Config::AdjustKeyMapValues() {
key_config_map.clear();
if (inputScaleFactors.empty()) {
inputScaleFactors.push_back(1.0);
}
key_config_map[GNA_CONFIG_KEY(SCALE_FACTOR)] = std::to_string(inputScaleFactors[0]);
for (int n = 0; n < inputScaleFactors.size(); n++) {
key_config_map[GNA_CONFIG_KEY(SCALE_FACTOR) + std::string("_") + std::to_string(n)] =
std::to_string(inputScaleFactors[n]);
}
key_config_map[GNA_CONFIG_KEY(FIRMWARE_MODEL_IMAGE)] = dumpXNNPath;
key_config_map[GNA_CONFIG_KEY(FIRMWARE_MODEL_IMAGE_GENERATION)] = dumpXNNGeneration;
std::string device_mode;
if (gnaFlags.sw_fp32) {
device_mode = GNA_CONFIG_VALUE(SW_FP32);
} else {
for (auto&& value : supported_values) {
#if GNA_LIB_VER == 1
if (value.second == gna_proc_type) {
device_mode = value.first;
break;
}
#else
if (value.second.first == pluginGna2AccMode &&
value.second.second == pluginGna2DeviceConsistent) {
device_mode = value.first;
break;
}
#endif
}
}
IE_ASSERT(!device_mode.empty());
key_config_map[GNA_CONFIG_KEY(DEVICE_MODE)] = device_mode;
key_config_map[GNA_CONFIG_KEY(COMPACT_MODE)] =
gnaFlags.compact_mode ? PluginConfigParams::YES: PluginConfigParams::NO;
key_config_map[CONFIG_KEY(EXCLUSIVE_ASYNC_REQUESTS)] =
gnaFlags.exclusive_async_requests ? PluginConfigParams::YES: PluginConfigParams::NO;
key_config_map[GNA_CONFIG_KEY(PRECISION)] = gnaPrecision.name();
key_config_map[CONFIG_KEY(EXCLUSIVE_ASYNC_REQUESTS)] =
gnaFlags.exclusive_async_requests ? PluginConfigParams::YES: PluginConfigParams::NO;
key_config_map[GNA_CONFIG_KEY(PWL_UNIFORM_DESIGN)] =
gnaFlags.uniformPwlDesign ? PluginConfigParams::YES: PluginConfigParams::NO;
key_config_map[CONFIG_KEY(PERF_COUNT)] =
gnaFlags.performance_counting ? PluginConfigParams::YES: PluginConfigParams::NO;
key_config_map[GNA_CONFIG_KEY(LIB_N_THREADS)] = std::to_string(gnaFlags.gna_lib_async_threads_num);
key_config_map[CONFIG_KEY(SINGLE_THREAD)] =
gnaFlags.gna_openmp_multithreading ? PluginConfigParams::NO: PluginConfigParams::YES;
}
std::string Config::GetParameter(const std::string& name) const {
auto result = key_config_map.find(name);
if (result == key_config_map.end()) {
THROW_GNA_EXCEPTION << "Unsupported config key: " << name;
}
return result->second;
}
std::vector<std::string> Config::GetSupportedKeys() const {
std::vector<std::string> result;
for (auto&& configOption : key_config_map) {
result.push_back(configOption.first);
}
return result;
}
} // namespace GNAPluginNS

48
inference-engine/src/gna_plugin/gna_plugin_config.hpp Normal file
View File

@ -0,0 +1,48 @@
// Copyright (C) 2020 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
#pragma once
#if GNA_LIB_VER == 1
#include <gna-api.h>
#else
#include <gna2-inference-api.h>
#include <gna2-common-api.h>
#endif
#include "ie_precision.hpp"
#include "descriptions/gna_flags.hpp"
#include <vector>
#include <map>
namespace GNAPluginNS {
struct Config {
Config() {
AdjustKeyMapValues();
}
void UpdateFromMap(const std::map<std::string, std::string>& configMap);
void AdjustKeyMapValues();
std::string GetParameter(const std::string& name) const;
std::vector<std::string> GetSupportedKeys() const;
// precision of GNA hardware model
InferenceEngine::Precision gnaPrecision = InferenceEngine::Precision::I16;
std::string dumpXNNPath;
std::string dumpXNNGeneration;
#if GNA_LIB_VER == 1
intel_gna_proc_t gna_proc_type = static_cast<intel_gna_proc_t>(GNA_SOFTWARE & GNA_HARDWARE);
#else
Gna2AccelerationMode pluginGna2AccMode = Gna2AccelerationModeSoftware;
Gna2DeviceVersion pluginGna2DeviceConsistent = Gna2DeviceVersion1_0;
#endif
std::vector<float> inputScaleFactors;
GNAFlags gnaFlags;
std::map<std::string, std::string> key_config_map;
};
} // namespace GNAPluginNS

44
inference-engine/src/gna_plugin/gna_plugin_internal.hpp
View File

@ -10,36 +10,58 @@
#include <cpp_interfaces/impl/ie_plugin_internal.hpp>
#include <cpp_interfaces/impl/ie_executable_network_internal.hpp>
#include "gna_executable_network.hpp"
#include "gna_plugin_config.hpp"
namespace GNAPluginNS {
class GNAPluginInternal : public InferenceEngine::InferencePluginInternal {
public:
private:
Config defaultConfig;
std::weak_ptr <GNAPlugin> plgPtr;
std::shared_ptr<GNAPlugin> GetCurrentPlugin() const {
auto ptr = plgPtr.lock();
if (ptr == nullptr) {
return std::make_shared<GNAPlugin>();
} else {
return ptr;
}
}
public:
InferenceEngine::ExecutableNetworkInternal::Ptr LoadExeNetworkImpl(const InferenceEngine::ICore * core,
const InferenceEngine::ICNNNetwork &network,
const std::map<std::string, std::string> &config) override {
return std::make_shared<GNAExecutableNetwork>(*cloneNet(network), config);
Config updated_config(defaultConfig);
updated_config.UpdateFromMap(config);
auto plg = std::make_shared<GNAPlugin>(updated_config.key_config_map);
plgPtr = plg;
return std::make_shared<GNAExecutableNetwork>(*cloneNet(network), plg);
}
void SetConfig(const std::map<std::string, std::string> &config) override {
auto plg = std::make_shared<GNAPlugin>();
plg->SetConfig(config);
defaultConfig.UpdateFromMap(config);
}
InferenceEngine::IExecutableNetwork::Ptr ImportNetwork(
const std::string &modelFileName,
const std::map<std::string, std::string> &config) override {
return make_executable_network(std::make_shared<GNAExecutableNetwork>(modelFileName, config));
Config updated_config(defaultConfig);
updated_config.UpdateFromMap(config);
auto plg = std::make_shared<GNAPlugin>(updated_config.key_config_map);
plgPtr = plg;
return make_executable_network(std::make_shared<GNAExecutableNetwork>(modelFileName, plg));
}
using InferenceEngine::InferencePluginInternal::ImportNetwork;
std::string GetName() const noexcept override {
auto plg = std::make_shared<GNAPlugin>();
return plg->GetName();
return GetCurrentPlugin()->GetName();
}
void QueryNetwork(const InferenceEngine::ICNNNetwork& network,
const std::map<std::string, std::string>& config,
InferenceEngine::QueryNetworkResult& res) const override {
auto plg = std::make_shared<GNAPlugin>();
auto plg = GetCurrentPlugin();
try {
plg->SetConfig(config);
} catch (InferenceEngine::details::InferenceEngineException) {}
@ -48,13 +70,11 @@ class GNAPluginInternal : public InferenceEngine::InferencePluginInternal {
InferenceEngine::Parameter GetMetric(const std::string& name,
const std::map<std::string, InferenceEngine::Parameter> & options) const override {
GNAPlugin statelessPlugin;
return statelessPlugin.GetMetric(name, options);
return GetCurrentPlugin()->GetMetric(name, options);
}
InferenceEngine::Parameter GetConfig(const std::string& name, const std::map<std::string, InferenceEngine::Parameter> & options) const override {
GNAPlugin statelessPlugin;
return statelessPlugin.GetConfig(name, options);
return defaultConfig.GetParameter(name);
}
};

17
inference-engine/src/gna_plugin/gna_plugin_policy.hpp
View File

@ -37,8 +37,9 @@ class Policy {
enum class ConcatAlignment {
DISABLED,
DISABLED_FOR_FP32,
ENABLED
} ConcatAlignmentPolicy = ConcatAlignment::ENABLED;
ENABLED,
FAST
} ConcatAlignmentPolicy = ConcatAlignment::FAST;
};
inline std::ostream& operator<<(std::ostream& os, Policy::ScaleShift policy) {
@ -51,4 +52,16 @@ inline std::ostream& operator<<(std::ostream& os, Policy::ScaleShift policy) {
return os;
}
inline std::ostream& operator<<(std::ostream& os, Policy::ConcatAlignment policy) {
switch (policy) {
case Policy::ConcatAlignment::DISABLED : os << "DISABLED"; break;
case Policy::ConcatAlignment::DISABLED_FOR_FP32 : os << "DISABLED_FOR_FP32"; break;
case Policy::ConcatAlignment::ENABLED : os << "ENABLED"; break;
case Policy::ConcatAlignment::FAST : os << "FAST"; break;
default : os.setstate(std::ios_base::failbit);
}
return os;
}
} // namespace GNAPluginNS

37
inference-engine/src/gna_plugin/gna_plugin_query_api.cpp
View File

@ -16,19 +16,14 @@ using namespace GNAPluginNS;
using namespace InferenceEngine;
using namespace InferenceEngine::PluginConfigParams;
Parameter GNAPlugin::GetConfig(const std::string& name, const std::map<std::string, Parameter> & options) const {
auto configKeys = supportedConfigKeysWithDefaults();
auto result = configKeys.find(name);
if (result == configKeys.end()) {
THROW_GNA_EXCEPTION << "unsupported config key: " << name;
}
return result->second;
Parameter GNAPlugin::GetConfig(const std::string& name, const std::map<std::string, Parameter> & /*options*/) const {
return config.GetParameter(name);
}
Parameter GNAPlugin::GetMetric(const std::string& name, const std::map<std::string, InferenceEngine::Parameter> & options) const {
const std::unordered_map<std::string, std::function<Parameter()>> queryApiSupported = {
{METRIC_KEY(AVAILABLE_DEVICES), [this]() {return GetAvailableDevices();}},
{METRIC_KEY(SUPPORTED_CONFIG_KEYS), [this]() {return supportedConfigKeys();}},
{METRIC_KEY(SUPPORTED_CONFIG_KEYS), [this]() {return config.GetSupportedKeys();}},
{METRIC_KEY(FULL_DEVICE_NAME), [&options, this]() {
auto availableDevices = GetAvailableDevices().as<std::vector<std::string>>();
@ -100,29 +95,3 @@ Parameter GNAPlugin::GetAvailableDevices() const {
return devices;
}
std::map<std::string, std::string> GNAPlugin::supportedConfigKeysWithDefaults() const {
std::map<std::string, std::string> options = {
{GNA_CONFIG_KEY(SCALE_FACTOR), "1.0"},
{GNA_CONFIG_KEY(FIRMWARE_MODEL_IMAGE), ""},
{GNA_CONFIG_KEY(FIRMWARE_MODEL_IMAGE_GENERATION), ""},
{GNA_CONFIG_KEY(DEVICE_MODE), GNAConfigParams::GNA_AUTO},
{GNA_CONFIG_KEY(COMPACT_MODE), CONFIG_VALUE(NO)},
{CONFIG_KEY(EXCLUSIVE_ASYNC_REQUESTS), CONFIG_VALUE(NO)},
{GNA_CONFIG_KEY(PRECISION), Precision(Precision::I8).name()},
{GNA_CONFIG_KEY(PWL_UNIFORM_DESIGN), CONFIG_VALUE(YES)},
{CONFIG_KEY(PERF_COUNT), CONFIG_VALUE(NO)},
{GNA_CONFIG_KEY(LIB_N_THREADS), "1"},
{CONFIG_KEY(SINGLE_THREAD), CONFIG_VALUE(YES)}
};
return options;
}
std::vector<std::string> GNAPlugin::supportedConfigKeys()const {
std::vector<std::string> result;
for (auto && configOption : supportedConfigKeysWithDefaults()) {
result.push_back(configOption.first);
}
return result;
}

10
inference-engine/src/gna_plugin/optimizer/gna_pass_manager.cpp
View File

@ -675,6 +675,10 @@ void InsertCopyLayerPass::run() {
void InsertConcatAligningFilterPass::run() {
auto quantized = InferenceEngine::getInjectedData<QuantizedLayerParams>(pLayers->front());
if (getPassManager()->getPolicy().ConcatAlignmentPolicy == Policy::ConcatAlignment::DISABLED) {
return;
}
// aligning specific not required in fp32 mode
if (getPassManager()->getPolicy().ConcatAlignmentPolicy == Policy::ConcatAlignment::DISABLED_FOR_FP32 && !quantized) {
return;
@ -740,6 +744,10 @@ void InsertConcatAligningFilterPass::run() {
// encodes offset to beginning of split layer input
concatAligningFilter->params["output_offset"] =
std::to_string((aligned64_offset / bytesPerConcatElement) * (quantized ? bytesPerConcatElement : 4));
// for padded rows we cannot use copy layer - TBD how to implement
concatAligningFilter->params["num_rows_padded"] = std::to_string(num_rows_padded);
// encodes original output size
concatAligningFilter->params["original_num_rows"] = std::to_string(num_rows_in);
@ -1084,7 +1092,7 @@ int PassManager::run(int index) {
saveGraphToDot(*network.get(), out, [](const CNNLayerPtr layer,
ordered_properties &printed_properties,
ordered_properties &node_properties) {});
network->serialize(name + ".xml", "", nullptr);
network->serialize(name + ".xml", name + ".bin", nullptr);
};
#else
auto dumpNetworkAfterPass = [] (std::shared_ptr<Pass> ) {};

2
inference-engine/src/hetero_plugin/hetero_async_infer_request.cpp
View File

@ -48,7 +48,7 @@ HeteroAsyncInferRequest::HeteroAsyncInferRequest(const HeteroInferRequest::Ptr&
void HeteroAsyncInferRequest::StartAsync_ThreadUnsafe() {
_heteroInferRequest->updateInOutIfNeeded();
RunFirstStage();
RunFirstStage(_pipeline.begin(), _pipeline.end());
}
StatusCode HeteroAsyncInferRequest::Wait(int64_t millis_timeout) {

