Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 337
Star 4
Fork 1
P
Paddle-Lite
未验证 提交 5dd85c2c 编写于 作者: H HappyAngel 提交者： GitHub
浏览文件
操作

Merge pull request #89 from PaddlePaddle/develop 

pull code
上级 1a56bcba f81db03a
展开全部 隐藏空白更改
内联 并排
Showing with 35701 addition and 8838 deletion
.gitignore
浏览文件 @ 5dd85c2c
......@@ -105,3 +105,5 @@ metal/paddle-mobile-demo/paddle-mobile-demo/Resources
metal/paddle-mobile-demo/paddle-mobile-demo/Resources/images
metal/paddle-mobile-demo/paddle-mobile-demo/Resources/models
metal/MobileNetDemo/MobileNetDemo/Resources
build*
CMakeLists.txt
浏览文件 @ 5dd85c2c
......@@ -61,6 +61,7 @@ lite_option(LITE_WITH_ARM "Enable ARM in lite mode" OFF)
lite_option(LITE_WITH_NPU "Enable NPU in lite mode" OFF)
lite_option(LITE_WITH_XPU "Enable XPU in lite mode" OFF)
lite_option(LITE_WITH_BM "Enable BM in lite mode" OFF)
lite_option(LITE_WITH_TRAIN "Enable training operators and kernels in lite" OFF)
lite_option(LITE_WITH_OPENMP "Enable OpenMP in lite framework" ON)
lite_option(LITE_WITH_OPENCL "Enable OpenCL support in lite" OFF)
lite_option(LITE_WITH_FPGA "Enable FPGA support in lite" OFF)
......@@ -76,6 +77,7 @@ lite_option(LITE_BUILD_TAILOR "Enable tailoring library according to model" OFF)
# cv build options
lite_option(LITE_WITH_CV "Enable build cv image in lite" OFF)
lite_option(LITE_WITH_STATIC_CUDA "Statically link cuda libraries." ON)
lite_option(LITE_WITH_ARM_CLANG "when arm lang is clang, its ON." OFF)
# TODO(Superjomn) Remove WITH_ANAKIN option if not needed latter.
if(ANDROID OR IOS OR ARMLINUX)
......@@ -130,7 +132,8 @@ endif()
if (WITH_LITE AND LITE_WITH_LIGHT_WEIGHT_FRAMEWORK)
message(STATUS "Building the mobile framework")
include(cross_compiling/postproject)
include(cross_compiling/npu) # check and prepare NPU DDK
include(device/npu) # check and prepare NPU DDK
include(device/xpu) # check and prepare XPU SDK
# We compile the mobile deployment library when LITE_ON_TINY_PUBLISH=ON
# So the following third party dependencies are not needed.
......@@ -171,7 +174,7 @@ endif()
########################################################################################
if(LITE_WITH_XPU)
include(xpu)
include(device/xpu)
endif()
include(external/mklml) # download mklml package
......
cmake/configure.cmake
浏览文件 @ 5dd85c2c
......@@ -122,6 +122,9 @@ if (LITE_WITH_ARM)
endif()
endif()
if (LITE_WITH_TRAIN)
add_definitions("-DLITE_WITH_TRAIN")
endif()
if (WITH_ARM_DOTPROD)
add_definitions("-DWITH_ARM_DOTPROD")
......
cmake/cross_compiling/findar.cmake
浏览文件 @ 5dd85c2c
......@@ -23,7 +23,7 @@ endif()
get_filename_component(AR_PATH ${CMAKE_CXX_COMPILER} PATH)
find_file(AR_TOOL NAMES llvm-ar PATHS ${AR_PATH})
find_file(AR_TOOL NAMES llvm-ar PATHS ${AR_PATH} NO_DEFAULT_PATH)
if(NOT AR_TOOL)
message(ERROR "Failed to find AR_TOOL in ${AR_PATH}")
......
cmake/cross_compiling/postproject.cmake
浏览文件 @ 5dd85c2c
......@@ -57,10 +57,14 @@ function(check_linker_flag)
endforeach()
set(CMAKE_SHARED_LINKER_FLAGS ${CMAKE_SHARED_LINKER_FLAGS} PARENT_SCOPE)
endfunction()
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
if (LITE_ON_TINY_PUBLISH)
if(NOT LITE_WITH_PYTHON)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fno-exceptions")
if((NOT LITE_WITH_PYTHON))
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fno-exceptions")
endif()
if(LITE_WITH_OPENCL AND (ARM_TARGET_LANG STREQUAL "clang"))
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fexceptions")
endif()
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -ffast-math -Ofast -Os -fomit-frame-pointer -fno-asynchronous-unwind-tables -fno-unwind-tables")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fvisibility=hidden -fvisibility-inlines-hidden -ffunction-sections")
......
cmake/cross_compiling/npu.cmake → cmake/device/npu.cmake
浏览文件 @ 5dd85c2c
......@@ -17,15 +17,16 @@ if(NOT LITE_WITH_NPU)
endif()
if(NOT DEFINED NPU_DDK_ROOT)
set(NPU_DDK_ROOT $ENV{NPU_DDK_ROOT})
if(NOT NPU_DDK_ROOT)
message(FATAL_ERROR "Must set NPU_DDK_ROOT or env NPU_DDK_ROOT when LITE_WITH_NPU=ON")
endif()
set(NPU_DDK_ROOT $ENV{NPU_DDK_ROOT})
if(NOT NPU_DDK_ROOT)
message(FATAL_ERROR "Must set NPU_DDK_ROOT or env NPU_DDK_ROOT when LITE_WITH_NPU=ON")
endif()
endif()
message(STATUS "NPU_DDK_ROOT: ${NPU_DDK_ROOT}")
find_path(NPU_DDK_INC NAMES HiAiModelManagerService.h
PATHS ${NPU_DDK_ROOT}/include NO_DEFAULT_PATH)
PATHS ${NPU_DDK_ROOT}/include
NO_DEFAULT_PATH)
if(NOT NPU_DDK_INC)
message(FATAL_ERROR "Can not find HiAiModelManagerService.h in ${NPU_DDK_ROOT}/include")
endif()
......@@ -34,21 +35,24 @@ include_directories("${NPU_DDK_ROOT}/include")
set(NPU_SUB_LIB_PATH "lib64")
if(ARM_TARGET_ARCH_ABI STREQUAL "armv8")
set(NPU_SUB_LIB_PATH "lib64")
set(NPU_SUB_LIB_PATH "lib64")
endif()
if(ARM_TARGET_ARCH_ABI STREQUAL "armv7")
set(NPU_SUB_LIB_PATH "lib")
set(NPU_SUB_LIB_PATH "lib")
endif()
find_library(NPU_DDK_HIAI_FILE NAMES hiai
PATHS ${NPU_DDK_ROOT}/${NPU_SUB_LIB_PATH})
PATHS ${NPU_DDK_ROOT}/${NPU_SUB_LIB_PATH}
NO_DEFAULT_PATH)
find_library(NPU_DDK_IR_FILE NAMES hiai_ir
PATHS ${NPU_DDK_ROOT}/${NPU_SUB_LIB_PATH})
PATHS ${NPU_DDK_ROOT}/${NPU_SUB_LIB_PATH}
NO_DEFAULT_PATH)
find_library(NPU_DDK_IR_BUILD_FILE NAMES hiai_ir_build
PATHS ${NPU_DDK_ROOT}/${NPU_SUB_LIB_PATH})
PATHS ${NPU_DDK_ROOT}/${NPU_SUB_LIB_PATH}
NO_DEFAULT_PATH)
if(NOT NPU_DDK_HIAI_FILE)
message(FATAL_ERROR "Can not find NPU_DDK_HIAI_FILE in ${NPU_DDK_ROOT}")
......@@ -76,6 +80,3 @@ endif()
set(npu_runtime_libs npu_ddk_hiai CACHE INTERNAL "npu ddk runtime libs")
set(npu_builder_libs npu_ddk_ir npu_ddk_ir_build CACHE INTERNAL "npu ddk builder libs")
cmake/xpu.cmake → cmake/device/xpu.cmake
浏览文件 @ 5dd85c2c
......@@ -17,15 +17,16 @@ if(NOT LITE_WITH_XPU)
endif()
if(NOT DEFINED XPU_SDK_ROOT)
set(XPU_SDK_ROOT $ENV{XPU_SDK_ROOT})
if(NOT XPU_SDK_ROOT)
message(FATAL_ERROR "Must set XPU_SDK_ROOT or env XPU_SDK_ROOT when LITE_WITH_XPU=ON")
endif()
set(XPU_SDK_ROOT $ENV{XPU_SDK_ROOT})
if(NOT XPU_SDK_ROOT)
message(FATAL_ERROR "Must set XPU_SDK_ROOT or env XPU_SDK_ROOT when LITE_WITH_XPU=ON")
endif()
endif()
message(STATUS "XPU_SDK_ROOT: ${XPU_SDK_ROOT}")
find_path(XPU_SDK_INC NAMES xtcl.h
PATHS ${XPU_SDK_ROOT}/XTCL/include/xtcl NO_DEFAULT_PATH)
PATHS ${XPU_SDK_ROOT}/XTCL/include/xtcl
NO_DEFAULT_PATH)
if(NOT XPU_SDK_INC)
message(FATAL_ERROR "Can not find xtcl.h in ${XPU_SDK_ROOT}/include")
endif()
......@@ -34,7 +35,8 @@ include_directories("${XPU_SDK_ROOT}/XTCL/include")
include_directories("${XPU_SDK_ROOT}/XTDK/include")
find_library(XPU_SDK_XTCL_FILE NAMES xtcl
PATHS ${XPU_SDK_ROOT}/XTCL/so)
PATHS ${XPU_SDK_ROOT}/XTCL/so
NO_DEFAULT_PATH)
if(NOT XPU_SDK_XTCL_FILE)
message(FATAL_ERROR "Can not find XPU XTCL Library in ${XPU_SDK_ROOT}")
......@@ -45,7 +47,8 @@ else()
endif()
find_library(XPU_SDK_TVM_FILE NAMES tvm
PATHS ${XPU_SDK_ROOT}/XTCL/so)
PATHS ${XPU_SDK_ROOT}/XTCL/so
NO_DEFAULT_PATH)
if(NOT XPU_SDK_TVM_FILE)
message(FATAL_ERROR "Can not find XPU TVM Library in ${XPU_SDK_ROOT}")
......@@ -56,7 +59,8 @@ else()
endif()
find_library(XPU_SDK_XPU_API_FILE NAMES xpuapi
PATHS ${XPU_SDK_ROOT}/XTDK/shlib)
PATHS ${XPU_SDK_ROOT}/XTDK/shlib
NO_DEFAULT_PATH)
if(NOT XPU_SDK_XPU_API_FILE)
message(FATAL_ERROR "Can not find XPU API Library in ${XPU_SDK_ROOT}")
......@@ -67,7 +71,8 @@ else()
endif()
find_library(XPU_SDK_XPU_RT_FILE NAMES xpurt
PATHS ${XPU_SDK_ROOT}/XTDK/shlib)
PATHS ${XPU_SDK_ROOT}/XTDK/shlib
NO_DEFAULT_PATH)
if(NOT XPU_SDK_XPU_RT_FILE)
message(FATAL_ERROR "Can not find XPU RT Library in ${XPU_SDK_ROOT}")
......@@ -78,18 +83,12 @@ else()
endif()
find_library(XPU_SDK_XPU_JITC_FILE NAMES xpujitc
PATHS ${XPU_SDK_ROOT}/XTDK/shlib)
if(NOT XPU_SDK_XPU_JITC_FILE)
message(FATAL_ERROR "Can not find XPU JITC Library in ${XPU_SDK_ROOT}")
else()
message(STATUS "Found XPU JITC Library: ${XPU_SDK_XPU_JITC_FILE}")
add_library(xpu_sdk_xpu_jitc SHARED IMPORTED GLOBAL)
set_property(TARGET xpu_sdk_xpu_jitc PROPERTY IMPORTED_LOCATION ${XPU_SDK_XPU_JITC_FILE})
endif()
PATHS ${XPU_SDK_ROOT}/XTDK/shlib
NO_DEFAULT_PATH)
find_library(XPU_SDK_LLVM_FILE NAMES LLVM-8
PATHS ${XPU_SDK_ROOT}/XTDK/shlib)
PATHS ${XPU_SDK_ROOT}/XTDK/shlib
NO_DEFAULT_PATH)
if(NOT XPU_SDK_LLVM_FILE)
message(FATAL_ERROR "Can not find LLVM Library in ${XPU_SDK_ROOT}")
......@@ -99,7 +98,7 @@ else()
set_property(TARGET xpu_sdk_llvm PROPERTY IMPORTED_LOCATION ${XPU_SDK_LLVM_FILE})
endif()
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DDMLC_USE_GLOG=1 -D_GLIBCXX_USE_CXX11_ABI=0")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DDMLC_USE_GLOG=0")
set(xpu_runtime_libs xpu_sdk_xtcl xpu_sdk_tvm xpu_sdk_xpu_api xpu_sdk_xpu_rt xpu_sdk_xpu_jitc xpu_sdk_llvm CACHE INTERNAL "xpu runtime libs")
set(xpu_builder_libs xpu_sdk_xtcl xpu_sdk_tvm xpu_sdk_xpu_api xpu_sdk_xpu_rt xpu_sdk_xpu_jitc xpu_sdk_llvm CACHE INTERNAL "xpu builder libs")
set(xpu_runtime_libs xpu_sdk_xtcl xpu_sdk_tvm xpu_sdk_xpu_api xpu_sdk_xpu_rt xpu_sdk_llvm CACHE INTERNAL "xpu runtime libs")
set(xpu_builder_libs xpu_sdk_xtcl xpu_sdk_tvm xpu_sdk_xpu_api xpu_sdk_xpu_rt xpu_sdk_llvm CACHE INTERNAL "xpu builder libs")
cmake/lite.cmake
浏览文件 @ 5dd85c2c
......@@ -275,6 +275,11 @@ set(host_kernels CACHE INTERNAL "host kernels")
set(kernels_src_list "${CMAKE_BINARY_DIR}/kernels_src_list.txt")
file(WRITE ${kernels_src_list} "") # clean
# file to record faked kernels for opt python lib
set(fake_kernels_src_list "${CMAKE_BINARY_DIR}/fake_kernels_src_list.txt")
file(WRITE ${fake_kernels_src_list} "") # clean
if(LITE_BUILD_TAILOR)
set(tailored_kernels_list_path "${LITE_OPTMODEL_DIR}/.tailored_kernels_source_list")
file(STRINGS ${tailored_kernels_list_path} tailored_kernels_list)
......@@ -303,62 +308,74 @@ function(add_kernel TARGET device level)
return()
endif()
if (LITE_ON_MODEL_OPTIMIZE_TOOL)
# the source list will collect for model_optimize_tool to fake kernel generation.
foreach(src ${args_SRCS})
file(APPEND ${kernels_src_list} "${CMAKE_CURRENT_SOURCE_DIR}/${src}\n")
endforeach()
return()
endif()
# when compiling the model_optimize_tool, a source file with all the fake kernel definitions will be generated,
# no need to continue the compilation of the true kernel source.
if (LITE_ON_MODEL_OPTIMIZE_TOOL)
return()
endif(LITE_ON_MODEL_OPTIMIZE_TOOL)
if ("${device}" STREQUAL "Host")
set(host_kernels "${host_kernels};${TARGET}" CACHE INTERNAL "")
endif()
if ("${device}" STREQUAL "ARM")
if (NOT LITE_WITH_ARM)
foreach(src ${args_SRCS})
file(APPEND ${fake_kernels_src_list} "${CMAKE_CURRENT_SOURCE_DIR}/${src}\n")
endforeach()
return()
endif()
set(arm_kernels "${arm_kernels};${TARGET}" CACHE INTERNAL "")
endif()
if ("${device}" STREQUAL "X86")
if (NOT LITE_WITH_X86)
foreach(src ${args_SRCS})
file(APPEND ${fake_kernels_src_list} "${CMAKE_CURRENT_SOURCE_DIR}/${src}\n")
endforeach()
return()
elseif (LITE_ON_MODEL_OPTIMIZE_TOOL)
foreach(src ${args_SRCS})
file(APPEND ${kernels_src_list} "${CMAKE_CURRENT_SOURCE_DIR}/${src}\n")
endforeach()
return()
endif()
set(x86_kernels "${x86_kernels};${TARGET}" CACHE INTERNAL "")
endif()
if ("${device}" STREQUAL "NPU")
if (NOT LITE_WITH_NPU)
foreach(src ${args_SRCS})
file(APPEND ${fake_kernels_src_list} "${CMAKE_CURRENT_SOURCE_DIR}/${src}\n")
endforeach()
return()
endif()
set(npu_kernels "${npu_kernels};${TARGET}" CACHE INTERNAL "")
endif()
if ("${device}" STREQUAL "XPU")
if (NOT LITE_WITH_XPU)
foreach(src ${args_SRCS})
file(APPEND ${fake_kernels_src_list} "${CMAKE_CURRENT_SOURCE_DIR}/${src}\n")
endforeach()
return()
endif()
set(xpu_kernels "${xpu_kernels};${TARGET}" CACHE INTERNAL "")
endif()
if ("${device}" STREQUAL "FPGA")
if (NOT LITE_WITH_FPGA)
foreach(src ${args_SRCS})
file(APPEND ${fake_kernels_src_list} "${CMAKE_CURRENT_SOURCE_DIR}/${src}\n")
endforeach()
return()
endif()
set(fpga_kernels "${fpga_kernels};${TARGET}" CACHE INTERNAL "")
endif()
if ("${device}" STREQUAL "BM")
if (NOT LITE_WITH_BM)
foreach(src ${args_SRCS})
file(APPEND ${fake_kernels_src_list} "${CMAKE_CURRENT_SOURCE_DIR}/${src}\n")
endforeach()
return()
endif()
set(bm_kernels "${bm_kernels};${TARGET}" CACHE INTERNAL "")
endif()
if ("${device}" STREQUAL "OPENCL")
if (NOT LITE_WITH_OPENCL)
foreach(src ${args_SRCS})
file(APPEND ${fake_kernels_src_list} "${CMAKE_CURRENT_SOURCE_DIR}/${src}\n")
endforeach()
return()
endif()
set(opencl_kernels "${opencl_kernels};${TARGET}" CACHE INTERNAL "")
......@@ -366,6 +383,9 @@ function(add_kernel TARGET device level)
if ("${device}" STREQUAL "CUDA")
if (NOT LITE_WITH_CUDA)
foreach(src ${args_SRCS})
file(APPEND ${fake_kernels_src_list} "${CMAKE_CURRENT_SOURCE_DIR}/${src}\n")
endforeach()
return()
endif()
set(cuda_kernels "${cuda_kernels};${TARGET}" CACHE INTERNAL "")
......