21
inference-engine/src/hetero_plugin/hetero_infer_request.cpp
View File

@ -65,6 +65,27 @@ HeteroInferRequest::HeteroInferRequest(InferenceEngine::InputsDataMap networkInp
}
}
void HeteroInferRequest::SetBlob(const char* name, const InferenceEngine::Blob::Ptr& data) {
InferenceEngine::InferRequestInternal::SetBlob(name, data);
assert(!_inferRequests.empty());
for (auto &&desc : _inferRequests) {
auto &r = desc._request;
assert(nullptr != r);
InputInfo::Ptr foundInput;
DataPtr foundOutput;
try {
// if `name` is input blob
if (findInputAndOutputBlobByName(name, foundInput, foundOutput)) {
r->SetBlob(name, data, foundInput->getPreProcess());
}
} catch (const InferenceEngine::details::InferenceEngineException & ex) {
std::string message = ex.what();
if (message.find(NOT_FOUND_str) == std::string::npos)
throw ex;
}
}
}
void HeteroInferRequest::InferImpl() {
updateInOutIfNeeded();
size_t i = 0;

2
inference-engine/src/hetero_plugin/hetero_infer_request.hpp
View File

@ -39,6 +39,8 @@ public:
void InferImpl() override;
void SetBlob(const char* name, const InferenceEngine::Blob::Ptr& data) override;
void GetPerformanceCounts(std::map<std::string, InferenceEngine::InferenceEngineProfileInfo> &perfMap) const override;
void updateInOutIfNeeded();

15
inference-engine/src/inference_engine/CMakeLists.txt
View File

@ -17,6 +17,8 @@ set(IE_STATIC_DEPENDENT_FILES ${CMAKE_CURRENT_SOURCE_DIR}/file_utils.cpp)
list(REMOVE_ITEM LIBRARY_SRC ${IE_STATIC_DEPENDENT_FILES})
set(IE_BASE_SOURCE_FILES
${CMAKE_CURRENT_SOURCE_DIR}/cnn_network_ngraph_impl.cpp
${CMAKE_CURRENT_SOURCE_DIR}/generic_ie.cpp
${CMAKE_CURRENT_SOURCE_DIR}/blob_factory.cpp
${CMAKE_CURRENT_SOURCE_DIR}/ie_data.cpp
${CMAKE_CURRENT_SOURCE_DIR}/ie_layouts.cpp
@ -94,7 +96,6 @@ file(GLOB_RECURSE plugin_api_src "${IE_MAIN_SOURCE_DIR}/src/plugin_api/*.hpp"
"${IE_MAIN_SOURCE_DIR}/src/plugin_api/*.h")
add_cpplint_target(${TARGET_NAME}_plugin_api_cpplint FOR_SOURCES ${plugin_api_src})
add_clang_format_target(${TARGET_NAME}_plugin_api_clang_format FOR_SOURCES ${plugin_api_src})
# Create common base object library
@ -103,6 +104,7 @@ add_library(${TARGET_NAME}_common_obj OBJECT
target_compile_definitions(${TARGET_NAME}_common_obj PRIVATE IMPLEMENT_INFERENCE_ENGINE_API)
target_include_directories(${TARGET_NAME}_common_obj PRIVATE
$<TARGET_PROPERTY:${TARGET_NAME}_transformations,INTERFACE_INCLUDE_DIRECTORIES>
$<TARGET_PROPERTY:${TARGET_NAME}_plugin_api,INTERFACE_INCLUDE_DIRECTORIES>)
target_include_directories(${TARGET_NAME}_common_obj SYSTEM PRIVATE
@ -121,6 +123,7 @@ target_include_directories(${TARGET_NAME}_obj SYSTEM PRIVATE $<TARGET_PROPERTY:n
$<TARGET_PROPERTY:pugixml,INTERFACE_INCLUDE_DIRECTORIES>)
target_include_directories(${TARGET_NAME}_obj PRIVATE "${CMAKE_CURRENT_SOURCE_DIR}"
$<TARGET_PROPERTY:${TARGET_NAME}_ir_readers,INTERFACE_INCLUDE_DIRECTORIES>
$<TARGET_PROPERTY:${TARGET_NAME}_plugin_api,INTERFACE_INCLUDE_DIRECTORIES>)
if(ENABLE_PROFILING_ITT AND INTEL_ITT_LIBS)
@ -146,7 +149,6 @@ if(WIN32)
endif()
add_cpplint_target(${TARGET_NAME}_cpplint FOR_TARGETS ${TARGET_NAME}_obj)
add_clang_format_target(${TARGET_NAME}_clang_format FOR_TARGETS ${TARGET_NAME}_obj)
# Create shared library file from object library
@ -196,7 +198,6 @@ target_include_directories(${TARGET_NAME}_nn_builder PRIVATE "${CMAKE_CURRENT_SO
"${IE_MAIN_SOURCE_DIR}/src/legacy_api/src")
add_cpplint_target(${TARGET_NAME}_nn_builder_cpplint FOR_TARGETS ${TARGET_NAME}_nn_builder)
add_clang_format_target(${TARGET_NAME}_nn_builder_clang_format FOR_TARGETS ${TARGET_NAME}_nn_builder)
# Static library used for unit tests which are always built
@ -272,6 +273,10 @@ if(THREADING STREQUAL "TBB" OR THREADING STREQUAL "TBB_AUTO")
install(FILES "${TBB}/LICENSE"
DESTINATION ${IE_CPACK_IE_DIR}/external/tbb
COMPONENT tbb)
install(FILES "${TBB}/cmake/TBBConfig.cmake"
"${TBB}/cmake/TBBConfigVersion.cmake"
DESTINATION ${IE_CPACK_IE_DIR}/external/tbb/cmake
COMPONENT tbb)
endif()
ie_cpack_add_component(core REQUIRED DEPENDS ${core_components})
@ -279,8 +284,8 @@ ie_cpack_add_component(core REQUIRED DEPENDS ${core_components})
install(DIRECTORY "${IE_MAIN_SOURCE_DIR}/include" DESTINATION ${IE_CPACK_IE_DIR}
COMPONENT core)
install(TARGETS ${TARGET_NAME} ${TARGET_NAME}_nn_builder
RUNTIME DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT core
ARCHIVE DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT core
RUNTIME DESTINATION ${IE_CPACK_RUNTIME_PATH} COMPONENT core
ARCHIVE DESTINATION ${IE_CPACK_ARCHIVE_PATH} COMPONENT core
LIBRARY DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT core)
install(FILES "${OpenVINO_BINARY_DIR}/share/ie_parallel.cmake"
"${OpenVINO_BINARY_DIR}/share/InferenceEngineConfig.cmake"

2
inference-engine/src/inference_engine/blob_factory.cpp
View File

@ -46,6 +46,8 @@ InferenceEngine::Blob::Ptr CreateBlobFromData(const InferenceEngine::DataPtr& da
return std::make_shared<InferenceEngine::TBlob<int8_t>>(desc);
case InferenceEngine::Precision::I32:
return std::make_shared<InferenceEngine::TBlob<int32_t>>(desc);
case InferenceEngine::Precision::BF16:
return std::make_shared<InferenceEngine::TBlob<short>>(desc);
default:
THROW_IE_EXCEPTION << "precision is no set";
}

287
inference-engine/src/inference_engine/cnn_network_ngraph_impl.cpp
View File

@ -28,7 +28,6 @@
#include "graph_tools.hpp"
#include "graph_transformer.h"
#include "ie_util_internal.hpp"
#include "ie_cnn_layer_builder_ngraph.h"
#include "ie_ngraph_utils.hpp"
#include "ie_profiling.hpp"
#include "network_serializer.h"
@ -97,7 +96,7 @@ void CNNNetworkNGraphImpl::createDataForResult(const ::ngraph::Output<::ngraph::
if (ptr) {
ptr->reshape(dims, ptr->getTensorDesc().getLayout());
} else {
const auto precision = details::ngraph::convertPrecision(output.get_element_type());
const auto precision = details::convertPrecision(output.get_element_type());
const auto layout = TensorDesc::getLayoutByDims(dims);
ptr.reset(new NGraphData(this, outName, {precision, dims, layout}));
}
@ -520,287 +519,3 @@ void CNNNetworkNGraphImpl::convertToCNNNetworkImpl() {
::ngraph::pass::ConvertOpSet1ToLegacy().run_on_function(graph);
cnnNetwork = InferenceEngine::details::convertFunctionToICNNNetwork(graph, *this);
}
std::shared_ptr<CNNNetworkNGraphImpl> CNNNetworkNGraphImpl::cloneNGraphImpl() const {
auto result = std::make_shared<CNNNetworkNGraphImpl>(cloneFunction());
for (const auto& outputInfo : _outputData) {
result->_outputData[outputInfo.first]->setPrecision(outputInfo.second->getPrecision());
result->_outputData[outputInfo.first]->setLayout(outputInfo.second->getLayout());
}
for (const auto& inputInfo : _inputData) {
result->_inputData[inputInfo.first]->setPrecision(inputInfo.second->getPrecision());
result->_inputData[inputInfo.first]->setLayout(inputInfo.second->getLayout());
result->_inputData[inputInfo.first]->getPreProcess() = inputInfo.second->getPreProcess();
}
if (cnnNetwork)
result->cnnNetwork = cloneNet(*cnnNetwork);
return result;
}
void CNNNetworkNGraphImpl::transformConstants() {
if (!cnnNetwork)
convertToCNNNetworkImpl();
// Remove all redundant constant and convert unsupported precisions
ConstTransformer transformator(cnnNetwork.get());
transformator.fullTrim();
}
void InferenceEngine::details::CNNLayerCreator::on_adapter(const std::string& name,
::ngraph::ValueAccessor<void>& adapter) {
if (auto a = ::ngraph::as_type<::ngraph::AttributeAdapter<::ngraph::element::Type>>(&adapter)) {
auto type = static_cast<::ngraph::element::Type&>(*a);
params[name] = details::ngraph::convertPrecision(type).name();
} else if (auto a = ::ngraph::as_type<::ngraph::AttributeAdapter<::ngraph::PartialShape>>(&adapter)) {
std::string dims;
auto shape = static_cast<::ngraph::PartialShape&>(*a);
for (size_t i = 0; i < shape.rank().get_length(); i++) {
if (!dims.empty()) dims += ",";
dims += std::to_string(shape[i].get_length());
}
params[name] = dims;
} else if (auto a = ::ngraph::as_type<::ngraph::AttributeAdapter<::ngraph::Shape>>(&adapter)) {
std::string dims;
auto shape = static_cast<::ngraph::Shape&>(*a);
for (size_t i = 0; i < shape.size(); i++) {
if (!dims.empty()) dims += ",";
dims += std::to_string(shape[i]);
}
params[name] = dims;
} else if (auto a = ::ngraph::as_type<::ngraph::AttributeAdapter<::ngraph::Strides>>(&adapter)) {
std::string dims;
auto shape = static_cast<::ngraph::Strides&>(*a);
for (size_t i = 0; i < shape.size(); i++) {
if (!dims.empty()) dims += ",";
dims += std::to_string(shape[i]);
}
params[name] = dims;
}
}
InferenceEngine::details::CNNLayerCreator::CNNLayerCreator(const std::shared_ptr<::ngraph::Node>& node): node(node) {
addSpecificCreator({"Parameter"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), "Input",
details::ngraph::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<CNNLayer>(attrs);
return res;
});
// TODO - Remove "GreaterEq" once ngraph transitions to GreaterEqual
addSpecificCreator({"Eltwise", "Subtract", "Power", "Maximum", "Divide", "Greater", "GreaterEqual", "FloorMod", "LogicalOr", "LogicalAnd", "LogicalXor",
"GreaterEq", "Less", "LessEqual", "Equal", "NotEqual", "Multiply", "Add"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), "Eltwise",
details::ngraph::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<EltwiseLayer>(attrs);
res->params = params;
if (node->description() == "Maximum") {
res->params["operation"] = "max";
} else if (node->description() == "Power") {
res->params["operation"] = "pow";
} else if (node->description() == "Subtract") {
res->params["operation"] = "sub";
} else if (node->description() == "Divide") {
res->params["operation"] = "div";
} else if (node->description() == "LessEqual") {
res->params["operation"] = "less_equal";
} else if (node->description() == "Less") {
res->params["operation"] = "less";
} else if (node->description() == "Equal") {
res->params["operation"] = "equal";
} else if (node->description() == "NotEqual") {
res->params["operation"] = "not_equal";
} else if (node->description() == "FloorMod") {
res->params["operation"] = "floor_mod";
} else if (node->description() == "Multiply") {
res->params["operation"] = "prod";
} else if (node->description() == "Add") {
res->params["operation"] = "sum";
} else if (node->description() == "Greater") {
res->params["operation"] = "greater";
} else if (node->description() == "GreaterEq") {
res->params["operation"] = "greater_equal";
} else if (node->description() == "GreaterEqual") {
res->params["operation"] = "greater_equal";
} else if (node->description() == "LogicalOr") {
res->params["operation"] = "logical_or";
} else if (node->description() == "LogicalAnd") {
res->params["operation"] = "logical_and";
} else if (node->description() == "LogicalXor") {
res->params["operation"] = "logical_xor";
} else if (node->description() == "Eltwise") {
auto castedLayer = std::dynamic_pointer_cast<::ngraph::op::Eltwise>(node);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << attrs.type << " layer " << attrs.name;
std::string type;
switch (castedLayer->eltwise_type) {
case ELTWISE_TYPE::Sum:
type = "sum";
break;
case ELTWISE_TYPE::Prod:
type = "prod";
break;
default:
THROW_IE_EXCEPTION << "Not supported eltwise type!";
}
res->params["operation"] = type;
}
return res;
});
addSpecificCreator({"Concat"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), node->description(),
details::ngraph::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<ConcatLayer>(attrs);
res->params = params;
return res;
});
addSpecificCreator({"AvgPool", "MaxPool"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), "Pooling",
details::ngraph::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<PoolingLayer>(attrs);
res->params = params;
if (res->params.find("auto_pad") != res->params.end() &&
details::CaselessEq<std::string>()(res->params["auto_pad"], "EXPLICIT"))
res->params.erase("auto_pad");
if (res->params.find("exclude_pad") != res->params.end()) {
res->params["exclude-pad"] = res->params["exclude_pad"];
res->params.erase("exclude_pad");
}
if (node->description() == "MaxPool") {
res->params["pool-method"] = "max";
} else if (node->description() == "AvgPool") {
res->params["pool-method"] = "avg";
}
return res;
});
addSpecificCreator({"Select"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), node->description(),
details::ngraph::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<SelectLayer>(attrs);
res->params = params;
return res;
});
addSpecificCreator({"BinaryConvolution"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), node->description(),
details::ngraph::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<BinaryConvolutionLayer>(attrs);
// todo: investigate difference between ngraph parameters for BinConvolution and the implementation above
// this leads to accuracy issue for Precollected_ONNX_ResNet50_88percentinto1bit e2e test
// res->params = params;
auto castedLayer = ::ngraph::as_type_ptr<::ngraph::op::v1::BinaryConvolution>(node);
std::string value;
for (const auto& val : castedLayer->get_pads_begin()) {
if (!value.empty()) value += ",";
value += Builder::asString(val);
}
res->params["pads_begin"] = value;
value.clear();
for (const auto& val : castedLayer->get_pads_end()) {
if (!value.empty()) value += ",";
value += Builder::asString(val);
}
res->params["pads_end"] = value;
switch (castedLayer->get_auto_pad()) {
case ::ngraph::op::PadType::SAME_UPPER:
res->params["auto_pad"] = "same_upper";
break;
case ::ngraph::op::PadType::SAME_LOWER:
res->params["auto_pad"] = "same_lower";
break;
case ::ngraph::op::PadType::VALID:
res->params["auto_pad"] = "valid";
break;
default:
break;
}
value.clear();
for (const auto& val : castedLayer->get_strides()) {
if (!value.empty()) value += ",";
value += Builder::asString(val);
}
res->params["strides"] = value;
value.clear();
for (const auto& val : castedLayer->get_dilations()) {
if (!value.empty()) value += ",";
value += Builder::asString(val);
}
res->params["dilations"] = value;
// Restore kernel size and output
const auto& shape = castedLayer->get_input_shape(1);
res->params["output"] = Builder::asString(shape[0]);
value.clear();
for (size_t i = 2; i < shape.size(); i++) {
if (!value.empty()) value += ",";
value += Builder::asString(shape[i]);
}
res->params["kernel"] = value;
switch (castedLayer->get_mode()) {
case ::ngraph::op::v1::BinaryConvolution::BinaryConvolutionMode::XNOR_POPCOUNT:
res->params["mode"] = "xnor-popcount";
}
auto weights_shape = castedLayer->input(1).get_source_output().get_shape();
res->params["input"] = Builder::asString(weights_shape[1]);
res->params["pad_value"] = Builder::asString(castedLayer->get_pad_value());
Builder::NodeConverter<::ngraph::op::Constant> converter;
const auto weightsNode = castedLayer->get_inputs()[1].get_output().get_node();
if (converter.canCreate(weightsNode)) {
const auto& weights = converter.createLayer(weightsNode);
res->blobs["weights"] = weights->blobs["custom"];
res->_weights = weights->blobs["custom"];
}
return res;
});
addSpecificCreator({"SpaceToBatch"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), node->description(),
details::ngraph::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<SpaceToBatchLayer>(attrs);
res->params = params;
return res;
});
addSpecificCreator({"BatchToSpace"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), node->description(),
details::ngraph::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<BatchToSpaceLayer>(attrs);
res->params = params;
return res;
});
}
CNNLayerPtr InferenceEngine::details::CNNLayerCreator::create() {
auto one_from = [](const std::string& desc, const std::vector<std::string>& descs) -> bool {
for (const auto& d : descs) {
if (details::CaselessEq<std::string>()(d, desc)) return true;
}
return false;
};
LayerParams attrs = {node->get_friendly_name(), node->description(),
details::ngraph::convertPrecision(node->get_output_element_type(0))};
if (creators.find(node->description()) != creators.end())
return creators[node->description()](node, params);
auto res = std::make_shared<CNNLayer>(attrs);
res->params = params;
return res;
}