cmake/mlu.cmake 0 → 100644
浏览文件 @ 5dd85c2c
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
if(NOT LITE_WITH_MLU)
return()
endif()
if(NOT DEFINED NEUWARE_HOME)
set(NEUWARE_HOME $ENV{NEUWARE_HOME})
if(NOT NEUWARE_HOME)
message(FATAL_ERROR "Must set NEUWARE_HOME or env NEUWARE_HOME when LITE_WITH_MLU=ON")
endif()
endif()
message(STATUS "LITE_WITH_MLU: ${LITE_WITH_MLU}")
find_path(CNML_INC NAMES cnml.h
PATHS ${NEUWARE_HOME}/include NO_DEFAULT_PATH)
if(NOT CNML_INC)
message(FATAL_ERROR "Can not find cnml.h in ${NEUWARE_HOME}/include")
endif()
find_path(CNRT_INC NAMES cnrt.h
PATHS ${NEUWARE_HOME}/include NO_DEFAULT_PATH)
if(NOT CNRT_INC)
message(FATAL_ERROR "Can not find cnrt.h in ${NEUWARE_HOME}/include")
endif()
include_directories("${NEUWARE_HOME}/include")
find_library(CNML_LIB_FILE NAMES cnml
PATHS ${NEUWARE_HOME}/lib64)
if(NOT CNML_LIB_FILE)
message(FATAL_ERROR "Can not find CNML Library in ${NEUWARE_HOME}/lib64")
else()
message(STATUS "Found CNML Library: ${CNML_LIB_FILE}")
add_library(cnml_lib SHARED IMPORTED GLOBAL)
set_property(TARGET cnml_lib PROPERTY IMPORTED_LOCATION ${CNML_LIB_FILE})
endif()
find_library(CNRT_LIB_FILE NAMES cnrt
PATHS ${NEUWARE_HOME}/lib64)
if(NOT CNRT_LIB_FILE)
message(FATAL_ERROR "Can not find CNRT Library in ${NEUWARE_HOME}/lib64")
else()
message(STATUS "Found CNRT Library: ${CNRT_LIB_FILE}")
add_library(cnrt_lib SHARED IMPORTED GLOBAL)
set_property(TARGET cnrt_lib PROPERTY IMPORTED_LOCATION ${CNRT_LIB_FILE})
endif()
docs/advanced_user_guides/test_tools.md 已删除 100644 → 0
浏览文件 @ 1a56bcba
# 测试工具
Basic profiler 用于 CPU 上kernel 耗时的统计。
## 开启方法:
参照 [编译安装](../installation/source_compile) 中的**full_publish**部分进行环境配置，在 cmake 时添加 `-DLITE_WITH_PROFILER=ON` ，就可以开启相应支持。
## 使用示例：
在模型执行完毕后，会自动打印类似如下 profiler 的日志
```
kernel average min max count
feed/def/1/4/2 0 0 0 1
conv2d/def/4/1/1 1175 1175 1175 1
conv2d/def/4/1/1 1253 1253 1253 1
depthwise_conv2d/def/4/1/1 519 519 519 1
conv2d/def/4/1/1 721 721 721 1
elementwise_add/def/4/1/1 18 18 18 1
conv2d/def/4/1/1 2174 2174 2174 1
depthwise_conv2d/def/4/1/1 380 380 380 1
conv2d/def/4/1/1 773 773 773 1
elementwise_add/def/4/1/1 2 2 2 1
conv2d/def/4/1/1 1248 1248 1248 1
depthwise_conv2d/def/4/1/1 492 492 492 1
conv2d/def/4/1/1 1150 1150 1150 1
elementwise_add/def/4/1/1 33 33 33 1
elementwise_add/def/4/1/1 3 3 3 1
conv2d/def/4/1/1 1254 1254 1254 1
depthwise_conv2d/def/4/1/1 126 126 126 1
```
docs/advanced_user_guides/x2paddle_models_doc.md 已删除 100644 → 0
浏览文件 @ 1a56bcba
# X2Paddle 支持模型列表
## 多框架支持
|模型 | caffe | tensorflow | onnx |
|---|---|---|---|
|mobilenetv1 | Y | Y | |
|mobilenetv2 | Y | Y | Y |
|resnet18 | Y | Y | |
|resnet50 | Y | Y | Y |
|mnasnet | Y | Y | |
|efficientnet | Y | Y | Y |
|squeezenetv1.1 | Y | Y | Y |
|shufflenet | Y | Y | |
|mobilenet_ssd | Y | Y | |
|mobilenet_yolov3 | | Y | |
|inceptionv4 | | | |
|mtcnn | Y | Y | |
|facedetection | Y | | |
|unet | Y | Y | |
|ocr_attention | | | |
|vgg16 | | | |
docs/advanced_user_guides/cv.md → docs/api_reference/cv.md
浏览文件 @ 5dd85c2c
# CV图像预处理库
# CV图像预处理API
请把编译脚本`Paddle-Lite/lite/too/build.sh`中`BUILD_CV`变量设置为`ON`， 其他编译参数设置请参考[源码编译](../source_compile)， 以确保 Lite 可以正确编译。这样`CV`图像的加速库就会编译进去，且会生成`paddle_image_preprocess.h`的API文件
请把编译脚本`Paddle-Lite/lite/too/build.sh`中`BUILD_CV`变量设置为`ON`， 其他编译参数设置请参考[源码编译](../user_guides/source_compile)， 以确保 Lite 可以正确编译。这样`CV`图像的加速库就会编译进去，且会生成`paddle_image_preprocess.h`的API文件
- 硬件平台： `ARM`
- 操作系统：`MAC` 和 `LINUX`
......
docs/api_reference/cxx_api_doc.md
浏览文件 @ 5dd85c2c
# C++ API文档
# C++ API
## CreatePaddlePredictor
......@@ -260,7 +260,7 @@ class MobileConfig;
`MobileConfig`用来配置构建轻量级PaddlePredictor的配置信息，如NaiveBuffer格式的模型地址、模型的内存地址(从内存加载模型时使用)、能耗模式、工作线程数等等。
*注意：输入的模型需要使用[Model Optimize Tool](../model_optimize_tool)转化为NaiveBuffer格式的优化模型。*
*注意：输入的模型需要使用[Model Optimize Tool](../user_guides/model_optimize_tool)转化为NaiveBuffer格式的优化模型。*
示例：
......@@ -277,13 +277,13 @@ config.set_power_mode(LITE_POWER_HIGH);
std::shared_ptr<PaddlePredictor> predictor = CreatePaddlePredictor<MobileConfig>(config);
```
### `set_model_from_file(model_dir)`
### `set_model_from_file(model_file)`
设置模型文件，当需要从磁盘加载模型时使用。
参数：
- `model_dir(std::string)` - 模型文件路径
- `model_file(std::string)` - 模型文件路径
返回：`None`
......@@ -589,7 +589,7 @@ for (int i = 0; i < ShapeProduction(output_tensor->shape()); i += 100) {
根据名称获取输出Tensor的指针。
**注意**：`GetTensor`接口是为开发者设计的调试接口，可以输出[转化](../model_optimize_tool)后模型中的任一节点。如果出现`GetTensor(InputName)`返回值为空`Tensor`，可能原因是以该`InputName`命名的Tensor在模型转化的**子图融合**过程被融合替换了。
**注意**：`GetTensor`接口是为开发者设计的调试接口，可以输出[转化](../user_guides/model_optimize_tool)后模型中的任一节点。如果出现`GetTensor(InputName)`返回值为空`Tensor`，可能原因是以该`InputName`命名的Tensor在模型转化的**子图融合**过程被融合替换了。
参数：
......
docs/api_reference/java_api_doc.md
浏览文件 @ 5dd85c2c
# Java API 文档
# Java API
## MobileConfig
......@@ -301,6 +301,12 @@ Tensor是Paddle-Lite的数据组织形式，用于对底层数据进行封装并
示例：
```java
// 导入Java API
import com.baidu.paddle.lite.MobileConfig;
import com.baidu.paddle.lite.Tensor;
import com.baidu.paddle.lite.Predictor;
import com.baidu.paddle.lite.PowerMode;
// 设置MobileConfig
MobileConfig config = new MobileConfig();
config.setModelDir(modelPath);
......@@ -325,7 +331,7 @@ input.setData(inputBuffer);
predictor.run();
// 获取输出Tensor
Tensor output = predictor.getOutput(0);
Tensor result = predictor.getOutput(0);
// 获取输出数据
float[] output = result.getFloatData();
for (int i = 0; i < 1000; ++i) {
......
docs/api_reference/python_api_doc.md
浏览文件 @ 5dd85c2c
# Python API文档
# Python API
## create_paddle_predictor
......
docs/benchmark/benchmark.md
浏览文件 @ 5dd85c2c
# Benchmark 数据
# 性能数据
可以参考[benchmark_tools](benchmark_tools)，推荐**一键benchmark**。
......@@ -15,14 +15,12 @@
* int8模型
* mobilenet_v1
* mobilenet_v2
* resnet50
* 测试机器(android ndk ndk-r17c)
* 骁龙855
* xiaomi mi9, snapdragon 855
* 4xA76(1@2.84GHz + 3@2.4GHz) + 4xA55@1.78GHz
* 骁龙845
* xiaomi mi8, 845
* 2.8GHz（大四核），1.7GHz（小四核）
......@@ -30,20 +28,12 @@
* 骁龙835
* xiaomi mix2, snapdragon 835
* 2.45GHz（大四核），1.9GHz（小四核）
* 骁龙625
* oppo R9s, snapdragon625
* A53 x 8, big core@2.0GHz
* 骁龙653
* 360 N5, snapdragon 653
* 4 x A73@2.0GHz + 4 x A53@1.4GHz
* 麒麟970
* HUAWEI Mate10
* 测试说明
* branch: release/2.0.0
* branch: release/v2.3.0
* warmup=10, repeats=30，统计平均时间，单位是ms
* 当线程数为1时，```DeviceInfo::Global().SetRunMode```设置LITE_POWER_HIGH，否者设置LITE_POWER_NO_BIND
* 模型的输入图像的维度是{1, 3, 224, 224}，输入图像的每一位数值是1
......@@ -55,78 +45,59 @@
#### paddlepaddle model
骁龙855|armv7 | armv7 | armv7 |armv8 | armv8 |armv8
----| ---- | ---- | ---- | ---- |---- |----
threads num|1 |2 |4 |1 |2 |4
mobilenet_v1 |32.19 |18.81 |10.90 |30.92 |18.31 |10.15
mobilenet_v2 |22.91 |13.75 |8.64 |21.15 |12.79 |7.84
shufflenet_v2 |4.67 |3.37 |2.65 |4.43 |3.15 |2.66
squeezenet_v1.1 |25.10 |15.93 |9.68 |23.28 |14.61 |8.71
mnasnet |21.84 |13.14 |7.96 |19.61 |11.88 |7.55
mobilenet_v1 |33.27 |19.52 |11.14 |31.72 |18.76 |10.24 |
mobilenet_v2 |29.08 |15.79 |9.25 |25.89 |14.17 |8.38 |
shufflenet_v2 |4.40 |3.09 |2.30 |4.28 |3.02 |2.35 |
squeezenet_v1.1 |19.96 |12.61 |8.76 |18.25 |11.46 |7.97 |
mnasnet |21.00 |12.54 |7.28 |19.65 |11.65 |6.96 |
骁龙835|armv7 | armv7 | armv7 |armv8 | armv8 |armv8
骁龙845|armv7 | armv7 | armv7 |armv8 | armv8 |armv8
----| ---- | ---- | ---- | ---- |---- |----
threads num|1 |2 |4 |1 |2 |4
mobilenet_v1 |94.13 |52.17 |30.68 |88.28 |47.58 |26.64
mobilenet_v2 |61.24 |34.64 |22.36 |56.66 |32.19 |19.63
shufflenet_v2 |10.87 |6.92 |5.12 |10.41 |6.76 |4.97
squeezenet_v1.1 |73.61 |42.25 |24.44 |64.87 |38.43 |23.06
mnasnet |58.22 |33.43 |20.44 |53.43 |30.20 |18.09
mobilenet_v1 |66.36 |35.97 |19.45 |62.66 |33.87 |17.85 |
mobilenet_v2 |45.86 |25.53 |14.6 |41.58 |23.24 |13.39 |
shufflenet_v2 |7.58 |4.89 |3.41 |7.44 |4.91 |3.58 |
squeezenet_v1.1 |37.15 |22.74 |13.51 |34.69 |21.27 |12.74 |
mnasnet |40.09 |21.73 |11.91 |38.19 |21.02 |12.11 |
麒麟980|armv7 | armv7 | armv7 |armv8 | armv8 |armv8
----| ---- | ---- | ---- | ---- |---- |----
threads num|1 |2 |4 |1 |2 |4
mobilenet_v1 |55.11 |28.24 |13.27 |34.24 |17.74 |12.41
mobilenet_v2 |37.03 |19.80 |51.94 |23.64 |12.98 |9.38
shufflenet_v2 |7.26 |4.94 |15.06 |5.32 |3.33 |2.82
squeezenet_v1.1 |42.73 |23.66 |57.39 |26.03 |14.53 |13.66
mnasnet |36.87 |20.15 |46.04 |21.85 |12.06 |8.68
麒麟970|armv7 | armv7 | armv7 |armv8 | armv8 |armv8
骁龙835|armv7 | armv7 | armv7 |armv8 | armv8 |armv8
----| ---- | ---- | ---- | ---- |---- |----
threads num|1 |2 |4 |1 |2 |4
mobilenet_v1 |97.80 |52.64 |34.46 |94.51 |49.36 |28.43
mobilenet_v2 |66.55 |38.52 |23.19 |62.89 |34.93 |21.53
shufflenet_v2 |13.78 |8.11 |5.93 |11.95 |7.90 |5.91
squeezenet_v1.1 |77.64 |43.67 |25.72 |69.91 |40.66 |24.62
mnasnet |61.86 |34.62 |22.68 |59.61 |32.79 |19.56
mobilenet_v1 |96.98 |53.92 |32.24 |89.31 |48.02 |27.58 |
mobilenet_v2 |67.72 |37.66 |23.82 |60.10 |34.36 |21.05 |
shufflenet_v2 |10.72 |6.62 |4.63 |10.10 |6.44 |4.63 |
squeezenet_v1.1 |53.89 |33.28 |20.73 |50.83 |32.31 |19.51 |
mnasnet |59.55 |33.53 |20.32 |56.21 |31.58 |19.06 |
#### caffe model
骁龙855|armv7 | armv7 | armv7 |armv8 | armv8 |armv8
----| ---- | ---- | ---- | ---- |---- |----
threads num|1 |2 |4 |1 |2 |4 |
mobilenet_v1 |32.42 |18.68 |10.86 |30.92 |18.35 |10.07 |
mobilenet_v2 |29.53 |17.76 |10.89 |27.19 |16.53 |9.75 |
shufflenet_v2 |4.61 |3.29 |2.61 |4.36 |3.11 |2.51 |
骁龙835|armv7 | armv7 | armv7 |armv8 | armv8 |armv8
----| ---- | ---- | ---- | ---- |---- |----
threads num|1 |2 |4 |1 |2 |4 |
mobilenet_v1 |92.52 |52.34 |30.37 |88.31 |49.75 |27.29 |
mobilenet_v2 |79.50 |45.67 |28.79 |76.13 |44.01 |26.13 |
shufflenet_v2 |10.94 |7.08 |5.16 |10.64 |6.83 |5.01 |
mobilenet_v1 |33.36 |19.45 |11.26 |31.63 |18.74 |10.31 |
mobilenet_v2 |31.63 |19.21 |11.61 |28.34 |17.14 |10.16 |
shufflenet_v2 |4.46 |3.08 |2.32 |4.26 |2.98 |2.35 |
麒麟980|armv7 | armv7 | armv7 |armv8 | armv8 |armv8
骁龙845|armv7 | armv7 | armv7 |armv8 | armv8 |armv8
----| ---- | ---- | ---- | ---- |---- |----
threads num|1 |2 |4 |1 |2 |4 |
mobilenet_v1 |55.36 |28.18 |13.31 |34.42 |17.93 |12.52 |
mobilenet_v2 |49.17 |26.10 |65.49 |30.50 |16.66 |11.72 |
shufflenet_v2 |8.45 |5.00 |15.65 |4.58 |3.14 |2.83 |
mobilenet_v1 |66.32 |35.83 |19.56 |62.52 |33.79 |17.91 |
mobilenet_v2 |58.46 |32.69 |18.56 |53.72 |29.86 |16.80 |
shufflenet_v2 |7.65 |4.82 |3.46 |7.55 |4.97 |3.62 |
麒麟970|armv7 | armv7 | armv7 |armv8 | armv8 |armv8
骁龙835|armv7 | armv7 | armv7 |armv8 | armv8 |armv8
----| ---- | ---- | ---- | ---- |---- |----
threads num|1 |2 |4 |1 |2 |4 |
mobilenet_v1 |97.85 |53.38 |33.85 |94.29 |49.42 |28.29 |
mobilenet_v2 |87.40 |50.25 |31.85 |85.55 |48.11 |28.24 |
shufflenet_v2 |12.16 |8.39 |6.21 |12.21 |8.33 |6.32 |
mobilenet_v1 |95.38 |54.09 |32.03 |95.05 |48.33 |27.54 |
mobilenet_v2 |88.46 |48.98 |30.23 |79.28 |44.64 |27.10 |
shufflenet_v2 |10.07 |6.51 |4.61 |10.31 |6.50 |4.66 |
#### int8量化模型测试数据
......@@ -136,6 +107,7 @@ threads num|1 |2 |4 |1 |2 |4 |
mobilenet_v1 |36.80 |21.58 |11.12 | 14.01 |8.13 |4.32 |
mobilenet_v2 |28.72 |19.08 |12.49 | 17.24 |11.55 |7.82 |
骁龙835|armv7 | armv7 | armv7 |armv8 | armv8 |armv8
----| ---- | ---- | ---- | ---- |---- |----
threads num|1 |2 |4 |1 |2 |4 |
......