66
inference-engine/src/plugin_api/cnn_network_ngraph_impl.hpp → inference-engine/src/inference_engine/cnn_network_ngraph_impl.hpp
View File

@ -102,7 +102,7 @@ public:
std::shared_ptr<const ::ngraph::Function> getFunction() const noexcept override {
return !cnnNetwork ? _ngraph_function : nullptr;
}
std::shared_ptr<::ngraph::Function> getFunction() noexcept {
std::shared_ptr<::ngraph::Function> getFunction() noexcept override {
return !cnnNetwork ? _ngraph_function : nullptr;
}
@ -118,9 +118,6 @@ public:
noexcept override;
void convertToCNNNetworkImpl();
std::shared_ptr<CNNNetworkNGraphImpl> cloneNGraphImpl() const;
void transformConstants();
protected:
std::shared_ptr<::ngraph::Function> _ngraph_function;
virtual std::shared_ptr<::ngraph::Function> cloneFunction(bool constFolding = false, const std::map<std::string,
@ -142,7 +139,7 @@ private:
friend INFERENCE_ENGINE_API_CPP(std::shared_ptr<CNNNetworkImpl>)
convertFunctionToICNNNetwork(const std::shared_ptr<const ::ngraph::Function>& graph,
const CNNNetworkNGraphImpl & nGraphImpl);
const ICNNNetwork& nGraphImpl);
/**
* @brief Reshape on the same shape
@ -196,65 +193,6 @@ private:
IE_SUPPRESS_DEPRECATED_END
/**
* @brief Creator for CNNLayer from nGraph op
*/
class CNNLayerCreator : public ::ngraph::AttributeVisitor {
public:
using CreatorFor = std::function<CNNLayerPtr(const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> param)>;
explicit CNNLayerCreator(const std::shared_ptr<::ngraph::Node>& node);
CNNLayerPtr create();
void on_attribute(const std::string& name, std::string& value) override {
params[name] = value;
}
void on_attribute(const std::string& name, bool& value) override {
params[name] = value ? "true" : "false";
}
void addSpecificCreator(const std::vector<std::string>& forTypes, const CreatorFor& creator) {
for (const auto type : forTypes) {
creators[type] = creator;
}
}
void on_adapter(const std::string& name, ::ngraph::ValueAccessor<std::string>& adapter) override {
std::string data = adapter.get();
std::transform(data.begin(), data.end(), data.begin(), [](unsigned char c) {
return std::tolower(c);
});
params[name] = data;
}
void on_adapter(const std::string& name, ::ngraph::ValueAccessor<std::vector<int64_t>>& adapter) override {
std::string dims;
auto shape = adapter.get();
for (size_t i = 0; i < shape.size(); i++) {
if (!dims.empty()) dims += ",";
dims += std::to_string(shape[i]);
}
params[name] = dims;
}
void on_adapter(const std::string& name, ::ngraph::ValueAccessor<double>& adapter) override {
params[name] = std::to_string(adapter.get());
}
void on_adapter(const std::string& name, ::ngraph::ValueAccessor<int64_t>& adapter) override {
params[name] = std::to_string(adapter.get());
}
void on_adapter(const std::string& name, ::ngraph::ValueAccessor<void>& adapter) override;
private:
std::shared_ptr<::ngraph::Node> node;
std::map<std::string, std::string> params;
std::map<std::string, CreatorFor> creators;
};
typedef std::shared_ptr<CNNNetworkNGraphImpl> CNNNetworkNGraphImplPtr;
} // namespace details
} // namespace InferenceEngine

13
inference-engine/src/inference_engine/generic_ie.cpp
View File

@ -96,7 +96,7 @@ void ngraph::op::GenericIE::validate_and_infer_types() {
// Set dynamic output shapes if input shapes are not defined
for (size_t i = 0; i < outputs.size(); i++) {
const auto& port = outputs[i];
auto type = InferenceEngine::details::ngraph::convertPrecision(port.precision);
auto type = InferenceEngine::details::convertPrecision(port.precision);
set_output_type(i, type, PartialShape::dynamic());
}
return;
@ -105,7 +105,7 @@ void ngraph::op::GenericIE::validate_and_infer_types() {
Shape this_ishape = get_input_shape(i);
InferenceEngine::SizeVector dims = this_ishape;
InferenceEngine::Blob::Ptr input = make_blob_with_precision(InferenceEngine::TensorDesc(
InferenceEngine::details::ngraph::convertPrecision(get_input_element_type(i)), dims,
InferenceEngine::details::convertPrecision(get_input_element_type(i)), dims,
InferenceEngine::TensorDesc::getLayoutByDims(dims)));
inputs.emplace_back(input);
}
@ -126,6 +126,11 @@ void ngraph::op::GenericIE::validate_and_infer_types() {
}
}
// WA: Proposal shape infer has to know number of outputs
if (type == "Proposal" && parameters.find("num_outputs") == parameters.end()) {
parameters["num_outputs"] = std::to_string(outputs.size());
}
ret = impl->inferShapes(inputs, parameters, blobs, outShapes, nullptr);
IE_SUPPRESS_DEPRECATED_END
@ -134,7 +139,7 @@ void ngraph::op::GenericIE::validate_and_infer_types() {
for (size_t i = 0; i < outputs.size(); i++) {
const auto& port = outputs[i];
ngraph::Shape outShape(outShapes[i]);
auto type = InferenceEngine::details::ngraph::convertPrecision(port.precision);
auto type = InferenceEngine::details::convertPrecision(port.precision);
set_output_type(i, type, PartialShape(outShape));
}
@ -149,7 +154,7 @@ void ngraph::op::GenericIE::validate_and_infer_types() {
for (size_t i = 0; i < outputs.size(); i++) {
const auto& port = outputs[i];
ngraph::Shape outShape(port.dims);
auto type = InferenceEngine::details::ngraph::convertPrecision(port.precision);
auto type = InferenceEngine::details::convertPrecision(port.precision);
set_output_type(i, type, PartialShape(outShape));
}
initialized++;

70
inference-engine/src/inference_engine/ie_core.cpp
View File

@ -16,13 +16,12 @@
#include <vector>
#include <ngraph/opsets/opset.hpp>
#include "cpp/ie_cnn_net_reader.h"
#include "cpp_interfaces/base/ie_plugin_base.hpp"
#include "details/ie_exception_conversion.hpp"
#include "details/ie_so_pointer.hpp"
#include "file_utils.h"
#include "ie_cnn_net_reader_impl.h"
#include "ie_icore.hpp"
#include "ie_ir_reader.hpp"
#include "ie_plugin.hpp"
#include "ie_plugin_config.hpp"
#include "ie_profiling.hpp"
@ -38,6 +37,27 @@ IE_SUPPRESS_DEPRECATED_START
namespace {
std::once_flag flag;
std::shared_ptr<InferenceEngine::details::SharedObjectLoader> cnnReaderLoader;
std::shared_ptr<InferenceEngine::details::SharedObjectLoader>
createCnnReaderLoader() {
std::call_once(flag, [&] () {
FileUtils::FilePath libraryName = FileUtils::toFilePath(std::string("inference_engine_ir_readers") + std::string(IE_BUILD_POSTFIX));
FileUtils::FilePath irReadersLibraryPath = FileUtils::makeSharedLibraryName(getInferenceEngineLibraryPath(), libraryName);
if (!FileUtils::fileExist(irReadersLibraryPath)) {
THROW_IE_EXCEPTION << "Please, make sure that Inference Engine IR readers library "
<< FileUtils::fromFilePath(::FileUtils::makeSharedLibraryName({}, libraryName)) << " is in "
<< getIELibraryPath();
}
cnnReaderLoader = std::shared_ptr<InferenceEngine::details::SharedObjectLoader>(
new InferenceEngine::details::SharedObjectLoader(irReadersLibraryPath.c_str()));
});
return cnnReaderLoader;
}
IInferencePluginAPI* getInferencePluginAPIInterface(IInferencePlugin* iplugin) {
return dynamic_cast<IInferencePluginAPI*>(iplugin);
}
@ -52,6 +72,11 @@ IInferencePluginAPI* getInferencePluginAPIInterface(InferencePlugin plugin) {
} // namespace
CNNNetReaderPtr CreateCNNNetReaderPtr() noexcept {
auto loader = createCnnReaderLoader();
return CNNNetReaderPtr(loader);
}
IE_SUPPRESS_DEPRECATED_END
DeviceIDParser::DeviceIDParser(const std::string& deviceNameWithID) {
@ -112,6 +137,7 @@ std::vector<std::string> DeviceIDParser::getMultiDevices(std::string devicesList
}
class Core::Impl : public ICore {
// Fields are ordered by deletion order
ITaskExecutor::Ptr _taskExecutor = nullptr;
IE_SUPPRESS_DEPRECATED_START
@ -124,10 +150,11 @@ class Core::Impl : public ICore {
std::vector<FileUtils::FilePath> listOfExtentions;
};
std::map<std::string, PluginDescriptor> pluginRegistry;
std::unordered_set<std::string> opsetNames;
std::vector<IExtensionPtr> extensions;
std::map<std::string, PluginDescriptor> pluginRegistry;
public:
Impl();
~Impl() override;
@ -395,12 +422,18 @@ std::map<std::string, Version> Core::GetVersions(const std::string& deviceName)
{
// for compatibility with samples / demo
if (deviceName.find("HETERO:") == 0) {
deviceNames = DeviceIDParser::getHeteroDevices(deviceName.substr(7));
if (deviceName.find("HETERO") == 0) {
auto pos = deviceName.find_first_of(":");
if (pos != std::string::npos) {
deviceNames = DeviceIDParser::getHeteroDevices(deviceName.substr(pos + 1));
}
deviceNames.push_back("HETERO");
} else if (deviceName.find("MULTI") == 0) {
auto pos = deviceName.find_first_of(":");
if (pos != std::string::npos) {
deviceNames = DeviceIDParser::getMultiDevices(deviceName.substr(pos + 1));
}
deviceNames.push_back("MULTI");
deviceNames = DeviceIDParser::getMultiDevices(deviceName.substr(6));
} else {
deviceNames.push_back(deviceName);
}
@ -457,13 +490,12 @@ Parsed<T> parseDeviceNameIntoConfig(const std::string& deviceName, const std::ma
CNNNetwork Core::ReadNetwork(const std::string& modelPath, const std::string& binPath) const {
IE_PROFILING_AUTO_SCOPE(Core::ReadNetwork)
IE_SUPPRESS_DEPRECATED_START
auto cnnReader = std::shared_ptr<ICNNNetReader>(CreateCNNNetReader());
ResponseDesc desc;
CNNNetReaderPtr cnnReader(createCnnReaderLoader());
StatusCode rt = cnnReader->ReadNetwork(modelPath.c_str(), &desc);
if (rt != OK) THROW_IE_EXCEPTION << desc.msg;
auto cnnNetReaderImpl = std::dynamic_pointer_cast<details::CNNNetReaderImpl>(cnnReader);
if (cnnNetReaderImpl && cnnReader->getVersion(&desc) >= 10) {
cnnNetReaderImpl->addExtensions(_impl->getExtensions());
if (cnnReader->getVersion(&desc) >= 10) {
cnnReader->addExtensions(_impl->getExtensions());
}
std::string bPath = binPath;
if (bPath.empty()) {
@ -491,13 +523,12 @@ CNNNetwork Core::ReadNetwork(const std::string& modelPath, const std::string& bi
CNNNetwork Core::ReadNetwork(const std::string& model, const Blob::CPtr& weights) const {
IE_PROFILING_AUTO_SCOPE(Core::ReadNetwork)
IE_SUPPRESS_DEPRECATED_START
auto cnnReader = std::shared_ptr<ICNNNetReader>(CreateCNNNetReader());
ResponseDesc desc;
CNNNetReaderPtr cnnReader(createCnnReaderLoader());
StatusCode rt = cnnReader->ReadNetwork(model.data(), model.length(), &desc);
if (rt != OK) THROW_IE_EXCEPTION << desc.msg;
auto cnnNetReaderImpl = std::dynamic_pointer_cast<details::CNNNetReaderImpl>(cnnReader);
if (cnnNetReaderImpl && cnnReader->getVersion(&desc) >= 10) {
cnnNetReaderImpl->addExtensions(_impl->getExtensions());
if (cnnReader->getVersion(&desc) >= 10) {
cnnReader->addExtensions(_impl->getExtensions());
}
TBlob<uint8_t>::Ptr weights_ptr;
if (weights) {
@ -507,6 +538,7 @@ CNNNetwork Core::ReadNetwork(const std::string& model, const Blob::CPtr& weights
rt = cnnReader->SetWeights(weights_ptr, &desc);
if (rt != OK) THROW_IE_EXCEPTION << desc.msg;
IE_SUPPRESS_DEPRECATED_END
return CNNNetwork(cnnReader);
}
@ -694,11 +726,6 @@ void Core::SetConfig(const std::map<std::string, std::string>& config, const std
THROW_IE_EXCEPTION << "SetConfig is supported only for HETERO itself (without devices). "
"You can configure the devices with SetConfig before creating the HETERO on top.";
}
if (config.find("TARGET_FALLBACK") != config.end()) {
THROW_IE_EXCEPTION << "Please, specify TARGET_FALLBACK to the LoadNetwork directly, "
"as you will need to pass the same TARGET_FALLBACK anyway.";
}
}
// MULTI case
@ -707,11 +734,6 @@ void Core::SetConfig(const std::map<std::string, std::string>& config, const std
THROW_IE_EXCEPTION << "SetConfig is supported only for MULTI itself (without devices). "
"You can configure the devices with SetConfig before creating the MULTI on top.";
}
if (config.find(MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES) != config.end()) {
THROW_IE_EXCEPTION << "Please, specify DEVICE_PRIORITIES to the LoadNetwork directly, "
"as you will need to pass the same DEVICE_PRIORITIES anyway.";
}
}
if (deviceName.empty()) {