docs/benchmark/benchmark_tools.md
浏览文件 @ 5dd85c2c
# Benchmark 测试方法
# 测试方法
本文将会介绍，在**Ubuntu:16.04交叉编译环境**下，用安卓手机在终端测试Paddle-Lite的性能，并介绍两种Benchmark方法：
......@@ -57,7 +57,7 @@ wget -c https://paddle-inference-dist.bj.bcebos.com/PaddleLite/benchmark_0/bench
#### 方式二：由源码编译benchmark_bin文件
根据[源码编译](../installation/source_compile)准备编译环境，拉取PaddleLite最新release发布版代码，并在仓库根目录下，执行：
根据[源码编译](../user_guides/source_compile)准备编译环境，拉取PaddleLite最新release发布版代码，并在仓库根目录下，执行：
```shell
###########################################
......@@ -135,53 +135,53 @@ sh benchmark.sh ./benchmark_bin_v8 ./benchmark_models result_armv8.txt true
> 不同手机，不同版本，测试模型的性能数据不同。
```shell
run benchmark armv7
run benchmark armv8
--------------------------------------
PaddleLite Benchmark
Threads=1 Warmup=10 Repeats=30
-- mnasnet avg = 159.8427 ms
-- mobilenet_v1 avg = 235.0072 ms
-- mobilenet_v2 avg = 173.0387 ms
-- shufflenet_v2 avg = 76.0040 ms
-- squeezenet_v11 avg = 164.2957 ms
mnasnet min = 19.83500 max = 19.38500 average = 19.65503
mobilenetv1 min = 32.00600 max = 31.56900 average = 31.81983
mobilenetv2 min = 22.37900 max = 22.08700 average = 22.28623
shufflenetv2 min = 10.80400 max = 10.62900 average = 10.68890
squeezenet min = 17.67400 max = 17.47900 average = 17.57677
Threads=2 Warmup=10 Repeats=30
-- mnasnet avg = 83.1287 ms
-- mobilenet_v1 avg = 121.6029 ms
-- mobilenet_v2 avg = 86.6175 ms
-- shufflenet_v2 avg = 41.5761 ms
-- squeezenet_v11 avg = 87.8678 ms
mnasnet min = 11.85600 max = 11.72000 average = 11.77127
mobilenetv1 min = 18.75000 max = 18.64300 average = 18.70593
mobilenetv2 min = 14.05100 max = 13.59900 average = 13.71450
shufflenetv2 min = 6.67200 max = 6.58300 average = 6.63400
squeezenet min = 12.07100 max = 11.33400 average = 11.41253
Threads=4 Warmup=10 Repeats=30
-- mnasnet avg = 73.3880 ms
-- mobilenet_v1 avg = 119.0739 ms
-- mobilenet_v2 avg = 85.3050 ms
-- shufflenet_v2 avg = 38.0762 ms
-- squeezenet_v11 avg = 64.2201 ms
mnasnet min = 7.19300 max = 7.02600 average = 7.08480
mobilenetv1 min = 10.42000 max = 10.29100 average = 10.34267
mobilenetv2 min = 8.61900 max = 8.46900 average = 8.54707
shufflenetv2 min = 4.55200 max = 4.41900 average = 4.46477
squeezenet min = 8.60000 max = 7.85200 average = 7.98407
--------------------------------------
run benchmark armv8
run benchmark armv7
--------------------------------------
PaddleLite Benchmark
Threads=1 Warmup=10 Repeats=30
-- mnasnet avg = 165.3073 ms
-- mobilenet_v1 avg = 306.0188 ms
-- mobilenet_v2 avg = 195.1884 ms
-- shufflenet_v2 avg = 99.3692 ms
-- squeezenet_v11 avg = 156.6971 ms
mnasnet min = 20.98300 max = 20.81400 average = 20.92527
mobilenetv1 min = 33.19000 max = 32.81700 average = 33.08490
mobilenetv2 min = 25.91400 max = 25.61700 average = 25.73097
shufflenetv2 min = 11.14300 max = 10.97600 average = 11.06757
squeezenet min = 19.31800 max = 19.20000 average = 19.26530
Threads=2 Warmup=10 Repeats=30
-- mnasnet avg = 90.2290 ms
-- mobilenet_v1 avg = 157.0007 ms
-- mobilenet_v2 avg = 118.1607 ms
-- shufflenet_v2 avg = 68.6804 ms
-- squeezenet_v11 avg = 91.3090 ms
mnasnet min = 12.59900 max = 12.46600 average = 12.52207
mobilenetv1 min = 19.05800 max = 18.94700 average = 18.97897
mobilenetv2 min = 15.28400 max = 15.11300 average = 15.19843
shufflenetv2 min = 6.97000 max = 6.81400 average = 6.90863
squeezenet min = 12.87900 max = 12.12900 average = 12.22530
Threads=4 Warmup=10 Repeats=30
-- mnasnet avg = 179.9730 ms
-- mobilenet_v1 avg = 204.0684 ms
-- mobilenet_v2 avg = 181.6486 ms
-- shufflenet_v2 avg = 123.2728 ms
-- squeezenet_v11 avg = 412.9046 ms
mnasnet min = 7.31400 max = 7.12900 average = 7.20357
mobilenetv1 min = 11.44000 max = 10.86900 average = 10.94383
mobilenetv2 min = 9.14900 max = 9.03800 average = 9.09907
shufflenetv2 min = 4.60600 max = 4.49400 average = 4.53360
squeezenet min = 8.27000 max = 8.10600 average = 8.19000
--------------------------------------
```
docs/demo_guides/android_app_demo.md 0 → 100644
浏览文件 @ 5dd85c2c
# Android Demo
## 多种应用场景
我们提供的Paddle-Lite示例工程[Paddle-Lite-Demo](https://github.com/PaddlePaddle/Paddle-Lite-Demo)，其中包含[Android](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-android-demo)、[iOS](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-ios-demo)和[Armlinux](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-armlinux-demo)平台的示例工程。涵盖[人脸识别](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-android-demo/face_detection_demo)、[人像分割](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-android-demo/human_segmentation_demo)、[图像分类](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-android-demo/image_classification_demo)、[目标检测](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-android-demo/object_detection_demo)4个应用场景。
### 1. 人脸识别
人脸检测是Paddle-Lite提供的人像检测demo。在移动端上提供了高精度、实时的人脸检测能力，能处理基于人脸检测的业务场景。在移动端预测的效果图如下：
<p align="center"><img width="300" height="250" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/face.jpg"/>&#8194;&#8194;&#8194;&#8194;&#8194;<img width="300" height="250" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/face2.jpg"/></p>
### 2. 人像分割
人像分割是Paddle-Lite 提供的图像分割demo ，在移动端上提供了实时的人像分割能力，可以应用证件照自动抠图、面积测量、智能交通（标记车道和交通标志）等场景。 在移动端预测的效果图如下：
<p align="center"><img width="250" height="250" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/human.jpg"/>&#8194;&#8194;&#8194;&#8194;&#8194;<img width="250" height="250" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/human2.jpg"/></p>
### 3. 图像分类
图像分类是Paddle-Lite 提供的图像处理demo ，在移动端上提供了实时的物体识别能力，可以应用到生产线自动分拣或质检、识别医疗图像、辅助医生肉眼诊断等场景。在移动端预测的效果图如下：
<p align="center"><img width="250" height="250" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/tabby_cat.jpg"/>&#8194;&#8194;&#8194;&#8194;&#8194;<img width="250" height="250" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/tabby_cat2.jpg"/></p>
### 4. 物体检测
物体检测是Paddle-Lite 提供的图像识别demo ，在移动端上提供了检测多个物体的位置、名称、位置及数量的能力。可以应用到视频监控（是否有违规物体或行为）、工业质检（微小瑕疵的数量和位置）、医疗诊断（细胞计数、中药识别）等场景。在移动端预测的效果图如下：
<p align="center"><img width="250" height="250" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/dog.jpg"/>&#8194;&#8194;&#8194;&#8194;&#8194;<img width="250" height="250" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/dog2.jpg"/></p>
## Android demo部署方法
下面我们以 **目标检测示例（object_detection_demo)** 为例讲解如何部署。
**目的**：将基于Paddle-Lite预测库的Android APP 部署到手机，实现物体检测
**需要的环境**： Android Studio、Android手机（开启USB调试模式）、下载到本地的[Paddle-Lite-Demo](https://github.com/PaddlePaddle/Paddle-Lite-Demo)工程
**部署步骤**：
1、 目标检测的Android示例位于 `Paddle-Lite-Demo\PaddleLite-android-demo\object_detection_demo`
2、用Android Studio 打开object_detection_demo工程 （本步骤需要联网）。
3、手机连接电脑，打开**USB调试**和**文件传输模式**，在Android Studio上连接自己的手机设备（手机需要开启允许从 USB安装软件权限）
![Android_studio](https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/android/Android_studio.png)
4、按下 Run按钮，自动编译APP并安装到手机。(该过程会自动下载Paddle-Lite预测库和模型，需要联网)
成功后效果如下，图一：APP安装到手机 图二： APP打开后的效果，会自动识别图片中的物体并标记
<p align="center"><img width="300" height="450" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/android/AndroidApp0.png"/>&#8194;&#8194;&#8194;&#8194;&#8194;<img width="300" height="450" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/android/AndroidApp1.jpg"/></p>
## Android demo结构讲解
Android 示例的代码结构如下图所示：
<p align="center"><img width="600" height="450" src="http://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/android/Android_struct.png"/>
1、 Predictor.java： 预测代码
```shell
# 位置：
object_detection_demo/app/src/main/java/com/baidu/paddle/lite/demo/object_detection/Predictor.java
```
2、 model.nb : 模型文件 (opt 工具转化后Paddle-Lite模型)；pascalvoc_label_list：训练模型时的`labels`文件
```shell
# 位置：
object_detection_demo/app/src/main/assets/models/ssd_mobilenet_v1_pascalvoc_for_cpu/model.nb
object_detection_demo/app/src/main/assets/labels/pascalvoc_label_list
```
3、 libpaddle_lite_jni.so、PaddlePredictor.jar：Paddle-Lite Java 预测库与Jar包
```shell
# 位置
object_detection_demo/app/src/main/jniLibs/arm64-v8a/libpaddle_lite_jni.so
object_detection_demo/app/libs/PaddlePredictor.jar
```
4、 build.gradle : 定义编译过程的 gradle 脚本。（不用改动，定义了自动下载Paddle-Lite预测和模型的过程）
```shell
# 位置
object_detection_demo/app/build.gradle
```
## 代码讲解 （使用Paddle-Lite Java API 执行预测）
Android 示例基于Java API 开发，调用Paddle-Lite Java API包括以下五步。更详细的API 描述参考： [Paddle-Lite Java API](https://paddle-lite.readthedocs.io/zh/latest/api_reference/java_api_doc.html)。
```c++
// 导入Java API
import com.baidu.paddle.lite.MobileConfig;
import com.baidu.paddle.lite.Tensor;
import com.baidu.paddle.lite.Predictor;
import com.baidu.paddle.lite.PowerMode;
// 1. 写入配置：设置MobileConfig
MobileConfig config = new MobileConfig();
config.setModelFromFile(<modelPath>); // 设置Paddle-Lite模型路径
config.setPowerMode(PowerMode.LITE_POWER_NO_BIND); // 设置CPU运行模式
config.setThreads(4); // 设置工作线程数
// 2. 创建 PaddlePredictor
PaddlePredictor predictor = PaddlePredictor.createPaddlePredictor(config);
// 3. 设置输入数据
long[] dims = {100, 100};
float[] inputBuffer = new float[10000];
for (int i = 0; i < 10000; ++i) {
inputBuffer[i] = i;
}
Tensor input = predictor.getInput(0);
input.resize(dims);
input.setData(inputBuffer);
// 4. 执行预测
predictor.run();
// 5. 获取输出数据
Tensor result = predictor.getOutput(0);
float[] output = result.getFloatData();
for (int i = 0; i < 1000; ++i) {
System.out.println(output[i]);
}
```
docs/user_guides/cpp_demo.md → docs/demo_guides/cpp_demo.md
浏览文件 @ 5dd85c2c
......@@ -2,7 +2,7 @@
## 1. 下载最新版本预测库
预测库下载界面位于[Paddle-Lite官方预编译库](../installation/release_lib)，可根据需求选择合适版本。
预测库下载界面位于[Paddle-Lite官方预编译库](../user_guides/release_lib)，可根据需求选择合适版本。
以**Android-ARMv8架构**为例，可以下载以下版本：
......@@ -173,7 +173,7 @@ predictor->Run();
std::unique_ptr<const Tensor> output_tensor(
std::move(predictor->GetOutput(0)));
// 转化为数据
auto output_data=output_tensor<float>();
auto output_data=output_tensor->data<float>();
```
......
docs/advanced_user_guides/cuda.md → docs/demo_guides/cuda.md
浏览文件 @ 5dd85c2c
# Lite基于CUDA的模型预测
# PaddleLite使用CUDA预测部署
Lite支持在x86_64，arm64架构上（如：TX2）进行CUDA的编译运行。
......@@ -28,7 +28,27 @@ cd Paddle-Lite
./lite/tools/build.sh --build_python=ON cuda
```
编译结束会在 `build_cuda/inference_lite_lib/python/lib/` 目录下生成 `lite_core.so`。
## 编译结果说明
cuda的编译结果位于 `build_cuda/inference_lite_lib`
**具体内容**说明：
1、 `bin`文件夹：可执行工具文件，目前为空
2、 `cxx`文件夹：包含c++的库文件与相应的头文件
- `include` : 头文件
- `lib` : 库文件
- 打包的静态库文件：
- `libpaddle_api_full_bundled.a` ：包含 full_api 和 light_api 功能的静态库
- 打包的动态态库文件：
- `libpaddle_full_api_shared.so` ：包含 full_api 和 light_api 功能的动态库
3、 `third_party` 文件夹：第三方库文件
4、 `demo` 文件夹：c++ demo.
如果编译打开了python选项，则会在 `build_cuda/inference_lite_lib/python/lib/` 目录下生成 `lite_core.so`。
## 运行
......@@ -36,7 +56,6 @@ cd Paddle-Lite
一： 下载darknet_yolov3模型，模型信息请参考[这里](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/yolov3)
```
# 下载模型
wget https://paddle-inference-dist.cdn.bcebos.com/PaddleLite/yolov3_infer.tar.gz
......@@ -47,7 +66,7 @@ wget https://paddle-inference-dist.cdn.bcebos.com/PaddleLite/kite.jpg
二： 运行
**NOTE:**此处示例使用的是python接口，后续会开放C++接口以及示例。
**NOTE:**此处示例使用的是python接口。
``` python
#-*- coding: utf-8 -*-
......@@ -107,4 +126,14 @@ print (output_tensor.float_data()[:6])
```
**NOTE：** 对CUDA的支持还在持续开发中。
**NOTE：** 此处示例使用的是C++接口。
```
cd build_cuda/inference_lite_lib/demo/cxx/
mkdir build && cd build
cmake ..
make
wget https://paddle-inference-dist.cdn.bcebos.com/PaddleLite/yolov3_infer.tar.gz
tar -zxf yolov3_infer.tar.gz
./demo yolov3_infer
```
docs/advanced_user_guides/fpga.md → docs/demo_guides/fpga.md
浏览文件 @ 5dd85c2c
# Lite基于FPGA的模型预测
# PaddleLite使用FPGA预测部署
Paddle Lite支持基于arm的FPGA zu3/zu5/zu9的模型预测，提供armv8的交叉编译
......@@ -22,7 +22,7 @@ CMAKE编译选项：
- 设置`LITE_WITH_FPGA=ON`和`LITE_WITH_ARM=ON`
其他编译选项与ARM编译相同，可以参考[“Paddle Lite在Docker下的ARM编译”](../source_compile)。
其他编译选项与ARM编译相同，可以参考[“Paddle Lite在Docker下的ARM编译”](../user_guides/source_compile)。
示例如下：
```shell
cmake .. \
......
docs/demo_guides/ios_app_demo.md 0 → 100644
浏览文件 @ 5dd85c2c
# iOS Demo
## 多种应用场景
我们提供Paddle-Lite示例工程[Paddle-Lite-Demo](https://github.com/PaddlePaddle/Paddle-Lite-Demo)，其中包含[Android](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-android-demo)、[iOS](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-ios-demo)和[Armlinux](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-armlinux-demo)平台的示例工程。iOS demo涵盖[图像分类](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-android-demo/image_classification_demo)、[目标检测](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-android-demo/object_detection_demo)2个应用场景。
### 1. 图像分类
图像分类是Paddle-Lite 提供的图像处理demo ，在移动端上提供了实时的物体识别能力，可以应用到生产线自动分拣或质检、识别医疗图像、辅助医生肉眼诊断等场景。在移动端预测的效果图如下：
<p align="center"><img width="250" height="250" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/tabby_cat.jpg"/>&#8194;&#8194;&#8194;&#8194;&#8194;<img width="250" height="250" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/tabby_cat2.jpg"/></p>
### 2. 物体检测
物体检测是Paddle-Lite 提供的图像识别demo ，在移动端上提供了检测多个物体的位置、名称、位置及数量的能力。可以应用到视频监控（是否有违规物体或行为）、工业质检（微小瑕疵的数量和位置）、医疗诊断（细胞计数、中药识别）等场景。在移动端预测的效果图如下：
<p align="center"><img width="250" height="250" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/dog.jpg"/>&#8194;&#8194;&#8194;&#8194;&#8194;<img width="250" height="250" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/dog2.jpg"/></p>
## iOS demo部署方法
下面我们以**目标检测（object_detection_demo)**为例讲解如何部署iOS工程。
**目的**：将基于Paddle-Lite预测库的iOS APP部署到苹果手机，实现物体检测。
**需要的环境**：Mac 电脑上安装Xcode、苹果手机、下载到本地的[Paddle-Lite-Demo](https://github.com/PaddlePaddle/Paddle-Lite-Demo)工程
**部署步骤**：
1、 目标检测的iOS示例位于 `Paddle-Lite-Demo\PaddleLite-ios-demo\object_detection_demo`
2、终端中执行 `download_dependencies.sh`脚本自动下载模型和Paddle-Lite预测库
```shell
cd PaddleLite-ios-demo # 1. 终端中进入 Paddle-Lite-Demo\PaddleLite-ios-demo
sh download_dependencies.sh # 2. 执行脚本下载依赖项 （需要联网）
```
下载完成后会出现提示： `Extract done `
3、用Xcode打开`object_detection_demo/detection_demo.xcodeproj`文件，修改工程配置。
依次修改 `General/Identity`和`Signing&Capabilities`属性，替换为自己的工程代号和团队名称。（必须修改，不然无法通过编译）
![Xcode1](https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/iOS/Xcode1.png)
![Xcode2](https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/iOS/Xcode2.png)
4、 IPhone手机连接电脑，在Xcode中连接自己的手机 （第一次连接IPhone到电脑时，需要在IPhone的`设置->通用->设备管理`中选择本电脑并信任）
<p align="center"><img width="600" height="250" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/iOS/Xcode-phone.jpg"/>
5、按下左上角的 Run按钮，自动编译APP并安装到手机。在苹果手机中设置信任该APP（进入`设置->通用->设备管理`，选中新安装的APP并`验证该应用`）
成功后效果如下，图一：APP安装到手机 图二： APP打开后的效果，会自动识别图片中的物体并标记
<p align="center"><img width="300" height="450" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/iOS/IOS2.jpeg"/>&#8194;&#8194;&#8194;&#8194;&#8194;<img width="300" height="450" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/iOS/IOS3.jpeg"/></p>
## iOS demo结构讲解
iOS 示例的代码结构如下图所示：
<p align="center"><img width="600" height="450" src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/iOS/IOS-struct.png"/>
1、 mobilenetv1-ssd： 模型文件 (opt 工具转化后Paddle-Lite模型)
```shell
# 位置：
ios-detection_demo/detection_demo/models/mobilenetv1-ssd
```
2、 libpaddle_api_light_bundled.a、paddle_api.h : Paddle-Lite C++ 预测库和头文件
```shell
# 位置：
# iOS预测库
ios-detection_demo/detection_demo/lib/libpaddle_api_light_bundled.a
# 预测库头文件
ios-detection_demo/detection_demo/include/paddle_api.h
ios-detection_demo/detection_demo/include/paddle_use_kernels.h
ios-detection_demo/detection_demo/include/paddle_use_ops.h
```
3、 ViewController.mm：主要预测代码
```shell
# 位置
ios-detection_demo/detection_demo/ViewController.mm
```
## 代码讲解 （如何使用Paddle-Lite C++ API 执行预测）
IOS 示例基于C++ API 开发，调用Paddle-Lite C++ API包括以下五步。更详细的API 描述参考： [Paddle-Lite C++ API](https://paddle-lite.readthedocs.io/zh/latest/api_reference/java_api_doc.html)。
```c++
#include <iostream>
// 引入C++ API
#include "paddle_lite/paddle_api.h"
#include "paddle_lite/paddle_use_ops.h"
#include "paddle_lite/paddle_use_kernels.h"
// 1. 设置MobileConfig
MobileConfig config;
config.set_model_from_file(<modelPath>); // 设置NaiveBuffer格式模型路径
config.set_power_mode(LITE_POWER_NO_BIND); // 设置CPU运行模式
config.set_threads(4); // 设置工作线程数
// 2. 创建PaddlePredictor
std::shared_ptr<PaddlePredictor> predictor = CreatePaddlePredictor<MobileConfig>(config);
// 3. 设置输入数据
std::unique_ptr<Tensor> input_tensor(std::move(predictor->GetInput(0)));
input_tensor->Resize({1, 3, 224, 224});
auto* data = input_tensor->mutable_data<float>();
for (int i = 0; i < ShapeProduction(input_tensor->shape()); ++i) {
data[i] = 1;
}
// 4. 执行预测
predictor->run();
// 5. 获取输出数据
std::unique_ptr<const Tensor> output_tensor(std::move(predictor->GetOutput(0)));
std::cout << "Output shape " << output_tensor->shape()[1] << std::endl;
for (int i = 0; i < ShapeProduction(output_tensor->shape()); i += 100) {
std::cout << "Output[" << i << "]: " << output_tensor->data<float>()[i]
<< std::endl;
}
```
docs/user_guides/java_demo.md → docs/demo_guides/java_demo.md
浏览文件 @ 5dd85c2c
......@@ -9,7 +9,7 @@
## 编译
首先在PaddleLite的开发 [Docker镜像](../installation/source_compile) 中，拉取最新PaddleLite代码，编译对应你手机架构的预测库，
首先在PaddleLite的开发 [Docker镜像](../user_guides/source_compile) 中，拉取最新PaddleLite代码，编译对应你手机架构的预测库，
下面我们以arm8 架构举例。进入paddlelite 目录，运行以下命令：
```shell
......
docs/advanced_user_guides/npu.md → docs/demo_guides/npu.md
浏览文件 @ 5dd85c2c
# 使用华为NPU
# PaddleLite使用NPU(华为)预测部署
Paddle Lite是首款支持华为自研达芬奇架构NPU（Kirin 810/990 SoC搭载的NPU）的预测框架。
原理是在线分析Paddle模型，将Paddle算子转成HiAI IR后，调用HiAI IR/Builder/Runtime APIs生成并执行HiAI模型。
......@@ -91,7 +91,7 @@ $ ./lite/tools/build_npu.sh --arm_os=android --arm_abi=armv7 --arm_lang=gcc --an
$ ./lite/tools/build_npu.sh --arm_os=android --arm_abi=armv7 --arm_lang=gcc --android_stl=c++_shared tiny_publish
```
注意：为了保证编译环境一致，建议参考[源码编译](../installation/source_compile)中的Docker开发环境进行配置，然后再执行上述命令。
注意：为了保证编译环境一致，建议参考[源码编译](../user_guides/source_compile)中的Docker开发环境进行配置，然后再执行上述命令。
## 优化生成NPU模型
......@@ -103,7 +103,6 @@ $ ./lite/tools/build_npu.sh --arm_os=android --arm_abi=armv7 --arm_lang=gcc --an
--optimize_out_type=(protobuf|naive_buffer) \
--optimize_out=<output_optimize_model_dir> \
--valid_targets=npu,arm \
--prefer_int8_kernel=(true|false) \
--record_tailoring_info =(true|false)
```
- model_optimize_tool生成的模型只是标记了NPU支持的Paddle算子，并没有真正生成NPU HiAI模型，只有在执行时才会将标记的Paddle算子转成HiAI IR，最终生成并执行HiAI模型，具体实现参考PR[2576](https://github.com/PaddlePaddle/Paddle-Lite/pull/2576)。
......@@ -111,13 +110,13 @@ $ ./lite/tools/build_npu.sh --arm_os=android --arm_abi=armv7 --arm_lang=gcc --an
## 通过JAVA接口加载并执行NPU模型
- 使用方法和[Java实例](../user_guides/java_demo)一致，无需额外设置任何参数，只需将模型换成NPU模型即可。[Paddle-Lite-Demo](https://github.com/PaddlePaddle/Paddle-Lite-Demo)中的Image Classification Demo for Android是同时支持CPU和NPU两种模型的图像分类Demo。
- 使用方法和[Java实例](java_demo)一致，无需额外设置任何参数，只需将模型换成NPU模型即可。[Paddle-Lite-Demo](https://github.com/PaddlePaddle/Paddle-Lite-Demo)中的Image Classification Demo for Android是同时支持CPU和NPU两种模型的图像分类Demo。
注意：在拷贝libpaddle_lite_jni.so的时候，由于依赖HiAI DDK so和libc++_shared.so库，需要将HiAI DDK中ai_ddk_lib/lib或ai_ddk_lib/lib64目录下的所有so和libc++_shared.so，拷到libpaddle_lite_jni.so同级目录下。
## 通过C++接口加载并执行NPU模型
- 使用方法和[C++实例](../user_guides/cpp_demo)一致，同样无需额外设置任何参数，只需将模型换成NPU模型即可。
- 使用方法和[C++实例](cpp_demo)一致，同样无需额外设置任何参数，只需将模型换成NPU模型即可。
注意：1）不能使用安卓模拟器，需要使用真实设备，且必须是支持NPU的华为手机。2）在使用adb push命令向手机推送目标程序时，需要将HiAI DDK中ai_ddk_lib/lib或ai_ddk_lib/lib64目录下的所有so和libc++_shared.so，推送到目标程序同级目录下。
......
docs/advanced_user_guides/opencl.md → docs/demo_guides/opencl.md
浏览文件 @ 5dd85c2c
# Lite基于OpenCL的模型预测
# PaddleLite使用OpenCL预测部署
Lite支持在Android系统上运行基于OpenCL的程序，目前支持Ubuntu环境下armv8、armv7的交叉编译。
......@@ -11,18 +11,45 @@ Lite支持在Android系统上运行基于OpenCL的程序，目前支持Ubuntu环
详见 **源码编译指南-环境准备** 章节。
### 编译选项
|参数|介绍|值|
|--------|--------|--------|
|--arm_os|代表目标操作系统|目前仅支持且默认为`android`|
|--arm_abi|代表体系结构类型，支持armv8和armv7|默认为`armv8`即arm64-v8a；`armv7`即armeabi-v7a|
|--arm_lang|代表编译目标文件所使用的编译器|默认为gcc，支持 gcc和clang两种|
### 编译Paddle-Lite OpenCL库范例
注：以android-armv8-opencl的目标、Docker容器的编译开发环境为例，CMake3.10，android-ndk-r17c位于`/opt/`目录下。
#### 针对 Lite 用户的编译命令(无单元测试,有编译产物)
- `arm_os`: `[android]`，目前不支持linux；
- `arm_abi`: `[armv7 | armv8]`；
- `arm_lang`: `[gcc]`，目前不支持clang；
- `build_extra`: `[OFF | ON]`，编译全量op和kernel，体积会大，编译时间长；
- `build_cv`: `[OFF | ON]`，编译arm cpu neon实现的的cv预处理模块；
- `android_stl`: `[c++_shared | c++_static]`，paddlelite的库以何种方式链接`android_stl`，选择`c++_shared`得到的动态库体积更小，但使用时候记得上传paddlelite所编译版本（armv7或armv8）一致的`libc++_shared.so`（来自Android-NDK）；
注：调用`./lite/tools/build.sh`执行编译。
```bash
# 假设当前位于处于Lite源码根目录下
# 导入NDK_ROOT变量，注意检查您的安装目录若与本示例不同
export NDK_ROOT=/opt/android-ndk-r17c
# 删除上一次CMake自动生成的.h文件
rm ./lite/api/paddle_use_kernels.h
rm ./lite/api/paddle_use_ops.h
# 根据指定编译参数编译
./lite/tools/build.sh \
--arm_os=android \
--arm_abi=armv8 \
--arm_lang=gcc \
--build_extra=OFF \
--build_cv=OFF \
--android_stl=c++_shared \
opencl
```
#### 针对 Lite 开发者的编译命令(有单元测试,编译产物)
注：调用`./lite/tools/ci_build.sh`执行编译，该命令会编译armv7和armv8的opencl库。虽然有编译产物，但因编译单元测试，编译产物包体积可能较大，不推荐使用。
```bash
# 假设当前位于处于Lite源码根目录下
......@@ -38,16 +65,20 @@ rm ./lite/api/paddle_use_ops.h
--arm_os=android \
--arm_abi=armv8 \
--arm_lang=gcc \
build_test_arm_opencl
build_opencl
```
注：如果要调试cl kernel，假设已经完成上述脚本编译(已生成cmake文件)。调试只需要修改`./lite/backends/opencl/cl_kernel/`下对应的kernel文件，保存后在项目根目录执行`python ./lite/tools/cmake_tools/gen_opencl_code.py ./lite/backends/opencl/cl_kernel ./lite/backends/opencl/opencl_kernels_source.cc`，该命令会自动将修改后，再切到build目录下执行`make publish_inference`或者你要编译的单测的可执行文件名，cl kernel文件的内容会随着编译自动打包到产物包如 .so 中或者对应单测可执行文件中。
### 编译产物说明
编译产物位于`build.lite.android.armv8.gcc.opencl`下的`inference_lite_lib.android.armv8.opencl`文件夹内，这里仅罗列关键产物：
- `cxx`:该目录是编译目标的C++的头文件和库文件;
- `demo`:该目录包含了两个demo，用来调用使用`libpaddle_api_full_bundled.a`和`libpaddle_api_light_bundled.a`，分别对应`mobile_full`和`mobile_light`文件夹。编译对应的demo仅需在`mobile_full`或`mobile_light`文
- `mobile_full`:使用cxx config，可直接加载fluid模型，若使用OpenCL需要在`mobilenetv1_full_api.cc`代码里开启`DEMO_USE_OPENCL`的宏，详细见代码注释;
- `mobile_light`:使用mobile config，只能加载`model_optimize_tool`优化过的模型;
- `opencl`:该目录存放opencl实现的相关kernel。
- `mobile_light`:使用mobile config，只能加载`model_optimize_tool`优化过的模型。
注：`opencl`实现的相关kernel已经打包到动态库中。
```bash
.