2
inference-engine/src/inference_engine/ie_rtti.cpp
View File

@ -13,6 +13,7 @@
#include <ie_parameter.hpp>
#include <ie_iextension.h>
#include <ie_extension.h>
#include <ngraph/opsets/opset.hpp>
using namespace InferenceEngine;
@ -83,6 +84,7 @@ template struct InferenceEngine::Parameter::RealData<std::vector<std::string>>;
template struct InferenceEngine::Parameter::RealData<std::vector<unsigned long>>;
template struct InferenceEngine::Parameter::RealData<std::tuple<unsigned int, unsigned int>>;
template struct InferenceEngine::Parameter::RealData<std::tuple<unsigned int, unsigned int, unsigned int>>;
template struct InferenceEngine::Parameter::RealData<InferenceEngine::Blob::Ptr>;
#endif // __clang__
//
// ie_blob.h

7
inference-engine/src/inference_engine/ie_system_conf.cpp
View File

@ -71,6 +71,13 @@ bool with_cpu_x86_avx512_core() {
#endif
}
bool with_cpu_x86_bfloat16() {
#ifdef ENABLE_MKL_DNN
return cpu.has(Xbyak::util::Cpu::tAVX512_BF16);
#else
return false;
#endif
}
bool checkOpenMpEnvVars(bool includeOMPNumThreads) {
for (auto&& var : {

2
inference-engine/src/inference_engine/os/lin/lin_system_conf.cpp
View File

@ -1,4 +1,4 @@
// Copyright (C) 2018-2020 Intel Corporation
// Copyright (C) 2018-2019 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//

2
inference-engine/src/inference_engine/threading/ie_cpu_streams_executor.cpp
View File

@ -1,4 +1,4 @@
// Copyright (C) 2018-2020 Intel Corporation
// Copyright (C) 2018-2019 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//

43
inference-engine/src/inference_engine/threading/ie_executor_manager.cpp
View File

@ -12,6 +12,7 @@
namespace InferenceEngine {
ITaskExecutor::Ptr ExecutorManagerImpl::getExecutor(std::string id) {
std::lock_guard<std::mutex> guard(taskExecutorMutex);
auto foundEntry = executors.find(id);
if (foundEntry == executors.end()) {
auto newExec = std::make_shared<CPUStreamsExecutor>(IStreamsExecutor::Config{id});
@ -22,6 +23,7 @@ ITaskExecutor::Ptr ExecutorManagerImpl::getExecutor(std::string id) {
}
IStreamsExecutor::Ptr ExecutorManagerImpl::getIdleCPUStreamsExecutor(const IStreamsExecutor::Config& config) {
std::lock_guard<std::mutex> guard(streamExecutorMutex);
for (const auto& it : cpuStreamsExecutors) {
const auto& executor = it.second;
if (executor.use_count() != 1)
@ -52,6 +54,8 @@ size_t ExecutorManagerImpl::getIdleCPUStreamsExecutorsNumber() {
}
void ExecutorManagerImpl::clear(const std::string& id) {
std::lock_guard<std::mutex> stream_guard(streamExecutorMutex);
std::lock_guard<std::mutex> task_guard(taskExecutorMutex);
if (id.empty()) {
executors.clear();
cpuStreamsExecutors.clear();
@ -66,8 +70,47 @@ void ExecutorManagerImpl::clear(const std::string& id) {
}
}
std::mutex ExecutorManager::_mutex;
ExecutorManager* ExecutorManager::_instance = nullptr;
ExecutorManager* ExecutorManager::getInstance() {
/*
* 1) We do not use singleton implementation via STATIC LOCAL object like
*
* getInstance() {
* static ExecutorManager _instance;
* return &instance;
* }
*
* Because of problem with destruction order on program exit.
* Some IE classes like MKLDNN::Engine use this singleton in destructor.
* But they has no direct dependency from c++ runtime point of view and
* it's possible that _instance local static variable will be destroyed
* before MKLDNN::~Engine call. Any further manipulation with destroyed
* object will lead to exception or crashes.
*
* 2) We do not use singleton implementation via STATIC object like:
*
* ExecutorManager ExecutorManager::_instance;
* getInstance() {
* return &instance;
* }
*
* Because of problem with double destruction. In some test cases we use
* double link with IE module via static and dynamic version. Both modules
* have static object with same export name and it leads to double construction
* and double destruction of that object. For some c++ compilers (ex gcc 5.4)
* it lead to crash with "double free".
*
* That's why we use manual allocation of singleton instance on heap.
*/
std::lock_guard<std::mutex> guard(_mutex);
if (_instance == nullptr) {
_instance = new ExecutorManager();
}
return _instance;
}
ITaskExecutor::Ptr ExecutorManager::getExecutor(std::string id) {
return _impl.getExecutor(id);
}

2
inference-engine/src/inference_engine/threading/ie_thread_affinity.cpp
View File

@ -1,4 +1,4 @@
// Copyright (C) 2018-2020 Intel Corporation
// Copyright (C) 2018-2019 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//

49
inference-engine/src/ir_readers/CMakeLists.txt Normal file
View File

@ -0,0 +1,49 @@
# Copyright (C) 2018-2019 Intel Corporation
# SPDX-License-Identifier: Apache-2.0
#
set(TARGET_NAME "inference_engine_ir_readers")
if(ENABLE_LTO)
ie_enable_lto()
endif()
set(PUBLIC_HEADERS_DIR "${CMAKE_CURRENT_SOURCE_DIR}/")
file(GLOB_RECURSE LIBRARY_SRC ${CMAKE_CURRENT_SOURCE_DIR}/*.cpp)
file(GLOB_RECURSE PUBLIC_HEADERS ${PUBLIC_HEADERS_DIR}/*.h ${PUBLIC_HEADERS_DIR}/*.hpp)
# Create named folders for the sources within the .vcproj
# Empty name lists them directly under the .vcproj
source_group("src" FILES ${LIBRARY_SRC})
source_group("include" FILES ${PUBLIC_HEADERS})
# Create shared library
add_library(${TARGET_NAME} SHARED ${LIBRARY_SRC} ${PUBLIC_HEADERS})
target_compile_definitions(${TARGET_NAME} PRIVATE IMPLEMENT_INFERENCE_ENGINE_API
IMPLEMENT_INFERENCE_ENGINE_PLUGIN)
target_include_directories(${TARGET_NAME} PUBLIC ${PUBLIC_HEADERS_DIR})
target_include_directories(${TARGET_NAME} PRIVATE "${IE_MAIN_SOURCE_DIR}/src/inference_engine")
target_link_libraries(${TARGET_NAME} PUBLIC inference_engine_plugin_api ${NGRAPH_LIBRARIES} inference_engine)
target_link_libraries(${TARGET_NAME} PRIVATE pugixml)
# code style
add_cpplint_target(${TARGET_NAME}_cpplint FOR_TARGETS ${TARGET_NAME})
add_clang_format_target(${TARGET_NAME}_clang_format FOR_TARGETS ${TARGET_NAME})
# developer package
ie_developer_export_targets(${TARGET_NAME})
# install
install(TARGETS ${TARGET_NAME}
RUNTIME DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT core
ARCHIVE DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT core
LIBRARY DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT core)

0
inference-engine/src/inference_engine/ie_blob_proxy.hpp → inference-engine/src/ir_readers/ie_blob_proxy.hpp
View File

19
inference-engine/src/inference_engine/ie_cnn_net_reader_impl.cpp → inference-engine/src/ir_readers/ie_cnn_net_reader_impl.cpp
View File

@ -18,6 +18,7 @@
#include "ie_format_parser.h"
#include "ie_ir_reader.hpp"
#include "ie_profiling.hpp"
#include "ie_plugin.hpp"
#include "parsers.h"
#include "xml_parse_utils.h"
@ -29,21 +30,19 @@ IE_SUPPRESS_DEPRECATED_START
CNNNetReaderImpl::CNNNetReaderImpl(const FormatParserCreator::Ptr& _creator)
: parseSuccess(false), _version(0), parserCreator(_creator) {}
CNNNetReaderImpl::~CNNNetReaderImpl() { }
StatusCode CNNNetReaderImpl::SetWeights(const TBlob<uint8_t>::Ptr& weights, ResponseDesc* desc) noexcept {
if (!_parser && _version < 10) {
return DescriptionBuffer(desc) << "network must be read first";
}
try {
if (_version == 10) {
#if defined(ENABLE_IR_READER)
// It's time to perform actual reading of V10 network and instantiate CNNNetworkNGraphImpl
IRReader v10Reader(extensions);
std::stringstream model;
xmlDoc->save(model);
network = std::make_shared<CNNNetworkNGraphImpl>(v10Reader.read(model.str(), weights));
#else
return DescriptionBuffer(desc) << "Please, recompile Inference Engine with the ENABLE_IR_READER=ON Cmake option";
#endif
} else {
_parser->SetWeights(weights);
}
@ -173,15 +172,13 @@ void CNNNetReaderImpl::addExtensions(const std::vector<InferenceEngine::IExtensi
}
std::shared_ptr<IFormatParser> V2FormatParserCreator::create(size_t version) {
#ifdef ENABLE_IR_READER
return std::make_shared<FormatParser>(version);
#else
THROW_IE_EXCEPTION << "Please, recompile Inference Engine library with the ENABLE_IR_READER=ON Cmake option";
return nullptr;
#endif
}
InferenceEngine::ICNNNetReader* InferenceEngine::CreateCNNNetReader() noexcept {
return new CNNNetReaderImpl(std::make_shared<V2FormatParserCreator>());
INFERENCE_PLUGIN_API(InferenceEngine::StatusCode)
CreateICNNNetReader(ICNNNetReader *& data, ResponseDesc *resp) noexcept {
data = new CNNNetReaderImpl(std::make_shared<V2FormatParserCreator>());
return StatusCode::OK;
}
IE_SUPPRESS_DEPRECATED_END

13
inference-engine/src/inference_engine/ie_cnn_net_reader_impl.h → inference-engine/src/ir_readers/ie_cnn_net_reader_impl.h
View File

@ -11,10 +11,10 @@
#include <vector>
#include "cnn_network_impl.hpp"
#include "ie_icnn_net_reader.h"
#include "ie_memcpy.h"
#include "ie_profiling.hpp"
#include "parsers.h"
#include "ie_util_internal.hpp"
namespace pugi {
class xml_node;
@ -31,14 +31,14 @@ struct FormatParserCreator {
virtual ~FormatParserCreator() = default;
};
struct V2FormatParserCreator : public FormatParserCreator {
struct INFERENCE_ENGINE_API_CLASS(V2FormatParserCreator) : public FormatParserCreator {
std::shared_ptr<IFormatParser> create(size_t version) override;
};
IE_SUPPRESS_DEPRECATED_START
class CNNNetReaderImpl : public ICNNNetReader {
class INFERENCE_ENGINE_API_CLASS(CNNNetReaderImpl) : public ICNNNetReader {
public:
explicit CNNNetReaderImpl(const FormatParserCreator::Ptr& _parserCreator);
explicit CNNNetReaderImpl(const FormatParserCreator::Ptr& _creator);
StatusCode ReadNetwork(const char* filepath, ResponseDesc* resp) noexcept override;
@ -78,7 +78,9 @@ public:
delete this;
}
void addExtensions(const std::vector<InferenceEngine::IExtensionPtr>& ext);
void addExtensions(const std::vector<InferenceEngine::IExtensionPtr>& ext) override;
~CNNNetReaderImpl() override;
private:
std::shared_ptr<InferenceEngine::details::IFormatParser> _parser;
@ -96,6 +98,7 @@ private:
std::shared_ptr<pugi::xml_document> xmlDoc;
std::vector<InferenceEngine::IExtensionPtr> extensions;
};
IE_SUPPRESS_DEPRECATED_END
} // namespace details

5
inference-engine/src/inference_engine/ie_format_parser.cpp → inference-engine/src/ir_readers/ie_format_parser.cpp
View File

@ -267,9 +267,10 @@ FormatParser::FormatParser(size_t version): _version(version) {
std::make_shared<LayerCreator<TopKLayer>>("TopK"),
std::make_shared<LayerCreator<UniqueLayer>>("Unique"),
std::make_shared<LayerCreator<NonMaxSuppressionLayer>>("NonMaxSuppression"),
std::make_shared<LayerCreator<ScatterLayer>>("ScatterUpdate"),
std::make_shared<LayerCreator<ScatterUpdateLayer>>("ScatterUpdate"),
std::make_shared<LayerCreator<ExperimentalDetectronPriorGridGeneratorLayer>>("ExperimentalDetectronPriorGridGenerator"),
std::make_shared<LayerCreator<ExperimentalDetectronGenerateProposalsSingleImageLayer>>("ExperimentalDetectronGenerateProposalsSingleImage")};
std::make_shared<LayerCreator<ExperimentalDetectronGenerateProposalsSingleImageLayer>>("ExperimentalDetectronGenerateProposalsSingleImage"),
std::make_shared<LayerCreator<ExperimentalDetectronTopKROIs>>("ExperimentalDetectronTopKROIs")};
creators.emplace_back(_version < 6 ? std::make_shared<LayerCreator<QuantizeLayer>>("Quantize")
: std::make_shared<LayerCreator<QuantizeLayer>>("FakeQuantize"));
}

4
inference-engine/src/inference_engine/ie_format_parser.h → inference-engine/src/ir_readers/ie_format_parser.h
View File

@ -71,11 +71,7 @@ public:
}
};
#ifdef ENABLE_IR_READER
class INFERENCE_ENGINE_API_CLASS(FormatParser): public IFormatParser {
#else
class FormatParser : public IFormatParser {
#endif
public:
explicit FormatParser(size_t version);

6
inference-engine/src/inference_engine/ie_ir_parser.cpp → inference-engine/src/ir_readers/ie_ir_parser.cpp
View File

@ -187,7 +187,7 @@ V10Parser::GenericLayerParams V10Parser::parseGenericParams(const pugi::xml_node
// Input port hasn't precision
if (!input) {
const std::string& preStr = GetStrAttr(parentNode, "precision");
type = InferenceEngine::details::ngraph::convertPrecision(preStr);
type = InferenceEngine::details::convertPrecision(preStr);
}
port.precision = type;
return port;
@ -413,7 +413,7 @@ std::shared_ptr<ngraph::Node> V10Parser::createNode(const std::vector<ngraph::Ou
for (const auto& port : params.outputPorts) {
ngraph::op::GenericIE::PortIE iePort;
iePort.dims = port.dims;
iePort.precision = InferenceEngine::details::ngraph::convertPrecision(port.precision);
iePort.precision = InferenceEngine::details::convertPrecision(port.precision);
outputs.emplace_back(iePort);
}
@ -767,7 +767,7 @@ std::shared_ptr<ngraph::Node> V10Parser::LayerCreator<ngraph::op::Convert>::crea
THROW_IE_EXCEPTION << "Cannot read parameter for " << getType() << " layer with name: " << layerParsePrms.name;
return std::make_shared<ngraph::op::Convert>(inputs[0],
details::ngraph::convertPrecision(GetStrAttr(dn, "destination_type")));
details::convertPrecision(GetStrAttr(dn, "destination_type")));
}
// LSTMCell layer

2
inference-engine/src/inference_engine/ie_ir_parser.hpp → inference-engine/src/ir_readers/ie_ir_parser.hpp
View File

@ -192,7 +192,7 @@ private:
std::string val;
if (!getStrAttribute(node.child("data"), name, val)) return;
if (auto a = ngraph::as_type<ngraph::AttributeAdapter<ngraph::element::Type>>(&adapter)) {
static_cast<ngraph::element::Type&>(*a) = details::ngraph::convertPrecision(val);
static_cast<ngraph::element::Type&>(*a) = details::convertPrecision(val);
} else if (auto a = ngraph::as_type<ngraph::AttributeAdapter<ngraph::PartialShape>>(&adapter)) {
std::vector<int64_t> shape;
std::vector<ngraph::Dimension> dims;

0
inference-engine/src/inference_engine/ie_ir_reader.cpp → inference-engine/src/ir_readers/ie_ir_reader.cpp
View File

4
inference-engine/src/inference_engine/ie_ir_reader.hpp → inference-engine/src/ir_readers/ie_ir_reader.hpp
View File

@ -36,11 +36,7 @@ namespace InferenceEngine {
* All methods here do not throw exceptions and return a StatusCode and ResponseDesc object.
* Alternatively, to use methods that throw exceptions, refer to the CNNNetReader wrapper class.
*/
#ifdef ENABLE_IR_READER
class INFERENCE_ENGINE_API_CLASS(IRReader) {
#else
class IRReader {
#endif
public:
IRReader() = default;
explicit IRReader(const std::vector<IExtensionPtr>& exts): extensions(exts) {}

0
inference-engine/src/inference_engine/ie_layer_parsers.cpp → inference-engine/src/ir_readers/ie_layer_parsers.cpp
View File

0
inference-engine/src/inference_engine/ie_layer_parsers.h → inference-engine/src/ir_readers/ie_layer_parsers.h
View File

0
inference-engine/src/inference_engine/parsers.h → inference-engine/src/ir_readers/parsers.h
View File

9
inference-engine/src/legacy_api/CMakeLists.txt
View File

@ -32,6 +32,8 @@ set_ie_threading_interface_for(${TARGET_NAME}_obj)
target_compile_definitions(${TARGET_NAME}_obj PRIVATE IMPLEMENT_INFERENCE_ENGINE_API)
target_include_directories(${TARGET_NAME}_obj PRIVATE ${PUBLIC_HEADERS_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/src
${IE_MAIN_SOURCE_DIR}/src/inference_engine # For CNNNetworkNGraphImpl
$<TARGET_PROPERTY:inference_engine_transformations,INTERFACE_INCLUDE_DIRECTORIES>
$<TARGET_PROPERTY:inference_engine_plugin_api,INTERFACE_INCLUDE_DIRECTORIES>
$<TARGET_PROPERTY:ngraph::ngraph,INTERFACE_INCLUDE_DIRECTORIES>
$<TARGET_PROPERTY:pugixml,INTERFACE_INCLUDE_DIRECTORIES>)
@ -51,10 +53,9 @@ add_library(${TARGET_NAME} SHARED
set_ie_threading_interface_for(${TARGET_NAME})
target_link_libraries(${TARGET_NAME} PRIVATE ${NGRAPH_LIBRARIES} pugixml)
target_link_libraries(${TARGET_NAME} PRIVATE ${NGRAPH_LIBRARIES} inference_engine_transformations pugixml)
add_cpplint_target(${TARGET_NAME}_cpplint FOR_TARGETS ${TARGET_NAME})
add_clang_format_target(${TARGET_NAME}_clang_format FOR_TARGETS ${TARGET_NAME})
# export targets
@ -67,6 +68,6 @@ ie_developer_export_targets(${TARGET_NAME})
# install
install(TARGETS ${TARGET_NAME}
RUNTIME DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT core
ARCHIVE DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT core
RUNTIME DESTINATION ${IE_CPACK_RUNTIME_PATH} COMPONENT core
ARCHIVE DESTINATION ${IE_CPACK_ARCHIVE_PATH} COMPONENT core
LIBRARY DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT core)

4
inference-engine/src/legacy_api/include/cnn_network_impl.hpp
View File

@ -40,6 +40,10 @@ public:
precision = prec;
}
std::shared_ptr<::ngraph::Function> getFunction() noexcept override {
return nullptr;
}
std::shared_ptr<const ::ngraph::Function> getFunction() const noexcept override {
return nullptr;
}

9
inference-engine/src/plugin_api/convert_function_to_cnn_network.hpp → inference-engine/src/legacy_api/include/convert_function_to_cnn_network.hpp
View File

@ -4,13 +4,18 @@
#pragma once
#include "cnn_network_ngraph_impl.hpp"
#include "cnn_network_impl.hpp"
#include <ngraph/attribute_visitor.hpp>
#include <memory>
#include <string>
#include <vector>
namespace InferenceEngine {
namespace details {
INFERENCE_ENGINE_API_CPP(std::shared_ptr<CNNNetworkImpl>)
convertFunctionToICNNNetwork(const std::shared_ptr<const ::ngraph::Function>& graph, const CNNNetworkNGraphImpl &nGraphImpl);
convertFunctionToICNNNetwork(const std::shared_ptr<const ::ngraph::Function>& graph, const ICNNNetwork &network);
} // namespace details
} // namespace InferenceEngine

1
inference-engine/src/legacy_api/include/graph_transformer.h
View File

@ -24,6 +24,7 @@ namespace InferenceEngine {
*/
class INFERENCE_ENGINE_API_CLASS(ConstTransformer) {
public:
explicit ConstTransformer(ICNNNetwork* _network);
explicit ConstTransformer(details::CNNNetworkImpl* _network);
explicit ConstTransformer(std::vector<DataPtr> &_inputs, std::vector<DataPtr> &_outputs);

8
inference-engine/src/inference_engine/ie_ngraph_utils.hpp → inference-engine/src/legacy_api/include/ie_ngraph_utils.hpp
View File

@ -11,7 +11,6 @@
namespace InferenceEngine {
namespace details {
namespace ngraph {
inline ::ngraph::element::Type convertPrecision(const Precision& precision) {
Precision::ePrecision pType = precision;
@ -22,6 +21,8 @@ inline ::ngraph::element::Type convertPrecision(const Precision& precision) {
return ::ngraph::element::Type(::ngraph::element::Type_t::f32);
case Precision::FP16:
return ::ngraph::element::Type(::ngraph::element::Type_t::f16);
case Precision::BF16:
return ::ngraph::element::Type(::ngraph::element::Type_t::bf16);
case Precision::U8:
return ::ngraph::element::Type(::ngraph::element::Type_t::u8);
case Precision::I8:
@ -53,6 +54,8 @@ inline ::ngraph::element::Type convertPrecision(const std::string& precision) {
return ::ngraph::element::Type(::ngraph::element::Type_t::f16);
} else if (precision == "f32" || precision == "FP32") {
return ::ngraph::element::Type(::ngraph::element::Type_t::f32);
} else if (precision == "bf16" || precision == "BF16") {
return ::ngraph::element::Type(::ngraph::element::Type_t::bf16);
} else if (precision == "f64" || precision == "FP64") {
return ::ngraph::element::Type(::ngraph::element::Type_t::f64);
} else if (precision == "i8" || precision == "I8") {
@ -90,6 +93,8 @@ inline Precision convertPrecision(const ::ngraph::element::Type& precision) {
return Precision(Precision::FP16);
case ::ngraph::element::Type_t::f32:
return Precision(Precision::FP32);
case ::ngraph::element::Type_t::bf16:
return Precision(Precision::BF16);
case ::ngraph::element::Type_t::i8:
return Precision(Precision::I8);
case ::ngraph::element::Type_t::i16:
@ -113,6 +118,5 @@ inline Precision convertPrecision(const ::ngraph::element::Type& precision) {
}
}
} // namespace ngraph
} // namespace details
} // namespace InferenceEngine

10
inference-engine/src/legacy_api/include/ie_util_internal.hpp
View File

@ -6,6 +6,7 @@
#include <cpp/ie_cnn_network.h>
#include <ie_icnn_network.hpp>
#include <cnn_network_impl.hpp>
#include <file_utils.h>
#include <deque>
@ -49,6 +50,15 @@ cloneNet(const std::vector<InferenceEngine::CNNLayerPtr>& layers, const ICNNNetw
IE_SUPPRESS_DEPRECATED_END
/**
* @brief Clones the whole network without conversion to CNNNetworkImpl. All layers and data objects will be cloned
* @note Blobs inside layers are reused
* @param network A network to clone
* @return A cloned object
*/
INFERENCE_ENGINE_API_CPP(std::shared_ptr<InferenceEngine::ICNNNetwork>)
cloneNetwork(const InferenceEngine::ICNNNetwork& network);
/**
* @brief Clones the whole network. All layers and data objects will be cloned
* @note Blobs inside layers are reused

4
inference-engine/src/legacy_api/include/layer_transform.hpp
View File

@ -33,8 +33,8 @@ using AllLayers =
ReshapeLayer*, TileLayer*, ScaleShiftLayer*, PReLULayer*, PowerLayer*, BatchNormalizationLayer*,
ClampLayer*, TensorIterator*, LSTMCell*, GRUCell*, RNNCell*, RNNSequenceLayer*, QuantizeLayer*,
BinaryConvolutionLayer*, WeightableLayer*, OneHotLayer*, MathLayer*, ReduceLayer*, UniqueLayer*,
NonMaxSuppressionLayer*, ScatterLayer*, ExperimentalDetectronPriorGridGeneratorLayer*,
ExperimentalDetectronGenerateProposalsSingleImageLayer*, CNNLayer*>;
NonMaxSuppressionLayer*, ScatterUpdateLayer*, ExperimentalDetectronPriorGridGeneratorLayer*,
ExperimentalDetectronGenerateProposalsSingleImageLayer*, ExperimentalDetectronTopKROIs*, CNNLayer*>;
template <class Visitor, std::size_t I = 0, typename... Tp>
inline typename std::enable_if<I == sizeof...(Tp), void>::type visitActualLayer(std::tuple<Tp...>&& t,

4
inference-engine/src/legacy_api/src/cnn_network_int8_normalizer.cpp
View File

@ -1000,7 +1000,9 @@ void CNNNetworkInt8Normalizer::QuantizeConvolutionOrFullyConnected(CNNLayer::Ptr
}
prev = *it;
}
symQuant = *(intervals.begin());
if (!intervals.empty()) {
symQuant = *(intervals.begin());
}
std::set<double> divs;
prev = 0.f;
for (auto it = individualsG.begin(); it != individualsG.end(); it++) {

329
inference-engine/src/inference_engine/convert_function_to_cnn_network.cpp → inference-engine/src/legacy_api/src/convert_function_to_cnn_network.cpp
View File