......@@ -65,40 +96,23 @@ rm ./lite/api/paddle_use_ops.h
| |-- libpaddle_api_light_bundled.a
| |-- libpaddle_full_api_shared.so
| `-- libpaddle_light_api_shared.so
|-- demo
| `-- cxx
| |-- Makefile.def
| |-- README.md
| |-- include
| | |-- paddle_api.h
| | |-- paddle_lite_factory_helper.h
| | |-- paddle_place.h
| | |-- paddle_use_kernels.h
| | |-- paddle_use_ops.h
| | `-- paddle_use_passes.h
| |-- mobile_full
| | |-- Makefile
| | `-- mobilenetv1_full_api.cc
| `-- mobile_light
| |-- Makefile
| `-- mobilenetv1_light_api.cc
`-- opencl
`-- cl_kernel
|-- buffer
| |-- depthwise_conv2d_kernel.cl
| |-- elementwise_add_kernel.cl
| |-- fc_kernel.cl
| |-- im2col_kernel.cl
| |-- layout_kernel.cl
| |-- mat_mul_kernel.cl
| |-- pool_kernel.cl
| `-- relu_kernel.cl
|-- cl_common.h
`-- image
|-- channel_add_kernel.cl
|-- elementwise_add_kernel.cl
|-- pool_kernel.cl
`-- relu_kernel.cl
`-- demo
`-- cxx
|-- Makefile.def
|-- README.md
|-- include
| |-- paddle_api.h
| |-- paddle_lite_factory_helper.h
| |-- paddle_place.h
| |-- paddle_use_kernels.h
| |-- paddle_use_ops.h
| `-- paddle_use_passes.h
|-- mobile_full
| |-- Makefile
| `-- mobilenetv1_full_api.cc
`-- mobile_light
|-- Makefile
`-- mobilenetv1_light_api.cc
```
调用`libpaddle_api_full_bundled.a`和`libpaddle_api_light_bundled.a`见下一部分运行示例。
......@@ -109,48 +123,9 @@ rm ./lite/api/paddle_use_ops.h
下面以android、ARMv8、gcc的环境为例，介绍3个示例，分别如何在手机上执行基于OpenCL的ARM GPU推理过程。
**注意：** 以下命令均在Lite源码根目录下运行。在3个示例前，下面这段命令都先要执行用来准备环境:
```bash
# 在/data/local/tmp目录下创建OpenCL文件目录
adb shell mkdir -p /data/local/tmp/opencl
adb shell mkdir -p /data/local/tmp/opencl/cl_kernel/buffer
adb shell mkdir -p /data/local/tmp/opencl/cl_kernel/image
# 将OpenCL的kernels文件推送到/data/local/tmp/opencl目录下
adb push lite/backends/opencl/cl_kernel/cl_common.h /data/local/tmp/opencl/cl_kernel/
adb push lite/backends/opencl/cl_kernel/buffer/* /data/local/tmp/opencl/cl_kernel/buffer/
adb push lite/backends/opencl/cl_kernel/image/* /data/local/tmp/opencl/cl_kernel/image/
```
### 运行示例1: 编译产物demo示例
```bash
######################################################################
# 编译mobile_full的demo #
######################################################################
# 步骤: #
# 0.确保编译Paddle-Lite时编译了OpenCL; #
# 1.编辑`mobilenetv1_full_api.cc`代码, 开启`DEMO_USE_OPENCL`的宏; #
# 2.在产物目录`demo/cxx/mobile_full`下编译`mobile_full`的demo; #
# 3.上传demo, 模型, opencl kernel文件到手机; #
# 4.运行demo得到预期结果. #
######################################################################
adb shell mkdir /data/local/tmp/opencl/mobilenet_v1
chmod +x ./build.lite.android.armv8.gcc.opencl/inference_lite_lib.android.armv8.opencl/demo/cxx/mobile_full/mobilenetv1_full_api
adb push ./build.lite.android.armv8.gcc.opencl/inference_lite_lib.android.armv8.opencl/demo/cxx/mobile_full/mobilenetv1_full_api /data/local/tmp/opencl/
adb push ./build.lite.android.armv8.gcc.opencl/install/mobilenet_v1/* /data/local/tmp/opencl/mobilenet_v1
# use mobile_full run mobilenet_v1
# `GLOG_v` is log level
adb shell "export GLOG_v=0; \
/data/local/tmp/opencl/mobilenetv1_full_api \
--model_dir=/data/local/tmp/opencl/mobilenet_v1 \
--optimized_model_dir=/data/local/tmp/opencl/full_api_opt_model"
######################################################################
# 编译mobile_light的demo #
######################################################################
......@@ -158,33 +133,40 @@ adb shell "export GLOG_v=0; \
# 0.确保编译Paddle-Lite时编译了OpenCL; #
# 1.编译model_optimize_tool并对模型优化, `targets`参数为`opencl`; #
# 2.在产物目录`demo/cxx/mobile_light`下编译`mobile_light`的demo; #
# 3.上传demo, 模型, opencl kernel文件到手机; #
# 3.上传demo, 模型文件到手机; #
# 4.运行demo得到预期结果. #
######################################################################
# 在/data/local/tmp目录下创建OpenCL文件目录
adb shell mkdir -p /data/local/tmp/opencl
# use model_optimize_tool to optimize model
./build.model_optimize_tool/lite/api/model_optimize_tool \
--model_dir=./build.lite.android.armv8.gcc.opencl/install/mobilenet_v1/ \
--optimize_out_type=naive_buffer \
--optimize_out=./build.lite.android.armv8.gcc.opencl/install/mobilenet_v1/ \
--optimize_out=./build.lite.android.armv8.gcc.opencl/install/mobilenet_v1/mobilenetv1_opt \
--valid_targets=opencl
adb shell mkdir /data/local/tmp/opencl/mobilenet_v1
adb shell mkdir /data/local/tmp/opencl/mobilenet_v1/
chmod +x ./build.lite.android.armv8.gcc.opencl/inference_lite_lib.android.armv8.opencl/demo/cxx/mobile_light/mobilenetv1_light_api
adb push ./build.lite.android.armv8.gcc.opencl/inference_lite_lib.android.armv8.opencl/demo/cxx/mobile_light/mobilenetv1_light_api /data/local/tmp/opencl/
adb push ./build.lite.android.armv8.gcc.opencl/install/mobilenet_v1/* /data/local/tmp/opencl/mobilenet_v1
adb push ./build.lite.android.armv8.gcc.opencl/install/mobilenet_v1/mobilenetv1_opt.nb /data/local/tmp/opencl/
# use mobile_light run mobilenet_v1
adb shell "export GLOG_v=5; \
adb shell "export GLOG_v=1; \
/data/local/tmp/opencl/mobilenetv1_light_api \
--model_dir=/data/local/tmp/opencl/"
/data/local/tmp/opencl/mobilenetv1_opt.nb"
```
**注：** `GLOG_v`是指定需要显示VLOG的日志级别，默认为0。权重参数会在第一次运行时加载，所以第一次执行时间略长。一般将warmup的值设为10，repeats值设为多次。
### 运行示例2: test_mobilenetv1单元测试
- **运行文件准备**
```bash
# 在/data/local/tmp目录下创建OpenCL文件目录
adb shell mkdir -p /data/local/tmp/opencl
# 将mobilenet_v1的模型文件推送到/data/local/tmp/opencl目录下
adb shell mkdir -p /data/local/tmp/opencl/mobilenet_v1
adb push build.lite.android.armv8.gcc.opencl/third_party/install/mobilenet_v1/* /data/local/tmp/opencl/mobilenet_v1/
......@@ -195,41 +177,25 @@ adb push build.lite.android.armv8.gcc.opencl/lite/api/test_mobilenetv1 /data/loc
- **执行OpenCL推理过程**
使用如下命令运行OpenCL程序。其中：
- `--cl_path`指定了OpenCL的kernels文件即cl\_kernel所在目录；
- `--modle_dir`指定了模型文件所在目录。
```bash
adb shell chmod +x /data/local/tmp/opencl/test_mobilenetv1
adb shell /data/local/tmp/opencl/test_mobilenetv1 \
--cl_path=/data/local/tmp/opencl \
--model_dir=/data/local/tmp/opencl/mobilenet_v1 \
--warmup=1 \
--repeats=1
adb shell "export GLOG_v=1; \
/data/local/tmp/opencl-image/test_mobilenetv1 \
--model_dir=/data/local/tmp/opencl-image/mobilenetv1_fluid/ \
--warmup=10 \
--repeats=100"
```
**注意：** 因为权重参数均会在Op Kernel第一次运行时进行加载，所以第一次的执行时间会略长。一般将warmup的值设为1，repeats值设为多次。
### 运行示例3: test_layout_opencl单元测试
- **运行文件准备**
```bash
# 将OpenCL单元测试程序test_layout_opencl，推送到/data/local/tmp/opencl目录下
adb push build.lite.android.armv8.gcc.opencl/lite/kernels/opencl/test_layout_opencl /data/local/tmp/opencl/
```
OpenCL推理过程**
```bash
adb shell mkdir -p /data/local/tmp/opencl
adb shell chmod +x /data/local/tmp/opencl/test_layout_opencl
adb shell /data/local/tmp/opencl/test_layout_opencl
adb shell "export GLOG_v=4; \
/data/local/tmp/opencl/test_layout_opencl"
```
### 如何在Code中使用
见运行示例1的demo代码:
......
docs/advanced_user_guides/x86.md → docs/demo_guides/x86.md
浏览文件 @ 5dd85c2c
# 使用X86预测库
# PaddleLite使用X86预测部署
Paddle-Lite 支持在Docker或Linux环境编译x86预测库。环境搭建参考[环境准备](../installation/source_compile)。
Paddle-Lite 支持在Docker或Linux环境编译x86预测库。环境搭建参考[环境准备](../user_guides/source_compile)。
(注意：非docker Linux环境需要是Ubuntu16.04)
......@@ -9,8 +9,8 @@ Paddle-Lite 支持在Docker或Linux环境编译x86预测库。环境搭建参考
1、 下载代码
```bash
git clone https://github.com/PaddlePaddle/Paddle-Lite.git
#需要切换到 release/v2.0.0之后版本
git checkout <release_tag>
# 切换到release分支
git checkout release/v2.3
```
2、 源码编译
......@@ -42,43 +42,56 @@ x86编译结果位于 `build.lite.x86/inference_lite_lib`
## x86预测API使用示例
1、我们提供Linux环境下x86 API运行mobilenet_v1的示例：[mobilenet_full_x86demo](https://paddlelite-data.bj.bcebos.com/x86/mobilenet_full_x86demo.zip)。下载解压后内容如下：
![](https://paddlelite-data.bj.bcebos.com/x86/x86-doc/demo.png)
`mobilenet_v1`为模型文件、`lib`和`include`分别是Paddle-Lite的预测库和头文件、`third_party`下是编译时依赖的第三方库`mklml`、`mobilenet_full_api.cc`是x86示例的源代码、`build.sh`为编译的脚本。
2、demo内容与使用方法
``` bash
# 1、编译
sh build.sh
```
编译结果为当前目录下的 `mobilenet_full_api `
``` bash
# 2、执行预测
mobilenet_full_api mobilenet_v1
```
`mobilenet_v1`为当前目录下的模型路径，`mobilenet_full_api`为第一步编译出的可执行文件。
3、示例源码`mobilenet_full_api.cc`
```c++
#include <gflags/gflags.h>
#include <iostream>
#include <vector>
#include "paddle_api.h" // NOLINT
#include "paddle_use_kernels.h" // NOLINT
#include "paddle_use_ops.h" // NOLINT
#include "paddle_use_passes.h" // NOLINT
#include "paddle_api.h"
using namespace paddle::lite_api; // NOLINT
DEFINE_string(model_dir, "", "Model dir path.");
DEFINE_string(optimized_model_dir, "", "Optimized model dir.");
DEFINE_bool(prefer_int8_kernel, false, "Prefer to run model with int8 kernels");
using namespace paddle::lite_api; // NOLINT
int64_t ShapeProduction(const shape_t& shape) {
int64_t res = 1;
for (auto i : shape) res *= i;
return res;
}
void RunModel() {
// 1. Set CxxConfig
CxxConfig config;
config.set_model_file(FLAGS_model_dir + "model");
config.set_param_file(FLAGS_model_dir + "params");
config.set_valid_places({
lite_api::Place{TARGET(kX86), PRECISION(kFloat)}
});
void RunModel(std::string model_dir) {
// 1. Create CxxConfig
CxxConfig config;
config.set_model_dir(model_dir);
config.set_valid_places({
Place{TARGET(kX86), PRECISION(kFloat)},
Place{TARGET(kHost), PRECISION(kFloat)}
});
// 2. Create PaddlePredictor by CxxConfig
std::shared_ptr<PaddlePredictor> predictor =
CreatePaddlePredictor<CxxConfig>(config);
// 3. Prepare input data
std::unique_ptr<Tensor> input_tensor(std::move(predictor->GetInput(0)));
input_tensor->Resize(shape_t({1, 3, 224, 224}));
input_tensor->Resize({1, 3, 224, 224});
auto* data = input_tensor->mutable_data<float>();
for (int i = 0; i < ShapeProduction(input_tensor->shape()); ++i) {
data[i] = 1;
......@@ -90,15 +103,21 @@ void RunModel() {
// 5. Get output
std::unique_ptr<const Tensor> output_tensor(
std::move(predictor->GetOutput(0)));
std::cout << "Output dim: " << output_tensor->shape()[1] << std::endl;
std::cout << "Output shape " << output_tensor->shape()[1] << std::endl;
for (int i = 0; i < ShapeProduction(output_tensor->shape()); i += 100) {
std::cout << "Output[" << i << "]:" << output_tensor->data<float>()[i] << std::endl;
std::cout << "Output[" << i << "]: " << output_tensor->data<float>()[i]
<< std::endl;
}
}
int main(int argc, char** argv) {
google::ParseCommandLineFlags(&argc, &argv, true);
RunModel();
if (argc < 2) {
std::cerr << "[ERROR] usage: ./" << argv[0] << " naive_buffer_model_dir\n";
exit(1);
}
std::string model_dir = argv[1];
RunModel(model_dir);
return 0;
}
```
docs/advanced_user_guides/add_layout.md → docs/develop_guides/add_layout.md
浏览文件 @ 5dd85c2c
# 如何增加Layout
# 新增Layout
Paddle-Lite中Place包含了Target、Layout、Precision信息，用来注册和选择模型中的具体Kernel。下面以增加Place中的layout：`ImageDefault`、`ImageFolder`、`ImageNW`为例，讲解如何增加新Layout。
......
docs/advanced_user_guides/add_new_pass.md → docs/develop_guides/add_new_pass.md
浏览文件 @ 5dd85c2c
# 新增Pass方法
# 新增Pass
本文从三个方面介绍了`Lite`中的`Pass`结构：**Pass是什么**、**Pass的实现与接口**、**Pass的一般注册流程**。最后以`Fc_fuse_pass`为例介绍了`fusion_pass`的作用与注册方法。
......
docs/advanced_user_guides/add_operation.md → docs/develop_guides/add_operation.md
浏览文件 @ 5dd85c2c
# 新增OP的方法
# 新增OP
以下以添加argmax为例，详细说明新增op的方法。
......
docs/index.rst
浏览文件 @ 5dd85c2c
......@@ -14,11 +14,11 @@ Welcome to Paddle-Lite's documentation!
introduction/tech_highlights
introduction/architecture
introduction/support_hardware
introduction/roadmap
introduction/support_operation_list
.. toctree::
:maxdepth: 1
:caption: Benchmark数据和方法
:caption: Benchmark
:name: sec-benchmark
benchmark/benchmark
......@@ -26,60 +26,60 @@ Welcome to Paddle-Lite's documentation!
.. toctree::
:maxdepth: 1
:caption: 安装
:name: sec-install
installation/release_lib
installation/source_compile
.. toctree::
:maxdepth: 1
:caption: 使用指南
:caption: 使用方法
:name: sec-user-guides
user_guides/tutorial
user_guides/release_lib
user_guides/source_compile
user_guides/x2paddle
user_guides/model_optimize_tool
user_guides/cpp_demo
user_guides/java_demo
user_guides/android_ios_app_demo
user_guides/post_quant_with_data
user_guides/post_quant_no_data
user_guides/model_quantization
user_guides/debug
user_guides/library_tailoring
.. toctree::
:maxdepth: 1
:caption: 进阶使用指南
advanced_user_guides/x2paddle
advanced_user_guides/x2paddle_models_doc
advanced_user_guides/post_quant_with_data
advanced_user_guides/post_quant_no_data
advanced_user_guides/model_quantization
advanced_user_guides/support_operation_list
advanced_user_guides/add_operation
advanced_user_guides/add_layout
advanced_user_guides/add_new_pass
advanced_user_guides/test_tools
advanced_user_guides/debug_tools
advanced_user_guides/npu
advanced_user_guides/opencl
advanced_user_guides/fpga
advanced_user_guides/cuda
advanced_user_guides/x86
advanced_user_guides/cv
:caption: 部署示例
:name: sec-demo_guides
demo_guides/cpp_demo
demo_guides/java_demo
demo_guides/android_app_demo
demo_guides/ios_app_demo
demo_guides/x86
demo_guides/cuda
demo_guides/opencl
demo_guides/fpga
demo_guides/npu
.. toctree::
:maxdepth: 1
:caption: 开发者文档
:caption: API文档
api_reference/cxx_api_doc
api_reference/java_api_doc
api_reference/python_api_doc
api_reference/cv
.. toctree::
:maxdepth: 1
:caption: 开发者贡献
develop_guides/for-developer
develop_guides/architecture-intro
develop_guides/add_operation
develop_guides/add_layout
develop_guides/add_new_pass
.. toctree::
:maxdepth: 1
:caption: API文档
:caption: Roadmap
:name: sec-roadmap
api_reference/cxx_api_doc
api_reference/java_api_doc
api_reference/python_api_doc
introduction/roadmap
.. toctree::
:maxdepth: 1
......
docs/introduction/faq.md
浏览文件 @ 5dd85c2c
# FAQ
# FAQ 常见问题
问题或建议可以发Issue，为加快问题解决效率，可先检索是否有类似问题，我们也会及时解答！
欢迎加入Paddle-Lite百度官方QQ群：696965088
1. 在Host端采用交叉编译方式编译PaddleLite，将编译后的libpaddle_light_api_shared.so和可执行程序放到板卡上运行，出现了如下图所示的错误，怎么解决？
![host_target_compiling_env_miss_matched](https://user-images.githubusercontent.com/9973393/75761527-31b8b700-5d74-11ea-8a9a-0bc0253ee003.png)
- 原因是Host端的交叉编译环境与Target端板卡的运行环境不一致，导致libpaddle_light_api_shared.so链接的GLIBC库高于板卡环境的GLIBC库。目前有四种解决办法（为了保证编译环境与官方一致，推荐第一种方式）：1）在Host端，参考[源码编译](../user_guides/source_compile)中的Docker方式重新编译libpaddle_light_api_shared.so；2）在Host端，使用与Target端版本一致的ARM GCC和GLIBC库重新编译libpaddle_light_api_shared.so；3）在Target端板卡上，参考[源码编译](../user_guides/source_compile)中的ARM Linux本地编译方式重新编译libpaddle_light_api_shared.so；4）在Target端板卡上，将GLIBC库升级到和Host端一致的版本，即GLIBC2.27。
docs/introduction/roadmap.md
浏览文件 @ 5dd85c2c
......@@ -23,5 +23,10 @@
- 增加 INT8 量化，从 PaddleSlim 训练到 PaddleLite 部署完整案例
- 支持内存中加载模型，以支持 APP 的简易加密
## 2.0.0-rc ?
预计发布 *2019-9-16~7days*
## 2.3
[v2.3 project](https://github.com/PaddlePaddle/Paddle-Lite/milestone/3?closed=1)
## 2.6
[v2.6 project](https://github.com/PaddlePaddle/Paddle-Lite/milestones/v2.6)
docs/introduction/support_hardware.md
浏览文件 @ 5dd85c2c
# 支持硬件列表
# 支持硬件
## ARM CPU
......
docs/advanced_user_guides/support_operation_list.md → docs/introduction/support_operation_list.md
浏览文件 @ 5dd85c2c
# 支持OP列表
# 支持OP
## Ops （共计158个算子）
......@@ -110,6 +110,7 @@
- equal
- gather
- generate_proposals
- graph_op
- greater_equal
- greater_than
- gru
......@@ -153,7 +154,6 @@
- sequence_expand
- sequence_expand_as
- sequence_pool
- sequence_pool_concat
- sequence_reshape
- sequence_reverse
- sequence_softmax
......@@ -226,7 +226,6 @@
- generate_proposals
- greater_equal
- greater_than
- grid_sampler
- gru
- gru_unit
- hard_sigmoid
......@@ -312,9 +311,6 @@
- gelu
- gru
- layer_norm
- leaky_relu
- lookup_table
- lookup_table_v2
- match_matrix_tensor
- matmul
- mul
......@@ -392,11 +388,9 @@
- yolo_box
### OpenCL kernels
- concat
- conv2d
- depthwise_conv2d
- elementwise_add
- elementwise_mul
- fc
- fusion_elementwise_add_activation
- layout
......@@ -404,10 +398,5 @@
- io_copy
- io_copy_once
- mul
- nearest_interp
- pool2d
- relu
- reshape
- reshape2
- scale
- sigmoid
docs/user_guides/android_ios_app_demo.md 已删除 100644 → 0
浏览文件 @ 1a56bcba
# Android/IOS APP demo
请参考[Paddle-Lite-Demo](https://github.com/PaddlePaddle/Paddle-Lite-Demo)。
docs/advanced_user_guides/debug_tools.md → docs/user_guides/debug.md
浏览文件 @ 5dd85c2c
# 调试方法
# 调试
## Profiler工具
Basic profiler 用于 CPU 上kernel 耗时的统计。
### 开启方法:
参照 [编译安装](../user_guides/source_compile) 中的**full_publish**部分进行环境配置，在 cmake 时添加 `-DLITE_WITH_PROFILE=ON` ，就可以开启相应支持。
### 使用示例：
在模型执行完毕后，会自动打印类似如下 profiler 的日志
```
kernel average min max count
feed/def/1/4/2 0 0 0 1
conv2d/def/4/1/1 1175 1175 1175 1
conv2d/def/4/1/1 1253 1253 1253 1
depthwise_conv2d/def/4/1/1 519 519 519 1
conv2d/def/4/1/1 721 721 721 1
elementwise_add/def/4/1/1 18 18 18 1
conv2d/def/4/1/1 2174 2174 2174 1
depthwise_conv2d/def/4/1/1 380 380 380 1
conv2d/def/4/1/1 773 773 773 1
elementwise_add/def/4/1/1 2 2 2 1
conv2d/def/4/1/1 1248 1248 1248 1
depthwise_conv2d/def/4/1/1 492 492 492 1
conv2d/def/4/1/1 1150 1150 1150 1
elementwise_add/def/4/1/1 33 33 33 1
elementwise_add/def/4/1/1 3 3 3 1
conv2d/def/4/1/1 1254 1254 1254 1
depthwise_conv2d/def/4/1/1 126 126 126 1
```
## Debug工具
**Lite Model Debug Tool** 是用来检查Paddle-Lite框架与Paddle-Fluid框架运行时tensor(包括variable与weight)之间diff信息的基础工具。
## 编译方法:
### 编译方法:
1. 参照 [编译安装](../installation/source_compile) 中的**full_publish**部分进行环境配置和编译。
1. 参照 [编译安装](../user_guides/source_compile) 中的**full_publish**部分进行环境配置和编译。
2. 在生成的`build`目录下，执行`make lite_model_debug_tool`，`lite_model_debug_tool`产出在编译目录的`lite/tools/debug`目录下。
## 工作流程:
### 工作流程:
1. 运行 `/bin/bash check_model.sh --model_dir=<your_model_path> --build_root_dir=<your_cmake_root_dir> debug_cpp_stage` 获得模型在Paddle-Lite框架下的运行拓扑信息、varibles信息和weights信息。运行后拓扑信息将会存储在默认名为 `topo_file.txt` 的文件中，variables和weights信息将会存储在默认名为 `tensor_cpp.txt` 的文件中。
2. 运行 `/bin/bash check_model.sh --model_dir=<your_model_path> --build_root_dir=<your_cmake_root_dir> debug_py_stage`执行fluid框架预测以获取相同模型在fluid框架下的variable与weight信息(注意：我们使用fluid的python api运行fluid模型，因此您在运行此步之前应确保已正确安装fluid的python api)。然后debug tool将会自动比较Paddle-Lite框架输出的信息和Paddle-Fluid框架输出的信息来检查是否存在运行时diff。 执行Paddle-Fluid框架，输出的信息将会存储在默认名为 `tensor_py.txt` 的文件中，相应的diff信息将会存储在默认名为 `diff.txt`的文件中(默认情况下，只会输出执行拓扑序中第一个有diff的variable相关的信息)。
## 注意事项:
### 注意事项:
1. 输出的结果是在**执行完一次预测后**输出的相应变量/权重的最终值，因此如果您在预测过程进行过诸如变量复用/子图融合等优化方法，则相应的输出可能会出现偏差。
2. 默认情况下debug tools将以全1作为输入进行比对。
3. 默认情况下，为了保证与Paddle-Fluid框架的结果可比对，debug tool将会禁用掉所有的Paddle-Lite的优化策略。
4. Paddle-Lite框架的执行环境由与您的编译选项有关，比如您开启了LITE_WITH_ARM编译选项，那debug tool的`debug_cpp_stage`也需要在ARM平台下运行。
## Diff信息输出：
### Diff信息输出：
如果debug tool检测到diff信息，那么在`diff.txt`中将会输出类似以下结构信息
......@@ -37,7 +72,7 @@ dropout_0.tmp_0 {1,1536,1,1} 0.599913 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.01202671
其中第二行为op相关信息，标明了执行哪个op出现了diff及其对应的输入输出变量名。Tensor File info为Paddle-Lite框架的输出信息，而Fluid Tensor info为Paddle-Fluid框架的相应输出信息。
示例中的`dropout_0.tmp_1`没有相应的tensor信息是因为工具检测到其在预测的后序流程中未被使用，因此不会对预测结果造成影响，从而将其自动屏蔽掉以保证输出尽量简洁。
## 其他选项：
### 其他选项：
| Option | Description |
| --------------------------- | ------------------------------------------------------------ |
......
docs/user_guides/library_tailoring.md
浏览文件 @ 5dd85c2c
# 裁剪预测库方法
# 裁剪预测库
Paddle-Lite支持**根据模型裁剪预测库**功能。Paddle-Lite的一般编译会将所有已注册的operator打包到预测库中，造成库文件体积膨胀；**裁剪预测库**能针对具体的模型，只打包优化后该模型需要的operator，有效降低预测库文件大小。
......@@ -39,7 +39,7 @@ Paddle-Lite支持**根据模型裁剪预测库**功能。Paddle-Lite的一般编
例如：
```bash
./lite/tools/build.sh --arm_os=android --arm_abi=armv7 --arm_lang=gcc --android_stl=c++_static --build_extra=ON --build_tailor=ON --opt_model_dir=../mobilenet_v1NB full_publish
./lite/tools/build.sh --arm_os=android --arm_abi=armv7 --arm_lang=gcc --android_stl=c++_static --build_extra=ON --build_tailor=ON --opt_model_dir=../mobilenet_v1NB tiny_publish
```
**注意**：上面命令中的`../mobilenet_v1NB`是第1步得到的转化模型的输出路径
......@@ -88,9 +88,6 @@ Paddle-Lite支持**根据模型裁剪预测库**功能。Paddle-Lite的一般编
#include <stdio.h>
#include <vector>
#include "paddle_api.h" // NOLINT
#include "paddle_use_kernels.h" // NOLINT
#include "paddle_use_ops.h" // NOLINT
#include "paddle_use_passes.h" // NOLINT
using namespace paddle::lite_api; // NOLINT
......@@ -182,4 +179,4 @@ int main(int argc, char** argv) {
1. 模型集合**必须**均为combined参数模型或均为非combined参数模型。
2. 使用非combined参数模型时，模型拓扑文件名应为`__model__`，使用非combined参数模型时，集合中各模型的拓扑与参数名应相同，分别由`--model_filename`和`--param_filename`指定。
3. 模型集合**必须**均为INT8量化模型或均为非INT8量化模型。
4. 需要使用Paddle-Lite 最新版本（release/v2.1.0之后）代码编译出的model_optimize_tool。
4. 需要使用Paddle-Lite `release/v2.1.0`之后版本代码编译出的模型优化工具。
docs/user_guides/model_optimize_tool.md
浏览文件 @ 5dd85c2c
# 模型转化方法
# 模型优化工具 opt
Paddle-Lite 提供了多种策略来自动优化原始的训练模型，其中包括量化、子图融合、混合调度、Kernel优选等等方法。为了使优化过程更加方便易用，我们提供了**opt** 工具来自动完成优化步骤，输出一个轻量的、最优的可执行模型。
......@@ -12,9 +12,12 @@ Paddle-Lite 提供了多种策略来自动优化原始的训练模型，其中
1. **推荐！** 可以进入Paddle-Lite Github仓库的[release界面](https://github.com/PaddlePaddle/Paddle-Lite/releases)，选择release版本下载对应的转化工具`opt`
(release/v2.2.0之前的转化工具为model_optimize_tool、release/v2.3.0之后为opt)
2. 本文提供`release/v2.3`和`release/v2.2.0`版本的优化工具下载
2. 我们提供`release/v2.3`编译结果下载：[opt](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.3.0/opt)、[opt_mac](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.3.0/opt_mac)
`release/v2.2.0`版本的model_optimize_tool: [model_optimize_tool](https://paddlelite-data.bj.bcebos.com/model_optimize_tool/model_optimize_tool)、[model_optimize_tool_mac](https://paddlelite-data.bj.bcebos.com/model_optimize_tool/model_optimize_tool_mac)
|版本 | Linux | MacOS|
|---|---|---|
| `release/v2.3`| [opt](https://paddlelite-data.bj.bcebos.com/model_optimize_tool/opt) | [opt_mac](https://paddlelite-data.bj.bcebos.com/model_optimize_tool/opt_mac) |
|`release/v2.2.0` | [model_optimize_tool](https://paddlelite-data.bj.bcebos.com/model_optimize_tool/model_optimize_tool) | [model_optimize_tool_mac](https://paddlelite-data.bj.bcebos.com/model_optimize_tool/model_optimize_tool_mac) |
3. 如果 release 列表里的工具不符合您的环境，可以下载Paddle-Lite 源码，源码编译出opt工具
......@@ -25,7 +28,7 @@ git checkout <release-version-tag>
./lite/tools/build.sh build_optimize_tool
```
编译结果位于`Paddle-Lite/build.opt/lite/api/opt`
**注意**：从源码编译opt前需要先[安装Paddle-Lite的开发环境](../installation/source_compile)。
**注意**：从源码编译opt前需要先[安装Paddle-Lite的开发环境](source_compile)。
## 使用opt
......@@ -80,7 +83,6 @@ PaddlePaddle模型有两种保存格式：
--optimize_out_type=(protobuf|naive_buffer) \
--optimize_out=<output_optimize_model_dir> \
--valid_targets=(arm|opencl|x86|npu|xpu) \
--prefer_int8_kernel=(true|false) \
--record_tailoring_info =(true|false)
```
......@@ -92,12 +94,12 @@ PaddlePaddle模型有两种保存格式：
| --optimize_out_type | 输出模型类型，目前支持两种类型：protobuf和naive_buffer，其中naive_buffer是一种更轻量级的序列化/反序列化实现。若您需要在mobile端执行模型预测，请将此选项设置为naive_buffer。默认为protobuf。 |
| --optimize_out | 优化模型的输出路径。 |
| --valid_targets | 指定模型可执行的backend，默认为arm。目前可支持x86、arm、opencl、npu、xpu，可以同时指定多个backend(以空格分隔)，Model Optimize Tool将会自动选择最佳方式。如果需要支持华为NPU（Kirin 810/990 Soc搭载的达芬奇架构NPU），应当设置为npu, arm。 |
| --prefer_int8_kernel | 若待优化模型为int8量化模型（如量化训练得到的量化模型），则设置该选项为true以使用int8内核函数进行推理加速，默认为false。 |
| --record_tailoring_info | 当使用 [根据模型裁剪库文件](./library_tailoring.html) 功能时，则设置该选项为true，以记录优化后模型含有的kernel和OP信息，默认为false。 |
* 如果待优化的fluid模型是非combined形式，请设置`--model_dir`，忽略`--model_file`和`--param_file`。
* 如果待优化的fluid模型是combined形式，请设置`--model_file`和`--param_file`，忽略`--model_dir`。
* 优化后的模型为以`.nb`名称结尾的单个文件。
* 删除`prefer_int8_kernel`的输入参数，`opt`自动判别是否是量化模型，进行相应的优化操作。
### 功能二：统计模型算子信息、判断是否支持
......
docs/advanced_user_guides/model_quantization.md → docs/user_guides/model_quantization.md
浏览文件 @ 5dd85c2c
......@@ -237,7 +237,7 @@ python compress.py \
接下来，使用原始的量化模型生成适合在移动端直接部署的模型。
参考[源码编译](../source_compile)配置编译环境，确保可以编译成功。参考[模型转化方法](../model_optimize_tool)，首先编译model_optimize_tool工具，然后执行下面命令对量化训练的模型进行优化（注意，需要自行修改model_file、param_file和optimize_out）。
参考[源码编译](source_compile)配置编译环境，确保可以编译成功。参考[模型转化方法](model_optimize_tool)，首先编译model_optimize_tool工具，然后执行下面命令对量化训练的模型进行优化（注意，需要自行修改model_file、param_file和optimize_out）。
```bash
./model_optimize_tool \
--model_file=mobilenet_v1_quant/float/model \
......@@ -245,7 +245,6 @@ python compress.py \
--optimize_out_type=naive_buffer \
--optimize_out=mobilenet_v1_quant_opt \
--valid_targets=arm \
--prefer_int8_kernel=true
```
如前所述，量化训练后，float目录下的模型参数范围为int8，但参数数据类型仍为float32类型，这样确实没有起到模型参数压缩的效果。但是，经过model\_optimize\_tool工具优化后对应的量化参数均会以int8类型重新存储达到参数压缩的效果，且模型结构也被优化（如进行了各种operator fuse操作）。
......@@ -260,7 +259,7 @@ adb push mobilenet_v1_quant_opt /data/local/tmp
### 使用mobilenetv1\_light\_api运行优化后的量化模型
参考[源码编译](../source_compile)配置编译环境后，在Paddle-Lite执行如下命令获取轻量级API的demo：
参考[源码编译](source_compile)配置编译环境后，在Paddle-Lite执行如下命令获取轻量级API的demo：
```bash
cd /Paddle-Lite/build.lite.android.armv8.gcc/inference_lite_lib.android.armv8/demo/cxx/mobile_light
......@@ -288,7 +287,7 @@ Output[700]: 0.002509
Output[800]: 0.000538
Output[900]: 0.000969
```
在C++中使用Paddle-Lite API的方法请猛戳[此处](../cpp_demo)，用户也可参考[mobilenetv1_light_api.cc](https://github.com/PaddlePaddle/Paddle-Lite/blob/develop/lite/demo/cxx/mobile_light/mobilenetv1_light_api.cc)的代码示例。
在C++中使用Paddle-Lite API的方法请猛戳[此处](../demo_guides/cpp_demo)，用户也可参考[mobilenetv1_light_api.cc](https://github.com/PaddlePaddle/Paddle-Lite/blob/develop/lite/demo/cxx/mobile_light/mobilenetv1_light_api.cc)的代码示例。
## FAQ
......
docs/advanced_user_guides/post_quant_no_data.md → docs/user_guides/post_quant_no_data.md
浏览文件 @ 5dd85c2c
......@@ -37,11 +37,13 @@
对于调用无校准数据训练后量化，首先给出一个例子。
```python
from paddle.fluid.contrib.slim.quantization import WeightQuantization
model_dir = path/to/fp32_model_params
save_model_dir = path/to/save_model_path
weight_quant = WeightQuantization(model_dir=model_dir)
weight_quant.quantize_weight_to_int(save_model_dir=save_model_dir,
weight_bits=8,
weight_bits=16,
quantizable_op_type=['conv2d', 'depthwise_conv2d', 'mul'])
```
......@@ -73,14 +75,17 @@ WeightQuantization.quantize_weight_to_int(save_model_dir,
## 3 量化模型预测
首先，使用PaddleLite提供的模型转换工具（model_optimize_tool）将量化模型转换成移动端预测的模型，然后加载转换后的模型进行预测部署。
目前，对于无校准数据训练后量化产出的量化模型，不支持PaddlePaddle加载执行，只能使用PaddleLite进行预测部署。
很简单，首先使用PaddleLite提供的模型转换工具（opt）将量化模型转换成移动端预测的模型，然后加载转换后的模型进行预测部署。
注意，PaddleLite 2.3版本才支持无校准数据训练后量化产出的量化，所以转换工具和预测库必须是2.3及之后的版本。
### 3.1 模型转换
参考[模型转换](../user_guides/model_optimize_tool)准备模型转换工具，建议从Release页面下载。
参考[模型转换](../user_guides/model_optimize_tool)使用模型转换工具。
因为该模型会将量化的权重反量化，然后实际加载并执行FP32预测模型，所以opt命令的输入参数--prefer_int8_kernel不需要设置为true，同时其他参数按照实际情况参考文档设置。
比如在安卓手机ARM端进行预测，模型转换的命令为：
```bash
./opt --model_dir=./mobilenet_v1_quant \
......@@ -91,4 +96,4 @@ WeightQuantization.quantize_weight_to_int(save_model_dir,
### 3.2 量化模型预测
和FP32模型一样，转换后的量化模型可以在Android/IOS APP中加载预测，建议参考[C++ Demo](../user_guides/cpp_demo)、[Java Demo](../user_guides/java_demo)、[Android/IOS Demo](../user_guides/android_ios_app_demo)。
和FP32模型一样，转换后的量化模型可以在Android/IOS APP中加载预测，建议参考[C++ Demo](../demo_guides/cpp_demo)、[Java Demo](../demo_guides/java_demo)、[Android/IOS Demo](../demo_guides/android_app_demo)。
docs/advanced_user_guides/post_quant_with_data.md → docs/user_guides/post_quant_with_data.md
浏览文件 @ 5dd85c2c
......@@ -41,7 +41,7 @@
### 2.3 配置校准数据生成器
有校准数据训练后量化内部使用异步数据读取的方式读取校准数据，大家只需要根据模型的输入，配置读取数据的sample_generator。sample_generator是Python生成器，**必须每次返回单个样本数据**，会用作`DataLoader.set_sample_generator()`的数据源。
建议参考[异步数据读取文档](https://www.paddlepaddle.org.cn/documentation/docs/zh/user_guides/howto/prepare_data/use_py_reader.html)和本文示例，学习如何配置校准数据生成器。
建议参考[异步数据读取文档](https://www.paddlepaddle.org.cn/documentation/docs/zh/advanced_guide/data_preparing/use_py_reader.html)和本文示例，学习如何配置校准数据生成器。
### 2.4 调用有校准数据训练后量化
......@@ -147,18 +147,17 @@ with fluid.name_scope('skip_quant'):
参考[模型转换](../user_guides/model_optimize_tool)准备模型转换工具，建议从Release页面下载。
参考[模型转换](../user_guides/model_optimize_tool)使用模型转换工具。注意opt命令的输入参数--prefer_int8_kernel必须设置为true，其他参数按照实际情况参考文档设置。比如在安卓手机ARM端进行预测，模型转换的命令为：
参考[模型转换](../user_guides/model_optimize_tool)使用模型转换工具，参数按照实际情况设置。比如在安卓手机ARM端进行预测，模型转换的命令为：
```bash
./opt --model_dir=./mobilenet_v1_quant \
--optimize_out_type=naive_buffer \
--optimize_out=mobilenet_v1_quant_opt \
--valid_targets=arm \
--prefer_int8_kernel=true
--valid_targets=arm
```
### 3.2 量化模型预测
和FP32模型一样，转换后的量化模型可以在Android/IOS APP中加载预测，建议参考[C++ Demo](../user_guides/cpp_demo)、[Java Demo](../user_guides/java_demo)、[Android/IOS Demo](../user_guides/android_ios_app_demo)。
和FP32模型一样，转换后的量化模型可以在Android/IOS APP中加载预测，建议参考[C++ Demo](../demo_guides/cpp_demo)、[Java Demo](../demo_guides/java_demo)、[Android/IOS Demo](../demo_guides/android_app_demo)。
## 4 使用示例
......