@ -9,6 +9,7 @@
#include <vector>
#include <unordered_set>
#include <cnn_network_ngraph_impl.hpp>
#include "ngraph_ops/convolution_ie.hpp"
#include "ngraph_ops/deconvolution_ie.hpp"
#include "ngraph_ops/eltwise.hpp"
@ -39,12 +40,318 @@
#include "ie_profiling.hpp"
#include "ie_cnn_layer_builder_ngraph.h"
#include <debug.h>
#include "transformations/convert_opset1_to_legacy/convert_opset1_to_legacy.hpp"
#include "transformations/utils/utils.hpp"
namespace InferenceEngine {
namespace details {
std::shared_ptr<CNNNetworkImpl> convertFunctionToICNNNetwork(const std::shared_ptr<const ::ngraph::Function>& graph, const CNNNetworkNGraphImpl &nGraphImpl) {
/**
* @brief Creator for CNNLayer from nGraph op
*/
class CNNLayerCreator : public ::ngraph::AttributeVisitor {
public:
using CreatorFor = std::function<CNNLayerPtr(const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> param)>;
explicit CNNLayerCreator(const std::shared_ptr<::ngraph::Node>& node);
CNNLayerPtr create();
void on_attribute(const std::string& name, std::string& value) override {
params[name] = value;
}
void on_attribute(const std::string& name, bool& value) override {
params[name] = value ? "true" : "false";
}
void addSpecificCreator(const std::vector<std::string>& forTypes, const CreatorFor& creator) {
for (const auto type : forTypes) {
creators[type] = creator;
}
}
void on_adapter(const std::string& name, ::ngraph::ValueAccessor<std::string>& adapter) override {
std::string data = adapter.get();
std::transform(data.begin(), data.end(), data.begin(), [](unsigned char c) {
return std::tolower(c);
});
params[name] = data;
}
void on_adapter(const std::string& name, ::ngraph::ValueAccessor<std::vector<int64_t>>& adapter) override {
auto shape = adapter.get();
params[name] = joinVec(shape);
}
void on_adapter(const std::string& name, ::ngraph::ValueAccessor<double>& adapter) override {
params[name] = std::to_string(adapter.get());
}
void on_adapter(const std::string& name, ::ngraph::ValueAccessor<int64_t>& adapter) override {
params[name] = std::to_string(adapter.get());
}
void on_adapter(const std::string& name, ::ngraph::ValueAccessor<void>& adapter) override;
private:
std::shared_ptr<::ngraph::Node> node;
std::map<std::string, std::string> params;
std::map<std::string, CreatorFor> creators;
};
void InferenceEngine::details::CNNLayerCreator::on_adapter(const std::string& name,
::ngraph::ValueAccessor<void>& adapter) {
if (auto a = ::ngraph::as_type<::ngraph::AttributeAdapter<::ngraph::element::Type>>(&adapter)) {
auto type = static_cast<::ngraph::element::Type&>(*a);
params[name] = details::convertPrecision(type).name();
} else if (auto a = ::ngraph::as_type<::ngraph::AttributeAdapter<::ngraph::PartialShape>>(&adapter)) {
std::string dims;
auto shape = static_cast<::ngraph::PartialShape&>(*a);
for (size_t i = 0; i < shape.rank().get_length(); i++) {
if (!dims.empty()) dims += ",";
dims += std::to_string(shape[i].get_length());
}
params[name] = dims;
} else if (auto a = ::ngraph::as_type<::ngraph::AttributeAdapter<::ngraph::Shape>>(&adapter)) {
auto shape = static_cast<::ngraph::Shape&>(*a);
params[name] = joinVec(shape);
} else if (auto a = ::ngraph::as_type<::ngraph::AttributeAdapter<::ngraph::Strides>>(&adapter)) {
auto shape = static_cast<::ngraph::Strides&>(*a);
params[name] = joinVec(shape);
}
}
InferenceEngine::details::CNNLayerCreator::CNNLayerCreator(const std::shared_ptr<::ngraph::Node>& node): node(node) {
addSpecificCreator({"Parameter"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), "Input",
details::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<CNNLayer>(attrs);
return res;
});
// TODO - Remove "GreaterEq" once ngraph transitions to GreaterEqual
addSpecificCreator({"Eltwise", "Subtract", "Power", "Maximum", "Divide", "Greater", "GreaterEqual", "FloorMod", "LogicalOr", "LogicalAnd", "LogicalXor",
"GreaterEq", "Less", "LessEqual", "Equal", "NotEqual", "Multiply", "Add"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), "Eltwise",
details::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<EltwiseLayer>(attrs);
res->params = params;
if (node->description() == "Maximum") {
res->params["operation"] = "max";
} else if (node->description() == "Power") {
res->params["operation"] = "pow";
} else if (node->description() == "Subtract") {
res->params["operation"] = "sub";
} else if (node->description() == "Divide") {
res->params["operation"] = "div";
} else if (node->description() == "LessEqual") {
res->params["operation"] = "less_equal";
} else if (node->description() == "Less") {
res->params["operation"] = "less";
} else if (node->description() == "Equal") {
res->params["operation"] = "equal";
} else if (node->description() == "NotEqual") {
res->params["operation"] = "not_equal";
} else if (node->description() == "FloorMod") {
res->params["operation"] = "floor_mod";
} else if (node->description() == "Multiply") {
res->params["operation"] = "prod";
} else if (node->description() == "Add") {
res->params["operation"] = "sum";
} else if (node->description() == "Greater") {
res->params["operation"] = "greater";
} else if (node->description() == "GreaterEq") {
res->params["operation"] = "greater_equal";
} else if (node->description() == "GreaterEqual") {
res->params["operation"] = "greater_equal";
} else if (node->description() == "LogicalOr") {
res->params["operation"] = "logical_or";
} else if (node->description() == "LogicalAnd") {
res->params["operation"] = "logical_and";
} else if (node->description() == "LogicalXor") {
res->params["operation"] = "logical_xor";
} else if (node->description() == "Eltwise") {
auto castedLayer = std::dynamic_pointer_cast<::ngraph::op::Eltwise>(node);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << attrs.type << " layer " << attrs.name;
std::string type;
switch (castedLayer->eltwise_type) {
case ELTWISE_TYPE::Sum:
type = "sum";
break;
case ELTWISE_TYPE::Prod:
type = "prod";
break;
default:
THROW_IE_EXCEPTION << "Not supported eltwise type!";
}
res->params["operation"] = type;
}
return res;
});
addSpecificCreator({"Concat"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), node->description(),
details::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<ConcatLayer>(attrs);
res->params = params;
return res;
});
addSpecificCreator({"AvgPool", "MaxPool"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), "Pooling",
details::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<PoolingLayer>(attrs);
res->params = params;
if (res->params.find("auto_pad") != res->params.end() &&
details::CaselessEq<std::string>()(res->params["auto_pad"], "EXPLICIT"))
res->params.erase("auto_pad");
if (res->params.find("exclude_pad") != res->params.end()) {
res->params["exclude-pad"] = res->params["exclude_pad"];
res->params.erase("exclude_pad");
}
if (node->description() == "MaxPool") {
res->params["pool-method"] = "max";
} else if (node->description() == "AvgPool") {
res->params["pool-method"] = "avg";
}
return res;
});
addSpecificCreator({"Select"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), node->description(),
details::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<SelectLayer>(attrs);
res->params = params;
return res;
});
addSpecificCreator({"BinaryConvolution"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), node->description(),
details::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<BinaryConvolutionLayer>(attrs);
// todo: investigate difference between ngraph parameters for BinConvolution and the implementation above
// this leads to accuracy issue for Precollected_ONNX_ResNet50_88percentinto1bit e2e test
// res->params = params;
auto castedLayer = ::ngraph::as_type_ptr<::ngraph::op::v1::BinaryConvolution>(node);
std::string value;
for (const auto& val : castedLayer->get_pads_begin()) {
if (!value.empty()) value += ",";
value += Builder::asString(val);
}
res->params["pads_begin"] = value;
value.clear();
for (const auto& val : castedLayer->get_pads_end()) {
if (!value.empty()) value += ",";
value += Builder::asString(val);
}
res->params["pads_end"] = value;
switch (castedLayer->get_auto_pad()) {
case ::ngraph::op::PadType::SAME_UPPER:
res->params["auto_pad"] = "same_upper";
break;
case ::ngraph::op::PadType::SAME_LOWER:
res->params["auto_pad"] = "same_lower";
break;
case ::ngraph::op::PadType::VALID:
res->params["auto_pad"] = "valid";
break;
default:
break;
}
value.clear();
for (const auto& val : castedLayer->get_strides()) {
if (!value.empty()) value += ",";
value += Builder::asString(val);
}
res->params["strides"] = value;
value.clear();
for (const auto& val : castedLayer->get_dilations()) {
if (!value.empty()) value += ",";
value += Builder::asString(val);
}
res->params["dilations"] = value;
// Restore kernel size and output
const auto& shape = castedLayer->get_input_shape(1);
res->params["output"] = Builder::asString(shape[0]);
value.clear();
for (size_t i = 2; i < shape.size(); i++) {
if (!value.empty()) value += ",";
value += Builder::asString(shape[i]);
}
res->params["kernel"] = value;
switch (castedLayer->get_mode()) {
case ::ngraph::op::v1::BinaryConvolution::BinaryConvolutionMode::XNOR_POPCOUNT:
res->params["mode"] = "xnor-popcount";
}
auto weights_shape = castedLayer->input(1).get_source_output().get_shape();
res->params["input"] = Builder::asString(weights_shape[1]);
res->params["pad_value"] = Builder::asString(castedLayer->get_pad_value());
Builder::NodeConverter<::ngraph::op::Constant> converter;
const auto weightsNode = castedLayer->get_inputs()[1].get_output().get_node();
if (converter.canCreate(weightsNode)) {
const auto& weights = converter.createLayer(weightsNode);
res->blobs["weights"] = weights->blobs["custom"];
res->_weights = weights->blobs["custom"];
}
return res;
});
addSpecificCreator({"SpaceToBatch"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), node->description(),
details::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<SpaceToBatchLayer>(attrs);
res->params = params;
return res;
});
addSpecificCreator({"BatchToSpace"}, [](const std::shared_ptr<::ngraph::Node>& node,
const std::map<std::string, std::string> params) -> CNNLayerPtr {
LayerParams attrs = {node->get_friendly_name(), node->description(),
details::convertPrecision(node->get_output_element_type(0))};
auto res = std::make_shared<BatchToSpaceLayer>(attrs);
res->params = params;
return res;
});
}
CNNLayerPtr InferenceEngine::details::CNNLayerCreator::create() {
auto one_from = [](const std::string& desc, const std::vector<std::string>& descs) -> bool {
for (const auto& d : descs) {
if (details::CaselessEq<std::string>()(d, desc)) return true;
}
return false;
};
LayerParams attrs = {node->get_friendly_name(), node->description(),
details::convertPrecision(node->get_output_element_type(0))};
if (creators.find(node->description()) != creators.end())
return creators[node->description()](node, params);
auto res = std::make_shared<CNNLayer>(attrs);
res->params = params;
return res;
}
std::shared_ptr<CNNNetworkImpl> convertFunctionToICNNNetwork(const std::shared_ptr<const ::ngraph::Function>& graph, const ICNNNetwork &network) {
IE_PROFILING_AUTO_SCOPE(convertFunctionToICNNNetwork)
const auto createCNNLayer = [](const std::shared_ptr<::ngraph::Node> &node) -> CNNLayerPtr {
class NGraphCNNLayer: public CNNLayer {
@ -240,8 +547,10 @@ std::shared_ptr<CNNNetworkImpl> convertFunctionToICNNNetwork(const std::shared_p
network->setInputInfo(info);
};
const CNNNetworkNGraphImpl* nGraphImpl = dynamic_cast<const CNNNetworkNGraphImpl*>(&network);
InputsDataMap thisInputDataMap;
nGraphImpl.getInputsInfo(thisInputDataMap);
network.getInputsInfo(thisInputDataMap);
// Create network
auto cnnNetworkImpl = std::make_shared<details::CNNNetworkImpl>();
@ -295,25 +604,25 @@ std::shared_ptr<CNNNetworkImpl> convertFunctionToICNNNetwork(const std::shared_p
for (const auto &dim : dims) {
if (!dim)
THROW_IE_EXCEPTION << cnnLayer->type << " layer " << cnnLayer->name
<< " has incorrect dimensions in the output data " << i;
<< " has incorrect dimensions in the output data " << i;
}
if (!ptr && nGraphImpl._data.find(outName) != nGraphImpl._data.end()) {
ptr = nGraphImpl._data.at(outName);
if (!ptr && nGraphImpl && nGraphImpl->_data.find(outName) != nGraphImpl->_data.end()) {
ptr = nGraphImpl->_data.at(outName);
if (auto nData = std::dynamic_pointer_cast<InferenceEngine::details::NGraphData>(ptr)) {
const auto layout =
dims.size() == nData->getTensorDesc().getDims().size() ?
nData->getTensorDesc().getLayout() :
TensorDesc::getLayoutByDims(dims);
dims.size() == nData->getTensorDesc().getDims().size() ?
nData->getTensorDesc().getLayout() :
TensorDesc::getLayoutByDims(dims);
nData->reset();
nData->reshape(dims, layout);
}
cnnNetworkImpl->addData(outName.c_str(), ptr);
}
if (!ptr) {
ptr.reset(new Data(outName,
{details::ngraph::convertPrecision(layer->get_output_element_type(i)), dims,
{details::convertPrecision(layer->get_output_element_type(i)), dims,
TensorDesc::getLayoutByDims(dims)}));
}

14
inference-engine/src/legacy_api/src/graph_transformer.cpp
View File

@ -17,6 +17,7 @@
#include <vector>
#include <mutex>
#include <cnn_network_ngraph_impl.hpp>
#include "blob_factory.hpp"
#include "cnn_network_impl.hpp"
#include "graph_tools.hpp"
@ -70,6 +71,19 @@ ConstTransformer::ConstTransformer(details::CNNNetworkImpl* _network)
THROW_IE_EXCEPTION << "[ERROR]: Failed to init ConstTransformer with null pointer of network";
}
ConstTransformer::ConstTransformer(ICNNNetwork* _network) {
if (auto cnnNet = dynamic_cast<InferenceEngine::details::CNNNetworkImpl *>(_network)) {
network = cnnNet;
} else if (auto nGraphNet = dynamic_cast<InferenceEngine::details::CNNNetworkNGraphImpl *>(_network)) {
if (auto cnnNet = dynamic_cast<InferenceEngine::details::CNNNetworkImpl *>(nGraphNet->getCNNNetwork().get()))
network = cnnNet;
}
if (!network)
THROW_IE_EXCEPTION << "[ERROR]: Failed to init ConstTransformer with unsupported network type";
inputs = get_inputs(network);
outputs = get_outputs(network);
}
ConstTransformer::ConstTransformer(std::vector<DataPtr> &_inputs, std::vector<DataPtr> &_outputs)
: inputs(_inputs), outputs(_outputs), network(nullptr) {
if (inputs.empty() || outputs.empty())

192
inference-engine/src/inference_engine/ie_cnn_layer_builder_ngraph.cpp → inference-engine/src/legacy_api/src/ie_cnn_layer_builder_ngraph.cpp
View File

@ -3,7 +3,7 @@
//
#include <ie_cnn_layer_builder_ngraph.h>
#include "cnn_network_ngraph_impl.hpp"
#include <cnn_network_ngraph_impl.hpp>
#include <precision_utils.h>
#include <cpp/ie_cnn_network.h>
@ -72,7 +72,7 @@ std::string asString<float>(const float& value) {
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Abs>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Abs",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -83,7 +83,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::GenericIE>::createLayer(const std::share
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get layer " << layer->get_friendly_name();
LayerParams params = {layer->get_friendly_name(), castedLayer->getType(),
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
if (castedLayer->getType() == "RNNCell")
res = std::make_shared<InferenceEngine::RNNCell>(params);
@ -232,7 +232,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::TensorIterator>::createLayer(const std::
// Create Inference Engine representation of TensorIterator
LayerParams params = {layer->get_friendly_name(), "TensorIterator",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::TensorIterator>(params);
// Body: inputs
@ -365,7 +365,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::TensorIterator>::createLayer(const std::
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Constant>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Const",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::Constant>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -377,9 +377,9 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Constant>::createLayer(const std::shared
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Convert>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Convert",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto p = details::ngraph::convertPrecision(layer->get_output_element_type(0));
auto p = details::convertPrecision(layer->get_output_element_type(0));
std::string precision_str;
switch (p) {
case Precision::FP16:
@ -423,7 +423,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Convert>::createLayer(const std::shared_
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Ceiling>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Ceiling",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -431,7 +431,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Ceiling>::createLayer(const std::shared_
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Floor>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Floor",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -439,7 +439,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Floor>::createLayer(const std::shared_pt
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Sigmoid>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Sigmoid",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -447,7 +447,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Sigmoid>::createLayer(const std::shared_
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Tanh>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "TanH",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -455,7 +455,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Tanh>::createLayer(const std::shared_ptr
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Relu>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "ReLU",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ReLULayer>(params);
return res;
}
@ -463,7 +463,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Relu>::createLayer(const std::shared_ptr
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::SeluIE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Selu",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::SeluIE>(layer);
@ -477,7 +477,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::SeluIE>::createLayer(const std::shared_p
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::ReLUIE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "ReLU",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ReLULayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::ReLUIE>(layer);
@ -490,7 +490,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::ReLUIE>::createLayer(const std::shared_p
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Range>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Range",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -498,7 +498,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Range>::createLayer(const std::shared_pt
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Exp>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Exp",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -506,7 +506,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Exp>::createLayer(const std::shared_ptr<
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::MVN>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "MVN",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::MVNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::MVN>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -529,7 +529,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::LRN>::createLayer(const std::shared_ptr<
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::LRN_IE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Norm",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::NormLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::LRN_IE>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -545,7 +545,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::LRN_IE>::createLayer(const std::shared_p
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::CropIE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Crop",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CropLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::CropIE>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -577,7 +577,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::CropIE>::createLayer(const std::shared_p
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Clamp>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Clamp",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ClampLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::Clamp>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -590,7 +590,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Clamp>::createLayer(const std::shared_pt
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::Softmax>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "SoftMax",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::SoftMaxLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::v1::Softmax>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -602,7 +602,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::Softmax>::createLayer(const std::sha
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Subtract>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Eltwise",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::EltwiseLayer>(params);
res->params["operation"] = "sub";
return res;
@ -611,7 +611,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Subtract>::createLayer(const std::shared
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::Power>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Eltwise",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::EltwiseLayer>(params);
res->params["operation"] = "pow";
return res;
@ -620,7 +620,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::Power>::createLayer(const std::share
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::Maximum>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Eltwise",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::EltwiseLayer>(params);
res->params["operation"] = "max";
return res;
@ -634,7 +634,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::Minimum>::createLayer(const std::sha
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::Divide>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Eltwise",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::EltwiseLayer>(params);
res->params["operation"] = "div";
return res;
@ -643,7 +643,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::Divide>::createLayer(const std::shar
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::Multiply>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Eltwise",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::EltwiseLayer>(params);
res->params["operation"] = "prod";
return res;
@ -652,7 +652,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::Multiply>::createLayer(const std::sh
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::Add>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Eltwise",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::EltwiseLayer>(params);
res->params["operation"] = "sum";
return res;
@ -673,7 +673,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::BatchNormInference>::createLayer(
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Squeeze>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Squeeze",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::Squeeze>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -684,7 +684,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Squeeze>::createLayer(const std::shared_
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Unsqueeze>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Unsqueeze",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::Unsqueeze>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -695,7 +695,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Unsqueeze>::createLayer(const std::share
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::FakeQuantize>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "FakeQuantize",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::QuantizeLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::FakeQuantize>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -707,7 +707,7 @@ template <>
CNNLayer::Ptr NodeConverter<ngraph::op::ConvolutionIE>::createLayer(
const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Convolution",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ConvolutionLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::ConvolutionIE>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -793,7 +793,7 @@ template <>
CNNLayer::Ptr NodeConverter<ngraph::op::DeconvolutionIE>::createLayer(
const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Deconvolution",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::DeconvolutionLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::DeconvolutionIE>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -861,7 +861,7 @@ template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::DeformableConvolution>::createLayer(
const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "DeformableConvolution",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::DeformableConvolutionLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::v1::DeformableConvolution>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -935,7 +935,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::DeformableConvolution>::createLayer(
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::AvgPool>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Pooling",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::PoolingLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::v1::AvgPool>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1001,7 +1001,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::AvgPool>::createLayer(const std::sha
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::MaxPool>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Pooling",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::PoolingLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::v1::MaxPool>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1066,7 +1066,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::MaxPool>::createLayer(const std::sha
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::ROIPooling>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "ROIPooling",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::ROIPooling>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1082,7 +1082,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::ROIPooling>::createLayer(const std::shar
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::PSROIPooling>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "PSROIPooling",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::PSROIPooling>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1101,7 +1101,7 @@ template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::DeformablePSROIPooling>::createLayer(
const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "PSROIPooling",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::v1::DeformablePSROIPooling>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1125,7 +1125,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::DeformablePSROIPooling>::createLayer
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::PRelu>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "PReLU",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::PReLULayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::PRelu>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1154,7 +1154,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::PRelu>::createLayer(const std::shared_pt
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::Split>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Split",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::SplitLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::v1::Split>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1175,7 +1175,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::Split>::createLayer(const std::share
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::VariadicSplit>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Split",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::SplitLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::VariadicSplit>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1196,7 +1196,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::VariadicSplit>::createLayer(const std::s
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Concat>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Concat",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ConcatLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::Concat>(layer);
@ -1210,7 +1210,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Concat>::createLayer(const std::shared_p
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::GatherIE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Gather",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::GatherLayer>(params);
auto castedLayer = std::dynamic_pointer_cast<ngraph::op::GatherIE>(layer);
@ -1229,14 +1229,14 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::GatherTree>::createLayer(const std::
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::GatherTreeIE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "GatherTree",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::ReverseSequence>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "ReverseSequence", details::ngraph::convertPrecision(layer->get_output_element_type(0))};
LayerParams params = {layer->get_friendly_name(), "ReverseSequence", details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ReverseSequenceLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::ReverseSequence>(layer);
@ -1252,7 +1252,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::ReverseSequence>::createLayer(const std:
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Reshape>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Reshape",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ReshapeLayer>(params);
return res;
}
@ -1260,7 +1260,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Reshape>::createLayer(const std::shared_
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::ShapeOf>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "ShapeOf",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -1268,7 +1268,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::ShapeOf>::createLayer(const std::shared_
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::Reshape>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Reshape",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto castedLayer = ngraph::as_type_ptr<ngraph::op::v1::Reshape>(layer);
if (castedLayer == nullptr)
@ -1293,7 +1293,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::Reshape>::createLayer(const std::sha
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::PadIE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Pad",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::PadLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::PadIE>(layer);
@ -1333,7 +1333,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::PadIE>::createLayer(const std::shared_pt
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::ScaleShiftIE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "ScaleShift",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ScaleShiftLayer>(params);
NodeConverter<ngraph::op::Constant> converter;
@ -1357,7 +1357,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::ScaleShiftIE>::createLayer(const std::sh
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Elu>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "elu",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::Elu>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1370,7 +1370,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Elu>::createLayer(const std::shared_ptr<
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::SquaredDifference>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Eltwise",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::EltwiseLayer>(params);
res->params["operation"] = "squared_diff";
return res;
@ -1380,7 +1380,7 @@ template <>
CNNLayer::Ptr NodeConverter<ngraph::op::DetectionOutput>::createLayer(
const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "DetectionOutput",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::DetectionOutput>(layer);
@ -1416,7 +1416,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::DetectionOutput>::createLayer(
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Transpose>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Permute",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
NodeConverter<ngraph::op::Constant> converter;
@ -1444,7 +1444,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Proposal>::createLayer(const std::shared
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::ProposalIE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Proposal",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::ProposalIE>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1484,7 +1484,7 @@ template <>
CNNLayer::Ptr NodeConverter<ngraph::op::PriorBoxClusteredIE>::createLayer(
const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "PriorBoxClustered",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::PriorBoxClusteredIE>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1533,7 +1533,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::PriorBoxClustered>::createLayer(
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::PriorBoxIE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "PriorBox",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::PriorBoxIE>(layer);
auto layer_info = params.type + " layer " + params.name;
@ -1613,7 +1613,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::PriorBox>::createLayer(const std::shared
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::PowerIE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Power",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::PowerLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::PowerIE>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1628,7 +1628,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::PowerIE>::createLayer(const std::shared_
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::TopK>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "TopK",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::TopKLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::v1::TopK>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1672,7 +1672,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::TopK>::createLayer(const std::shared
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::TopKIE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "TopK",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::TopKLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::TopKIE>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1687,7 +1687,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::TopKIE>::createLayer(const std::shared_p
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Eltwise>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Eltwise",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::EltwiseLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::Eltwise>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1712,7 +1712,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Eltwise>::createLayer(const std::shared_
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::TileIE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Tile",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::TileLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::TileIE>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1725,7 +1725,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::TileIE>::createLayer(const std::shared_p
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::ResampleV2>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Resample", details::ngraph::convertPrecision(layer->get_output_element_type(0))};
LayerParams params = {layer->get_friendly_name(), "Resample", details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::ResampleV2>(layer);
if (castedLayer == nullptr)
@ -1752,7 +1752,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::ResampleV2>::createLayer(const std::shar
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Interp>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Resample",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto castedLayer = ngraph::as_type_ptr<ngraph::op::Interp>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1766,7 +1766,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Interp>::createLayer(const std::shared_p
}
params = {layer->get_friendly_name(), "Interp",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
res->params["height"] = asString(attrs.height);
@ -1786,7 +1786,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Interpolate>::createLayer(const std::sha
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::FullyConnected>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "FullyConnected",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto castedLayer = ngraph::as_type_ptr<ngraph::op::FullyConnected>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1824,7 +1824,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::LSTMCell>::createLayer(const std::shared
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::LSTMCellIE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "LSTMCell",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto castedLayer = ngraph::as_type_ptr<ngraph::op::LSTMCellIE>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1872,7 +1872,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::LSTMCellIE>::createLayer(const std::shar
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::MatMul>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Gemm",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto castedLayer = ngraph::as_type_ptr<ngraph::op::MatMul>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1887,7 +1887,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::MatMul>::createLayer(const std::shared_p
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::RegionYolo>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "RegionYolo",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::RegionYolo>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1920,7 +1920,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::RegionYolo>::createLayer(const std::shar
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::ReorgYolo>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "ReorgYolo",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::ReorgYolo>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1938,7 +1938,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::ReorgYolo>::createLayer(const std::share
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::ReduceMin>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "ReduceMin",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ReduceLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::v1::ReduceMin>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1950,7 +1950,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::ReduceMin>::createLayer(const std::s
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::ReduceMax>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "ReduceMax",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ReduceLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::v1::ReduceMax>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1962,7 +1962,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::ReduceMax>::createLayer(const std::s
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::ReduceMean>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "ReduceMean",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ReduceLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::v1::ReduceMean>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1974,7 +1974,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::ReduceMean>::createLayer(const std::
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::ReduceProd>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "ReduceProd",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ReduceLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::v1::ReduceProd>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -1986,7 +1986,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::ReduceProd>::createLayer(const std::
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::ReduceSum>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "ReduceSum",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ReduceLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::v1::ReduceSum>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -2003,7 +2003,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::NormalizeL2>::createLayer(const std::sha
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Log>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Log",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -2011,7 +2011,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Log>::createLayer(const std::shared_ptr<
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::NormalizeIE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Normalize",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::NormLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::NormalizeIE>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -2025,6 +2025,8 @@ CNNLayer::Ptr NodeConverter<ngraph::op::NormalizeIE>::createLayer(const std::sha
if (converter.canCreate(weightsNode)) {
const auto& weights = converter.createLayer(weightsNode);
res->blobs["weights"] = weights->blobs["custom"];
} else {
THROW_IE_EXCEPTION << "Cannot convert weight node for NormalizeIE op";
}
return res;
@ -2034,7 +2036,7 @@ template <>
CNNLayer::Ptr NodeConverter<ngraph::op::CTCGreedyDecoder>::createLayer(
const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "CTCGreedyDecoder",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = ngraph::as_type_ptr<ngraph::op::CTCGreedyDecoder>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -2046,7 +2048,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::CTCGreedyDecoder>::createLayer(
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Erf>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Erf",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -2054,7 +2056,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Erf>::createLayer(const std::shared_ptr<
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Sign>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Sign",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -2062,7 +2064,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Sign>::createLayer(const std::shared_ptr
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Sin>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Sin",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -2070,7 +2072,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Sin>::createLayer(const std::shared_ptr<
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Sinh>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Sinh",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -2078,7 +2080,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Sinh>::createLayer(const std::shared_ptr
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Asin>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Asin",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -2086,7 +2088,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Asin>::createLayer(const std::shared_ptr
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Cos>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Cos",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -2094,7 +2096,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Cos>::createLayer(const std::shared_ptr<
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Cosh>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Cosh",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -2102,7 +2104,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Cosh>::createLayer(const std::shared_ptr
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Acos>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Acos",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -2110,7 +2112,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Acos>::createLayer(const std::shared_ptr
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Tan>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Tan",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -2118,7 +2120,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Tan>::createLayer(const std::shared_ptr<
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Atan>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Atan",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -2126,7 +2128,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::Atan>::createLayer(const std::shared_ptr
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::Sqrt>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Sqrt",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
return res;
}
@ -2142,7 +2144,7 @@ template <>
CNNLayer::Ptr NodeConverter<ngraph::op::StridedSliceIE>::createLayer(
const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "StridedSlice",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::StridedSliceLayer>(params);
auto castedLayer = std::dynamic_pointer_cast<ngraph::op::StridedSliceIE>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -2209,7 +2211,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::HardSigmoid>::createLayer(const std::sha
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::HardSigmoid_IE>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = { layer->get_friendly_name(), "HardSigmoid", details::ngraph::convertPrecision(layer->get_output_element_type(0)) };
LayerParams params = { layer->get_friendly_name(), "HardSigmoid", details::convertPrecision(layer->get_output_element_type(0)) };
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
auto castedLayer = std::dynamic_pointer_cast<ngraph::op::HardSigmoid_IE>(layer);
if (castedLayer == nullptr)
@ -2223,7 +2225,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::HardSigmoid_IE>::createLayer(const std::
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::GRN>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "GRN",
details::ngraph::convertPrecision(layer->get_output_element_type(0))};
details::convertPrecision(layer->get_output_element_type(0))};
auto castedLayer = std::dynamic_pointer_cast<ngraph::op::GRN>(layer);
if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
@ -2234,7 +2236,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::GRN>::createLayer(const std::shared_ptr<
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::LogicalNot>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "Activation", details::ngraph::convertPrecision(layer->get_output_element_type(0))};
LayerParams params = {layer->get_friendly_name(), "Activation", details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
res->params["type"] = "not";
return res;
@ -2242,7 +2244,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::LogicalNot>::createLayer(const std::
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::ReduceLogicalAnd>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "ReduceAnd", details::ngraph::convertPrecision(layer->get_output_element_type(0))};
LayerParams params = {layer->get_friendly_name(), "ReduceAnd", details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ReduceLayer>(params);
auto castedLayer = std::dynamic_pointer_cast<ngraph::op::v1::ReduceLogicalAnd>(layer);
@ -2254,7 +2256,7 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v1::ReduceLogicalAnd>::createLayer(const
template <>
CNNLayer::Ptr NodeConverter<ngraph::op::v1::ReduceLogicalOr>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
LayerParams params = {layer->get_friendly_name(), "ReduceOr", details::ngraph::convertPrecision(layer->get_output_element_type(0))};
LayerParams params = {layer->get_friendly_name(), "ReduceOr", details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::ReduceLayer>(params);
auto castedLayer = std::dynamic_pointer_cast<ngraph::op::v1::ReduceLogicalOr>(layer);

2
inference-engine/src/inference_engine/ie_cnn_layer_builder_ngraph.h → inference-engine/src/legacy_api/src/ie_cnn_layer_builder_ngraph.h
View File