docs/installation/release_lib.md → docs/user_guides/release_lib.md
浏览文件 @ 5dd85c2c
# 官方 release 预编译库
# 预编译库
## 编译版本介绍
......@@ -11,7 +11,7 @@
- arm_stl=`c++_static/c++_shared` Lite预测库链接STL库的方式，支持静态或动态链接
- build_extra=`ON/OFF` 是否编译全量OP，OFF时只编译CV相关基础OP，[参数详情](./library.html)
- build_extra=`ON/OFF` 是否编译全量OP，OFF时只编译CV相关基础OP，[参数详情](library)
- `tiny_publish/full_publish` 编译模式，`tiny_publish`编译移动端部署库、`full_publish`编译部署库的同时编译第三方依赖库
......@@ -52,8 +52,8 @@
| 运行系统 | 下载 |
| :---------: | :--------------: |
| Linux | [release/v2.3](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.3.0/opt) |
| MacOs | [release/v2.3](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.3.0/opt_mac) |
| Linux | [release/v2.3](https://paddlelite-data.bj.bcebos.com/model_optimize_tool/opt) |
| MacOs | [release/v2.3](https://paddlelite-data.bj.bcebos.com/model_optimize_tool/opt_mac) |
......@@ -63,7 +63,7 @@
- [Android源码编译](./source_compile.html#paddlelite)
- [iOS源码编译](./source_compile.html#paddlelite)
- [ArmLinux源码编译](./source_compile.html#paddlelite)
- [x86源码编译](../advanced_user_guides/x86)
- [opencl源码编译](../advanced_user_guides/opencl)
- [CUDA源码编译](../advanced_user_guides/cuda)
- [FPGA源码编译](../advanced_user_guides/fpga)
- [x86源码编译](../demo_guides/x86)
- [opencl源码编译](../demo_guides/opencl)
- [CUDA源码编译](../demo_guides/cuda)
- [FPGA源码编译](../demo_guides/fpga)
docs/installation/source_compile.md → docs/user_guides/source_compile.md
浏览文件 @ 5dd85c2c
# 源码编译
# 预测库编译
Paddle-Lite 提供了移动端的一键源码编译脚本 `lite/tools/build.sh`，编译流程如下：
PaddleLite已经提供官方Release预测库下载，请参考[文档](release_lib)。
PaddleLite 提供了移动端的一键源码编译脚本 `lite/tools/build.sh`，编译流程如下：
1. 环境准备（选择其一）：Docker交叉编译环境、Linux交叉编译环境
2. 编译：调用`build.sh`脚本一键编译
......@@ -234,6 +236,8 @@ brew cask install java
## 二、编译PaddleLite
**注：编译OpenCL、华为NPU、FPGA、CUDA、X86预测库、CV模块，见进阶使用指南的对应章节。**
### 下载代码
```shell
......
docs/user_guides/tutorial.md
浏览文件 @ 5dd85c2c
......@@ -24,8 +24,7 @@ $ ./opt \
--param_file=<param_path> \
--optimize_out_type=(protobuf|naive_buffer) \
--optimize_out=<output_optimize_model_dir> \
--valid_targets=(arm|opencl|x86) \
--prefer_int8_kernel=(ture|false)
--valid_targets=(arm|opencl|x86)
```
其中，optimize_out为您希望的优化模型的输出路径。optimize_out_type则可以指定输出模型的序列化方式，其目前支持Protobuf与Naive Buffer两种方式，其中Naive Buffer是一种更轻量级的序列化/反序列化实现。如果你需要使用Lite在mobile端进行预测，那么您需要设置optimize_out_type=naive_buffer。
......@@ -48,8 +47,8 @@ $ ./opt \
## 四. Lite API
为了方便您的使用，我们提供了C++、Java、Python三种API，并且提供了相应的api的完整使用示例:[C++完整示例](cpp_demo)、[Java完整示例](java_demo)、[Python完整示例](../advanced_user_guides/cuda)，您可以参考示例中的说明快速了解C++/Java/Python的API使用方法，并集成到您自己的项目中去。需要说明的是，为了减少第三方库的依赖、提高Lite预测框架的通用性，在移动端使用Lite API您需要准备Naive Buffer存储格式的模型，具体方法可参考第2节`模型优化`。
为了方便您的使用，我们提供了C++、Java、Python三种API，并且提供了相应的api的完整使用示例:[C++完整示例](../demo_guides/cpp_demo)、[Java完整示例](../demo_guides/java_demo)、[Python完整示例](../demo_guides/cuda)，您可以参考示例中的说明快速了解C++/Java/Python的API使用方法，并集成到您自己的项目中去。需要说明的是，为了减少第三方库的依赖、提高Lite预测框架的通用性，在移动端使用Lite API您需要准备Naive Buffer存储格式的模型，具体方法可参考第2节`模型优化`。
## 五. 测试工具
为了使您更好的了解并使用Lite框架，我们向有进一步使用需求的用户开放了 [Lite Model Debug Tool](../advanced_user_guides/debug_tools) 和 [Profile Monitor Tool](../advanced_user_guides/test_tools)。Lite Model Debug Tool可以用来查找Lite框架与PaddlePaddle框架在执行预测时模型中的对应变量值是否有差异，进一步快速定位问题Op，方便复现与排查问题。Profile Monitor Tool可以帮助您了解每个Op的执行时间消耗，其会自动统计Op执行的次数，最长、最短、平均执行时间等等信息，为性能调优做一个基础参考。您可以通过 [相关专题](../advanced_user_guides/debug_tools) 了解更多内容。
为了使您更好的了解并使用Lite框架，我们向有进一步使用需求的用户开放了 [Debug工具](debug#debug) 和 [Profile工具](debug#profiler)。Lite Model Debug Tool可以用来查找Lite框架与PaddlePaddle框架在执行预测时模型中的对应变量值是否有差异，进一步快速定位问题Op，方便复现与排查问题。Profile Monitor Tool可以帮助您了解每个Op的执行时间消耗，其会自动统计Op执行的次数，最长、最短、平均执行时间等等信息，为性能调优做一个基础参考。您可以通过 [相关专题](debug) 了解更多内容。
docs/advanced_user_guides/x2paddle.md → docs/user_guides/x2paddle.md
浏览文件 @ 5dd85c2c
# 通过 X2Paddle 转换模型
# 模型转换工具 X2Paddle
X2Paddle可以将caffe、tensorflow、onnx模型转换成Paddle支持的模型。
[X2Paddle](https://github.com/PaddlePaddle/X2Paddle)支持将Caffe/TensorFlow模型转换为PaddlePaddle模型。目前X2Paddle支持的模型参考[x2paddle_model_zoo](https://github.com/PaddlePaddle/X2Paddle/blob/develop/x2paddle_model_zoo.md)。
## 多框架支持
|模型 | caffe | tensorflow | onnx |
|---|---|---|---|
|mobilenetv1 | Y | Y | |
|mobilenetv2 | Y | Y | Y |
|resnet18 | Y | Y | |
|resnet50 | Y | Y | Y |
|mnasnet | Y | Y | |
|efficientnet | Y | Y | Y |
|squeezenetv1.1 | Y | Y | Y |
|shufflenet | Y | Y | |
|mobilenet_ssd | Y | Y | |
|mobilenet_yolov3 | | Y | |
|inceptionv4 | | | |
|mtcnn | Y | Y | |
|facedetection | Y | | |
|unet | Y | Y | |
|ocr_attention | | | |
|vgg16 | | | |
## 安装
```
......
lite/CMakeLists.txt
浏览文件 @ 5dd85c2c
......@@ -12,6 +12,7 @@ message(STATUS "LITE_WITH_FPGA:\t${LITE_WITH_FPGA}")
message(STATUS "LITE_WITH_BM:\t${LITE_WITH_BM}")
message(STATUS "LITE_WITH_PROFILE:\t${LITE_WITH_PROFILE}")
message(STATUS "LITE_WITH_CV:\t${LITE_WITH_CV}")
message(STATUS "LITE_WITH_ARM_LANG:\t${LITE_WITH_ARM_LANG}")
set(LITE_MODEL_DIR "${THIRD_PARTY_PATH}/install")
set(LITE_ON_MOBILE ${LITE_WITH_LIGHT_WEIGHT_FRAMEWORK})
......@@ -64,6 +65,9 @@ if (LITE_WITH_LIGHT_WEIGHT_FRAMEWORK AND LITE_WITH_ARM)
if (LITE_WITH_NPU)
set(INFER_LITE_PUBLISH_ROOT "${INFER_LITE_PUBLISH_ROOT}.npu")
endif(LITE_WITH_NPU)
if (LITE_WITH_XPU)
set(INFER_LITE_PUBLISH_ROOT "${INFER_LITE_PUBLISH_ROOT}.xpu")
endif(LITE_WITH_XPU)
if (LITE_WITH_FPGA)
set(INFER_LITE_PUBLISH_ROOT "${INFER_LITE_PUBLISH_ROOT}.fpga")
endif(LITE_WITH_FPGA)
......@@ -79,7 +83,16 @@ message(STATUS "publish inference lib to ${INFER_LITE_PUBLISH_ROOT}")
if (LITE_WITH_PYTHON)
add_custom_target(publish_inference_python_lib ${TARGET}
COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/python/lib"
COMMAND cp "${CMAKE_BINARY_DIR}/lite/api/python/pybind/liblite_pybind.so" "${INFER_LITE_PUBLISH_ROOT}/python/lib/lite_core.so")
COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/python/install/libs"
COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/python/install/lite"
COMMAND cp "${CMAKE_BINARY_DIR}/lite/api/python/setup.py" "${INFER_LITE_PUBLISH_ROOT}/python/install/"
COMMAND cp "${CMAKE_SOURCE_DIR}/lite/api/python/__init__.py" "${INFER_LITE_PUBLISH_ROOT}/python/install/lite"
COMMAND cp "${CMAKE_BINARY_DIR}/lite/api/python/pybind/liblite_pybind.so" "${INFER_LITE_PUBLISH_ROOT}/python/install/lite/lite.so"
COMMAND cp "${CMAKE_BINARY_DIR}/lite/api/python/pybind/liblite_pybind.so" "${INFER_LITE_PUBLISH_ROOT}/python/lib/lite.so")
add_custom_target(publish_inference_python_installer ${TARGET}
COMMAND python setup.py bdist_wheel
WORKING_DIRECTORY ${INFER_LITE_PUBLISH_ROOT}/python/install/
DEPENDS publish_inference_python_lib)
add_custom_target(publish_inference_python_light_demo ${TARGET}
COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/demo/python"
COMMAND cp "${CMAKE_SOURCE_DIR}/lite/demo/python/mobilenetv1_light_api.py" "${INFER_LITE_PUBLISH_ROOT}/demo/python/")
......@@ -91,6 +104,7 @@ if (LITE_WITH_PYTHON)
endif()
add_dependencies(publish_inference_python_lib lite_pybind)
add_dependencies(publish_inference publish_inference_python_lib)
add_dependencies(publish_inference publish_inference_python_installer)
add_dependencies(publish_inference publish_inference_python_light_demo)
endif()
......@@ -123,7 +137,29 @@ if (LITE_WITH_X86)
endif()
if(LITE_WITH_CUDA)
add_dependencies(publish_inference paddle_full_api_shared)
add_custom_target(publish_inference_cuda_cxx_lib ${TARGET}
COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/bin"
COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/cxx/include"
COMMAND cp "${CMAKE_SOURCE_DIR}/lite/api/paddle_*.h" "${INFER_LITE_PUBLISH_ROOT}/cxx/include"
COMMAND cp "${CMAKE_BINARY_DIR}/libpaddle_api_full_bundled.a" "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
COMMAND cp "${CMAKE_BINARY_DIR}/libpaddle_api_light_bundled.a" "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
COMMAND cp "${CMAKE_BINARY_DIR}/lite/api/*.so" "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
)
add_dependencies(publish_inference_cuda_cxx_lib bundle_full_api)
add_dependencies(publish_inference_cuda_cxx_lib bundle_light_api)
add_dependencies(publish_inference_cuda_cxx_lib paddle_full_api_shared)
add_dependencies(publish_inference_cuda_cxx_lib paddle_light_api_shared)
add_dependencies(publish_inference publish_inference_cuda_cxx_lib)
add_custom_target(publish_inference_cuda_cxx_demos ${TARGET}
COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/third_party"
COMMAND cp -r "${CMAKE_BINARY_DIR}/third_party/install/*" "${INFER_LITE_PUBLISH_ROOT}/third_party"
COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/demo/cxx"
COMMAND cp -r "${CMAKE_SOURCE_DIR}/lite/demo/cxx/cuda_demo/*" "${INFER_LITE_PUBLISH_ROOT}/demo/cxx"
)
add_dependencies(publish_inference_cuda_cxx_lib publish_inference_cuda_cxx_demos)
add_dependencies(publish_inference_cuda_cxx_demos paddle_full_api_shared)
endif(LITE_WITH_CUDA)
if (LITE_WITH_LIGHT_WEIGHT_FRAMEWORK AND LITE_WITH_ARM)
if (NOT LITE_ON_TINY_PUBLISH)
......@@ -186,6 +222,7 @@ if (LITE_WITH_LIGHT_WEIGHT_FRAMEWORK AND LITE_WITH_ARM)
add_dependencies(publish_inference tiny_publish_cxx_lib)
if(NOT "${CMAKE_BUILD_TYPE}" STREQUAL "Debug")
add_custom_command(TARGET tiny_publish_cxx_lib POST_BUILD
COMMAND ${CMAKE_STRIP} "-s" ${INFER_LITE_PUBLISH_ROOT}/cxx/lib/libpaddle_api_light_bundled.a
COMMAND ${CMAKE_STRIP} "-s" ${INFER_LITE_PUBLISH_ROOT}/cxx/lib/libpaddle_light_api_shared.so)
endif()
endif()
......@@ -282,6 +319,10 @@ if (LITE_WITH_LIGHT_WEIGHT_FRAMEWORK AND LITE_WITH_ARM)
COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/opencl"
COMMAND cp -r "${CMAKE_SOURCE_DIR}/lite/backends/opencl/cl_kernel" "${INFER_LITE_PUBLISH_ROOT}/opencl"
)
if (NOT LITE_ON_TINY_PUBLISH)
add_dependencies(publish_inference_cxx_lib publish_inference_opencl)
else()
add_dependencies(tiny_publish_cxx_lib publish_inference_opencl)
endif()
endif()
endif()
lite/api/CMakeLists.txt
浏览文件 @ 5dd85c2c
......@@ -45,7 +45,11 @@ else()
if ((ARM_TARGET_OS STREQUAL "android") OR (ARM_TARGET_OS STREQUAL "armlinux"))
add_library(paddle_light_api_shared SHARED "")
target_sources(paddle_light_api_shared PUBLIC ${__lite_cc_files} paddle_api.cc light_api.cc light_api_impl.cc)
set_target_properties(paddle_light_api_shared PROPERTIES COMPILE_FLAGS "-flto -fdata-sections")
set(TARGET_COMIPILE_FLAGS "-fdata-sections")
if (NOT (ARM_TARGET_LANG STREQUAL "clang")) #gcc
set(TARGET_COMIPILE_FLAGS "${TARGET_COMIPILE_FLAGS} -flto")
endif()
set_target_properties(paddle_light_api_shared PROPERTIES COMPILE_FLAGS "${TARGET_COMIPILE_FLAGS}")
add_dependencies(paddle_light_api_shared op_list_h kernel_list_h)
if (LITE_WITH_NPU)
# Need to add HIAI runtime libs (libhiai.so) dependency
......@@ -78,6 +82,7 @@ message(STATUS "get X86 kernels ${x86_kernels}")
message(STATUS "get CUDA kernels ${cuda_kernels}")
message(STATUS "get Host kernels ${host_kernels}")
message(STATUS "get ARM kernels ${arm_kernels}")
message(STATUS "get OpenCL kernels ${opencl_kernels}")
message(STATUS "get NPU kernels ${npu_kernels}")
message(STATUS "get XPU kernels ${xpu_kernels}")
message(STATUS "get FPGA kernels ${fpga_kernels}")
......@@ -143,38 +148,40 @@ if(WITH_TESTING)
--optimized_model=${LITE_MODEL_DIR}/lite_naive_model_opt SERIAL)
add_dependencies(test_cxx_api extern_lite_download_lite_naive_model_tar_gz)
if(NOT LITE_WITH_LIGHT_WEIGHT_FRAMEWORK)
lite_cc_test(test_googlenet SRCS test_googlenet_lite.cc
DEPS mir_passes lite_api_test_helper paddle_api_full paddle_api_light gflags utils
${ops} ${host_kernels} ${x86_kernels}
ARGS --model_dir=${LITE_MODEL_DIR}/googlenet)
add_dependencies(test_googlenet extern_lite_download_GoogleNet_inference_tar_gz)
lite_cc_test(test_mobilenetv1_lite_x86 SRCS test_mobilenetv1_lite_x86.cc
DEPS mir_passes lite_api_test_helper paddle_api_full paddle_api_light gflags utils
${ops} ${host_kernels} ${x86_kernels}
ARGS --model_dir=${LITE_MODEL_DIR}/mobilenet_v1)
add_dependencies(test_mobilenetv1_lite_x86 extern_lite_download_mobilenet_v1_tar_gz)
lite_cc_test(test_mobilenetv2_lite_x86 SRCS test_mobilenetv2_lite_x86.cc
DEPS mir_passes lite_api_test_helper paddle_api_full paddle_api_light gflags utils
${ops} ${host_kernels} ${x86_kernels}
ARGS --model_dir=${LITE_MODEL_DIR}/mobilenet_v2_relu)
add_dependencies(test_mobilenetv2_lite_x86 extern_lite_download_mobilenet_v2_relu_tar_gz)
lite_cc_test(test_inceptionv4_lite_x86 SRCS test_inceptionv4_lite_x86.cc
DEPS mir_passes lite_api_test_helper paddle_api_full paddle_api_light gflags utils
${ops} ${host_kernels} ${x86_kernels}
ARGS --model_dir=${LITE_MODEL_DIR}/inception_v4_simple)
add_dependencies(test_inceptionv4_lite_x86 extern_lite_download_inception_v4_simple_tar_gz)
lite_cc_test(test_resnet50_lite_x86 SRCS test_resnet50_lite_x86.cc
DEPS mir_passes lite_api_test_helper paddle_api_full paddle_api_light gflags utils