@ -93,7 +93,7 @@ private:
Blob::Ptr shareWeights(const std::shared_ptr<ngraph::op::Constant>& constLayer) const {
if (!constLayer) THROW_IE_EXCEPTION << "Cannot share weights! Constant operation is empty!";
auto dataPrecision = details::ngraph::convertPrecision(constLayer->get_element_type());
auto dataPrecision = details::convertPrecision(constLayer->get_element_type());
size_t shapeSize = ngraph::shape_size(constLayer->get_shape());
if (dataPrecision == Precision::BIN) {

65
inference-engine/src/legacy_api/src/ie_layer_validators.cpp
View File

@ -2447,7 +2447,9 @@ void PriorBoxClusteredValidator::checkShapes(const CNNLayer* layer, const std::v
PriorBoxClusteredValidator::PriorBoxClusteredValidator(const std::string& _type): LayerValidator(_type) {}
void ProposalValidator::parseParams(CNNLayer* layer) {
layer->params["num_outputs"] = std::to_string(layer->outData.size());
if (layer->params.find("num_outputs") == layer->params.end()) {
layer->params["num_outputs"] = std::to_string(layer->outData.size());
}
}
void ProposalValidator::checkParams(const CNNLayer* layer) {
@ -3074,52 +3076,55 @@ void NMSValidator::checkShapes(const CNNLayer* layer, const vector<SizeVector>&
THROW_IE_EXCEPTION << layer->name << " 'score_threshold' should be scalar";
}
ScatterValidator::ScatterValidator(const std::string& _type): LayerValidator(_type) {}
ScatterUpdateValidator::ScatterUpdateValidator(const std::string& _type): LayerValidator(_type) {}
void ScatterValidator::parseParams(CNNLayer* layer) {
auto casted = dynamic_cast<ScatterLayer*>(layer);
void ScatterUpdateValidator::parseParams(CNNLayer* layer) {
auto casted = dynamic_cast<ScatterUpdateLayer*>(layer);
if (!casted) {
THROW_IE_EXCEPTION << layer->name << " Layer is not instance of ScatterLayer class";
THROW_IE_EXCEPTION << layer->name << " Layer is not instance of ScatterUpdateLayer class";
}
casted->axis = casted->GetParamAsInt("axis", 0);
}
void ScatterValidator::checkShapes(const CNNLayer* layer, const vector<SizeVector>& inShapes) const {
auto casted = dynamic_cast<const ScatterLayer*>(layer);
void ScatterUpdateValidator::checkShapes(const CNNLayer* layer, const vector<SizeVector>& inShapes) const {
auto casted = dynamic_cast<const ScatterUpdateLayer*>(layer);
if (!casted) {
THROW_IE_EXCEPTION << layer->name << " Layer is not instance of ScatterLayer class";
THROW_IE_EXCEPTION << layer->name << " Layer is not instance of ScatterUpdateLayer class";
}
size_t numInputs = inShapes.size();
if (numInputs != 3)
THROW_IE_EXCEPTION << layer->name << " Scatter can take only 3 inputs, but actually it has: " << numInputs;
if (numInputs != 4)
THROW_IE_EXCEPTION << layer->name << " Scatter can take only 4 inputs, but actually it has: " << numInputs;
if (!(-static_cast<int>(inShapes[0].size()) <= casted->axis && casted->axis < static_cast<int>(inShapes[0].size())))
THROW_IE_EXCEPTION << layer->name << " Incorrect input parameters dimensions and axis number!";
static constexpr int DATA = 0;
static constexpr int INDICES = 1;
static constexpr int UPDATES = 2;
static constexpr int AXIS = 3;
if (inShapes[0].size() == 0 || (inShapes[0].size() == 1 && inShapes[0][0] == 1))
THROW_IE_EXCEPTION << layer->name << " 'Data' tensor rank should be >= 1";
if (inShapes[DATA].size() < 1)
THROW_IE_EXCEPTION << layer->name << " 'Data' tensor rank must be >= 1";
if (inShapes[1].size() == 0 || (inShapes[1].size() == 1 && inShapes[1][0] == 1))
THROW_IE_EXCEPTION << layer->name << " 'Indexes' tensor rank should be >= 1";
if (inShapes[INDICES].size() < 1)
THROW_IE_EXCEPTION << layer->name << " 'Indices' tensor rank must be >= 1";
if (inShapes[1].size() == 0 || (inShapes[1].size() == 1 && inShapes[1][0] == 1))
THROW_IE_EXCEPTION << layer->name << " 'Updates' tensor rank should be >= 1";
if (inShapes[UPDATES].size() < 1)
THROW_IE_EXCEPTION << layer->name << " 'Updates' tensor rank must be >= 1";
if (inShapes[1] != inShapes[2])
if (!(inShapes[AXIS].size() == 1 && inShapes[AXIS][0] == 1))
THROW_IE_EXCEPTION << layer->name << " 'Axis' tensor must be 1D array of 1 element";
if (inShapes[UPDATES].size() != inShapes[INDICES].size() + inShapes[DATA].size() - 1)
THROW_IE_EXCEPTION << layer->name << " Incorrect number of 'indexes' and 'updates' tensors dimension";
const size_t SCATTER_DATA = 0;
const size_t SCATTER_INDEXES = 1;
const size_t SCATTER_UPDATES = 2;
Precision inIdxPrecision = layer->insData[SCATTER_INDEXES].lock()->getTensorDesc().getPrecision();
Precision inIdxPrecision = layer->insData[INDICES].lock()->getTensorDesc().getPrecision();
if (inIdxPrecision != Precision::FP32 && inIdxPrecision != Precision::I32)
THROW_IE_EXCEPTION << layer->name << " Incorrect input 'Indexes' precision. Only FP32 or I32 are supported!";
THROW_IE_EXCEPTION << layer->name << " Incorrect input 'Indices' precision. Only FP32 or I32 are supported!";
if (layer->insData[SCATTER_DATA].lock()->getTensorDesc().getPrecision() !=
layer->insData[SCATTER_UPDATES].lock()->getTensorDesc().getPrecision())
Precision inAxisPrecision = layer->insData[AXIS].lock()->getTensorDesc().getPrecision();
if (inAxisPrecision != Precision::FP32 && inAxisPrecision != Precision::I32)
THROW_IE_EXCEPTION << layer->name << " Incorrect input 'Axis' precision. Only FP32 or I32 are supported!";
if (layer->insData[DATA].lock()->getTensorDesc().getPrecision() !=
layer->insData[UPDATES].lock()->getTensorDesc().getPrecision())
THROW_IE_EXCEPTION << layer->name << " Precision should be equal for input tensors 'Data' and 'Updates'";
}
@ -3248,7 +3253,7 @@ LayerValidators::LayerValidators() {
REG_LAYER_VALIDATOR_FOR_TYPE(TopKValidator, TopK);
REG_LAYER_VALIDATOR_FOR_TYPE(UniqueValidator, Unique);
REG_LAYER_VALIDATOR_FOR_TYPE(NMSValidator, NonMaxSuppression);
REG_LAYER_VALIDATOR_FOR_TYPE(ScatterValidator, ScatterUpdate);
REG_LAYER_VALIDATOR_FOR_TYPE(ScatterUpdateValidator, ScatterUpdate);
}
} // namespace InferenceEngine

4
inference-engine/src/legacy_api/src/ie_layer_validators.hpp
View File

@ -969,9 +969,9 @@ public:
void checkShapes(const CNNLayer* layer, const std::vector<SizeVector>& inShapes) const override;
};
class ScatterValidator : public LayerValidator {
class ScatterUpdateValidator : public LayerValidator {
public:
explicit ScatterValidator(const std::string& _type);
explicit ScatterUpdateValidator(const std::string& _type);
void parseParams(CNNLayer* layer) override;

3
inference-engine/src/legacy_api/src/ie_layers.cpp
View File

@ -67,6 +67,7 @@ ReduceLayer::~ReduceLayer() {}
TopKLayer::~TopKLayer() {}
UniqueLayer::~UniqueLayer() {}
NonMaxSuppressionLayer::~NonMaxSuppressionLayer() {}
ScatterLayer::~ScatterLayer() {}
ScatterUpdateLayer::~ScatterUpdateLayer() {}
ExperimentalDetectronPriorGridGeneratorLayer::~ExperimentalDetectronPriorGridGeneratorLayer() {}
ExperimentalDetectronGenerateProposalsSingleImageLayer::~ExperimentalDetectronGenerateProposalsSingleImageLayer() {}
ExperimentalDetectronTopKROIs::~ExperimentalDetectronTopKROIs() {}

34
inference-engine/src/legacy_api/src/ie_util_internal.cpp
View File

@ -77,9 +77,10 @@ CNNLayerPtr layerCloneImpl<TensorIterator>(const CNNLayer* source) {
CNNLayerPtr clonelayer(const CNNLayer& source) {
using fptr = CNNLayerPtr (*)(const CNNLayer*);
// Most derived layers must go first in this list
static const fptr cloners[] = {&layerCloneImpl<ExperimentalDetectronGenerateProposalsSingleImageLayer>,
static const fptr cloners[] = {&layerCloneImpl<ExperimentalDetectronTopKROIs>,
&layerCloneImpl<ExperimentalDetectronGenerateProposalsSingleImageLayer>,
&layerCloneImpl<ExperimentalDetectronPriorGridGeneratorLayer>,
&layerCloneImpl<ScatterLayer>,
&layerCloneImpl<ScatterUpdateLayer>,
&layerCloneImpl<NonMaxSuppressionLayer>,
&layerCloneImpl<SelectLayer>,
&layerCloneImpl<BatchNormalizationLayer>,
@ -145,6 +146,35 @@ CNNLayerPtr clonelayer(const CNNLayer& source) {
return nullptr; // Silence "control may reach end of non-void function" warning
}
std::shared_ptr<ICNNNetwork> cloneNetwork(const ICNNNetwork& network) {
if (auto func = network.getFunction()) {
CNNNetwork net(func);
InputsDataMap originInputs;
OutputsDataMap originOutputs;
network.getInputsInfo(originInputs);
network.getOutputsInfo(originOutputs);
InputsDataMap clonedInputs = net.getInputsInfo();
OutputsDataMap clonedOutputs = net.getOutputsInfo();
for (const auto& outputInfo : originOutputs) {
if (clonedOutputs.find(outputInfo.first) == clonedOutputs.end())
THROW_IE_EXCEPTION << "Cannot clone network! Cloned network doesn't contain all outputs";
clonedOutputs[outputInfo.first]->setPrecision(outputInfo.second->getPrecision());
clonedOutputs[outputInfo.first]->setLayout(outputInfo.second->getLayout());
}
for (const auto& inputInfo : originInputs) {
if (clonedInputs.find(inputInfo.first) == clonedInputs.end())
THROW_IE_EXCEPTION << "Cannot clone network! Cloned network doesn't contain all inputs";
clonedInputs[inputInfo.first]->setPrecision(inputInfo.second->getPrecision());
clonedInputs[inputInfo.first]->setLayout(inputInfo.second->getLayout());
clonedInputs[inputInfo.first]->getPreProcess() = inputInfo.second->getPreProcess();
}
return net;
}
return cloneNet(network);
}
details::CNNNetworkImplPtr cloneNet(const ICNNNetwork& network) {
std::vector<CNNLayerPtr> layers;
details::CNNNetworkIterator i(&network);

2
inference-engine/src/legacy_api/src/layer_transform.cpp
View File

@ -1,4 +1,4 @@
// Copyright (C) 2018-2020 Intel Corporation
// Copyright (C) 2018-2019 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//

4
inference-engine/src/legacy_api/src/network_serializer.cpp
View File

@ -487,6 +487,7 @@ std::size_t FillXmlDoc(const InferenceEngine::ICNNNetwork& network, pugi::xml_do
if (dumpWeights && !node->blobs.empty()) {
auto blobsNode = layer.append_child("blobs");
for (const auto& dataIt : node->blobs) {
if (!dataIt.second) continue;
size_t dataSize = dataIt.second->byteSize();
pugi::xml_node data = blobsNode.append_child(dataIt.first.c_str());
data.append_attribute("offset").set_value(dataOffset);
@ -546,11 +547,12 @@ void SerializeBlobs(std::ostream& stream, const InferenceEngine::ICNNNetwork& ne
for (auto&& node : ordered) {
if (!node->blobs.empty()) {
for (const auto& dataIt : node->blobs) {
if (!dataIt.second) continue;
const char* dataPtr = dataIt.second->buffer().as<char*>();
size_t dataSize = dataIt.second->byteSize();
stream.write(dataPtr, dataSize);
if (!stream.good()) {
THROW_IE_EXCEPTION << "Error during writing blob waights";
THROW_IE_EXCEPTION << "Error during writing blob weights";
}
}
}

2
inference-engine/src/legacy_api/src/shape_infer/built-in/ie_built_in_holder.cpp
View File

@ -248,7 +248,7 @@ REG_SHAPE_INFER_FOR_TYPE(GatherTreeShapeProp, GatherTree);
REG_SHAPE_INFER_FOR_TYPE(TopKShapeProp, TopK);
REG_SHAPE_INFER_FOR_TYPE(UniqueShapeProp, Unique);
REG_SHAPE_INFER_FOR_TYPE(NMSShapeProp, NonMaxSuppression);
REG_SHAPE_INFER_FOR_TYPE(ScatterShapeProp, Scatter);
REG_SHAPE_INFER_FOR_TYPE(ScatterUpdateShapeProp, ScatterUpdate);
} // namespace ShapeInfer
} // namespace InferenceEngine

14
inference-engine/src/legacy_api/src/shape_infer/built-in/ie_scatter_shape_infer.hpp
View File

@ -15,19 +15,19 @@ namespace InferenceEngine {
namespace ShapeInfer {
/**
*@brief Implementation of Shape inference for Scatter layer
*@brief Implementation of Shape inference for ScatterUpdate layer
*/
class ScatterShapeProp : public BuiltInShapeInferImpl {
class ScatterUpdateShapeProp : public BuiltInShapeInferImpl {
public:
explicit ScatterShapeProp(const std::string& type): BuiltInShapeInferImpl(type) {}
explicit ScatterUpdateShapeProp(const std::string& type): BuiltInShapeInferImpl(type) {}
void inferShapesImpl(const std::vector<Blob::CPtr>& inBlobs, const std::map<std::string, std::string>& params,
const std::map<std::string, Blob::Ptr>& blobs, std::vector<SizeVector>& outShapes) override {
LayerParams lp {};
ScatterLayer scatterLayer(lp);
scatterLayer.params = params;
scatterLayer.type = _type;
validate(&scatterLayer, inBlobs, params, blobs);
ScatterUpdateLayer scatterUpdateLayer(lp);
scatterUpdateLayer.params = params;
scatterUpdateLayer.type = _type;
validate(&scatterUpdateLayer, inBlobs, params, blobs);
outShapes = {inShapes[0]};
}

Some files were not shown because too many files have changed in this diff Show More