${ops} ${host_kernels} ${x86_kernels}
ARGS --model_dir=${LITE_MODEL_DIR}/resnet50)
add_dependencies(test_resnet50_lite_x86 extern_lite_download_resnet50_tar_gz)
lite_cc_test(test_step_rnn_lite_x86 SRCS test_step_rnn_lite_x86.cc
DEPS mir_passes lite_api_test_helper paddle_api_full paddle_api_light gflags utils
${ops} ${host_kernels} ${x86_kernels}
ARGS --model_dir=${LITE_MODEL_DIR}/step_rnn)
add_dependencies(test_step_rnn_lite_x86 extern_lite_download_step_rnn_tar_gz)
if(LITE_WITH_X86)
lite_cc_test(test_googlenet SRCS test_googlenet_lite.cc
DEPS mir_passes lite_api_test_helper paddle_api_full paddle_api_light gflags utils
${ops} ${host_kernels} ${x86_kernels}
ARGS --model_dir=${LITE_MODEL_DIR}/googlenet)
add_dependencies(test_googlenet extern_lite_download_GoogleNet_inference_tar_gz)
lite_cc_test(test_mobilenetv1_lite_x86 SRCS test_mobilenetv1_lite_x86.cc
DEPS mir_passes lite_api_test_helper paddle_api_full paddle_api_light gflags utils
${ops} ${host_kernels} ${x86_kernels}
ARGS --model_dir=${LITE_MODEL_DIR}/mobilenet_v1)
add_dependencies(test_mobilenetv1_lite_x86 extern_lite_download_mobilenet_v1_tar_gz)
lite_cc_test(test_mobilenetv2_lite_x86 SRCS test_mobilenetv2_lite_x86.cc
DEPS mir_passes lite_api_test_helper paddle_api_full paddle_api_light gflags utils
${ops} ${host_kernels} ${x86_kernels}
ARGS --model_dir=${LITE_MODEL_DIR}/mobilenet_v2_relu)
add_dependencies(test_mobilenetv2_lite_x86 extern_lite_download_mobilenet_v2_relu_tar_gz)
lite_cc_test(test_inceptionv4_lite_x86 SRCS test_inceptionv4_lite_x86.cc
DEPS mir_passes lite_api_test_helper paddle_api_full paddle_api_light gflags utils
${ops} ${host_kernels} ${x86_kernels}
ARGS --model_dir=${LITE_MODEL_DIR}/inception_v4_simple)
add_dependencies(test_inceptionv4_lite_x86 extern_lite_download_inception_v4_simple_tar_gz)
lite_cc_test(test_resnet50_lite_x86 SRCS test_resnet50_lite_x86.cc
DEPS mir_passes lite_api_test_helper paddle_api_full paddle_api_light gflags utils
${ops} ${host_kernels} ${x86_kernels}
ARGS --model_dir=${LITE_MODEL_DIR}/resnet50)
add_dependencies(test_resnet50_lite_x86 extern_lite_download_resnet50_tar_gz)
lite_cc_test(test_step_rnn_lite_x86 SRCS test_step_rnn_lite_x86.cc
DEPS mir_passes lite_api_test_helper paddle_api_full paddle_api_light gflags utils
${ops} ${host_kernels} ${x86_kernels}
ARGS --model_dir=${LITE_MODEL_DIR}/step_rnn)
add_dependencies(test_step_rnn_lite_x86 extern_lite_download_step_rnn_tar_gz)
endif()
if(LITE_WITH_BM)
lite_cc_test(test_resnet50_lite_bm SRCS test_resnet50_lite_bm.cc
lite_cc_test(test_classify_lite_bm SRCS test_classify_lite_bm.cc
DEPS mir_passes lite_api_test_helper paddle_api_full paddle_api_light gflags utils
${ops} ${host_kernels} ${bm_kernels} ${bm_bridges}
ARGS --model_dir=${LITE_MODEL_DIR}/resnet50)
......@@ -229,6 +236,7 @@ if(LITE_WITH_LIGHT_WEIGHT_FRAMEWORK AND WITH_TESTING)
ARGS --cl_path=${CMAKE_SOURCE_DIR}/lite/backends/opencl
--model_dir=${LITE_MODEL_DIR}/inception_v4 SERIAL)
add_dependencies(test_inceptionv4 extern_lite_download_inception_v4_simple_tar_gz)
# brief: we comment ocr_test_ut because we do not supply ocr model to test, it is the reference to infer nlp model
# lite_cc_test(test_ocr_attention SRCS ocr_attention_test.cc
# DEPS ${lite_model_test_DEPS})
......@@ -295,6 +303,11 @@ if (LITE_ON_TINY_PUBLISH)
return()
endif()
# add library for opt_base
lite_cc_library(opt_base SRCS opt_base.cc cxx_api_impl.cc paddle_api.cc cxx_api.cc DEPS kernel op optimizer mir_passes utils)
add_dependencies(opt_base supported_kernel_op_info_h framework_proto all_kernel_faked_cc kernel_list_h)
if (LITE_ON_MODEL_OPTIMIZE_TOOL)
message(STATUS "Compiling opt")
lite_cc_binary(opt SRCS opt.cc cxx_api_impl.cc paddle_api.cc cxx_api.cc
......@@ -377,6 +390,16 @@ if(NOT IOS)
FPGA_DEPS ${fpga_kernels}
X86_DEPS ${x86_kernels}
CUDA_DEPS ${cuda_kernels})
lite_cc_binary(test_transformer SRCS transform_test.cc DEPS paddle_api_full paddle_api_light gflags utils
${ops} ${host_kernels}
ARM_DEPS ${arm_kernels}
CV_DEPS paddle_cv_arm
NPU_DEPS ${npu_kernels}
XPU_DEPS ${xpu_kernels}
CL_DEPS ${opencl_kernels}
FPGA_DEPS ${fpga_kernels}
X86_DEPS ${x86_kernels}
CUDA_DEPS ${cuda_kernels})
endif()
#lite_cc_binary(cxx_api_bin SRCS cxx_api_bin.cc
......
lite/api/android/jni/native/CMakeLists.txt
浏览文件 @ 5dd85c2c
......@@ -25,7 +25,11 @@ if (NOT LITE_ON_TINY_PUBLISH)
endif()
else()
add_library(paddle_lite_jni SHARED "")
set_target_properties(paddle_lite_jni PROPERTIES COMPILE_FLAGS "-flto -fdata-sections")
set(TARGET_COMIPILE_FLAGS "-fdata-sections")
if (NOT (ARM_TARGET_LANG STREQUAL "clang")) #gcc
set(TARGET_COMIPILE_FLAGS "${TARGET_COMIPILE_FLAGS} -flto")
endif()
set_target_properties(paddle_lite_jni PROPERTIES COMPILE_FLAGS ${TARGET_COMIPILE_FLAGS})
target_sources(paddle_lite_jni PUBLIC ${__lite_cc_files} paddle_lite_jni.cc tensor_jni.cc)
add_dependencies(paddle_lite_jni op_list_h kernel_list_h)
if (LITE_WITH_NPU)
......
lite/api/android/jni/native/paddle_lite_jni.h
浏览文件 @ 5dd85c2c
......@@ -17,11 +17,6 @@
#include <jni.h>
/* Header for class com_baidu_paddle_lite_PaddlePredictor */
#include "lite/api/paddle_lite_factory_helper.h"
#include "lite/api/paddle_use_kernels.h"
#include "lite/api/paddle_use_ops.h"
#ifndef LITE_ON_TINY_PUBLISH
#include "lite/api/paddle_use_passes.h"
#endif
#ifdef __cplusplus
extern "C" {
#endif
......
lite/api/apis_test.cc
浏览文件 @ 5dd85c2c
......@@ -21,9 +21,6 @@
#include <vector>
#include "lite/api/cxx_api.h"
#include "lite/api/light_api.h"
#include "lite/api/paddle_use_kernels.h"
#include "lite/api/paddle_use_ops.h"
#include "lite/api/paddle_use_passes.h"
#include "lite/core/mir/pass_registry.h"
DEFINE_string(model_dir, "", "");
......
lite/api/benchmark.cc
浏览文件 @ 5dd85c2c
......@@ -23,31 +23,28 @@
#include <string>
#include <vector>
#include "lite/api/paddle_api.h"
#include "lite/api/paddle_use_kernels.h"
#include "lite/api/paddle_use_ops.h"
#include "lite/api/paddle_use_passes.h"
#include "lite/core/device_info.h"
#include "lite/utils/cp_logging.h"
#include "lite/utils/string.h"
DEFINE_string(model_dir,
"",
"the path of the model, set model_dir when the model is no "
"combined formate. This option will be ignored if model_file "
"and param_file are exist.");
DEFINE_string(model_file,
"the path of the model, the model and param files is under "
"model_dir.");
DEFINE_string(model_filename,
"",
"the path of model file, set model_file when the model is "
"combined formate.");
DEFINE_string(param_file,
"the filename of model file. When the model is combined formate, "
"please set model_file.");
DEFINE_string(param_filename,
"",
"the path of param file, set param_file when the model is "
"the filename of param file, set param_file when the model is "
"combined formate.");
DEFINE_string(input_shape,
"1,3,224,224",
"set input shapes according to the model, "
"separated by colon and comma, "
"such as 1,3,244,244:1,3,300,300.");
"such as 1,3,244,244");
DEFINE_string(input_img_path, "", "the path of input image");
DEFINE_int32(warmup, 0, "warmup times");
DEFINE_int32(repeats, 1, "repeats times");
DEFINE_int32(power_mode,
......@@ -80,12 +77,13 @@ inline double GetCurrentUS() {
return 1e+6 * time.tv_sec + time.tv_usec;
}
void OutputOptModel(const std::string& save_optimized_model_dir,
const std::vector<std::vector<int64_t>>& input_shapes) {
void OutputOptModel(const std::string& save_optimized_model_dir) {
lite_api::CxxConfig config;
config.set_model_dir(FLAGS_model_dir);
config.set_model_file(FLAGS_model_file);
config.set_param_file(FLAGS_param_file);
if (!FLAGS_model_filename.empty() && !FLAGS_param_filename.empty()) {
config.set_model_file(FLAGS_model_dir + "/" + FLAGS_model_filename);
config.set_param_file(FLAGS_model_dir + "/" + FLAGS_param_filename);
}
std::vector<Place> vaild_places = {
Place{TARGET(kARM), PRECISION(kFloat)},
};
......@@ -109,7 +107,7 @@ void OutputOptModel(const std::string& save_optimized_model_dir,
}
#ifdef LITE_WITH_LIGHT_WEIGHT_FRAMEWORK
void Run(const std::vector<std::vector<int64_t>>& input_shapes,
void Run(const std::vector<int64_t>& input_shape,
const std::string& model_dir,
const std::string model_name) {
// set config and create predictor
......@@ -121,17 +119,27 @@ void Run(const std::vector<std::vector<int64_t>>& input_shapes,
auto predictor = lite_api::CreatePaddlePredictor(config);
// set input
for (int j = 0; j < input_shapes.size(); ++j) {
auto input_tensor = predictor->GetInput(j);
input_tensor->Resize(input_shapes[j]);
auto input_data = input_tensor->mutable_data<float>();
int input_num = 1;
for (size_t i = 0; i < input_shapes[j].size(); ++i) {
input_num *= input_shapes[j][i];
}
auto input_tensor = predictor->GetInput(0);
input_tensor->Resize(input_shape);
auto input_data = input_tensor->mutable_data<float>();
int input_num = 1;
for (size_t i = 0; i < input_shape.size(); ++i) {
input_num *= input_shape[i];
}
if (FLAGS_input_img_path.empty()) {
for (int i = 0; i < input_num; ++i) {
input_data[i] = 1.f;
}
} else {
std::fstream fs(FLAGS_input_img_path);
if (!fs.is_open()) {
LOG(FATAL) << "open input image " << FLAGS_input_img_path << " error.";
}
for (int i = 0; i < input_num; i++) {
fs >> input_data[i];
}
// LOG(INFO) << "input data:" << input_data[0] << " " <<
// input_data[input_num-1];
}
// warmup
......@@ -178,25 +186,12 @@ int main(int argc, char** argv) {
exit(0);
}
if (FLAGS_model_dir.back() == '/') {
FLAGS_model_dir.pop_back();
}
std::size_t found = FLAGS_model_dir.find_last_of("/");
std::string model_name = FLAGS_model_dir.substr(found + 1);
std::string save_optimized_model_dir = FLAGS_model_dir + "opt2";
auto split_string =
[](const std::string& str_in) -> std::vector<std::string> {
std::vector<std::string> str_out;
std::string tmp_str = str_in;
while (!tmp_str.empty()) {
size_t next_offset = tmp_str.find(":");
str_out.push_back(tmp_str.substr(0, next_offset));
if (next_offset == std::string::npos) {
break;
} else {
tmp_str = tmp_str.substr(next_offset + 1);
}
}
return str_out;
};
std::string save_optimized_model_dir = FLAGS_model_dir + "_opt2";
auto get_shape = [](const std::string& str_shape) -> std::vector<int64_t> {
std::vector<int64_t> shape;
......@@ -214,22 +209,18 @@ int main(int argc, char** argv) {
return shape;
};
std::vector<std::string> str_input_shapes = split_string(FLAGS_input_shape);
std::vector<std::vector<int64_t>> input_shapes;
for (size_t i = 0; i < str_input_shapes.size(); ++i) {
input_shapes.push_back(get_shape(str_input_shapes[i]));
}
std::vector<int64_t> input_shape = get_shape(FLAGS_input_shape);
// Output optimized model if needed
if (FLAGS_run_model_optimize) {
paddle::lite_api::OutputOptModel(save_optimized_model_dir, input_shapes);
paddle::lite_api::OutputOptModel(save_optimized_model_dir);
}
#ifdef LITE_WITH_LIGHT_WEIGHT_FRAMEWORK
// Run inference using optimized model
std::string run_model_dir =
FLAGS_run_model_optimize ? save_optimized_model_dir : FLAGS_model_dir;
paddle::lite_api::Run(input_shapes, run_model_dir, model_name);
paddle::lite_api::Run(input_shape, run_model_dir, model_name);
#endif
return 0;
}
lite/api/cxx_api.cc
浏览文件 @ 5dd85c2c
......@@ -294,6 +294,32 @@ void Predictor::Build(const cpp::ProgramDesc &desc,
inner_places.emplace_back(TARGET(kHost), PRECISION(kAny), DATALAYOUT(kAny));
inner_places.emplace_back(
TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW));
const std::vector<std::string> quant_dequant_op = {
"fake_quantize_abs_max",
"fake_quantize_range_abs_max",
"fake_quantize_moving_average_abs_max",
"fake_quantize_dequantize_moving_average_abs_max",
"fake_dequantize_max_abs",
"fake_channel_wise_dequantize_max_abs"};
bool is_quantized_model = false;
for (size_t i = 0; i < program_desc_.BlocksSize() && !is_quantized_model;
++i) {
auto *block_desc = program_desc_.GetBlock<cpp::BlockDesc>(i);
for (size_t j = 0; j < block_desc->OpsSize() && !is_quantized_model; ++j) {
auto *op_desc = block_desc->GetOp<cpp::OpDesc>(j);
std::string op_type = op_desc->Type();
if (std::find(quant_dequant_op.begin(),
quant_dequant_op.end(),
op_type) != quant_dequant_op.end()) {
is_quantized_model = true;
}
}
}
if (is_quantized_model) {
inner_places.emplace_back(Place{TARGET(kARM), PRECISION(kInt8)});
}
Program program(desc, scope_, inner_places);
core::KernelPickFactor factor;
......@@ -333,16 +359,16 @@ lite::Tensor *Predictor::GetInputByName(const std::string &name) {
}
}
#ifdef LITE_WITH_TRAIN
void Predictor::FeedVars(const std::vector<framework::Tensor> &tensors) {
auto var = scope_->FindVar("feed");
auto &feed_list = *(var->GetMutable<std::vector<lite::Tensor>>());
feed_list.resize(tensors.size());
// #ifdef LITE_WITH_TRAIN
// void Predictor::FeedVars(const std::vector<framework::Tensor> &tensors) {
// auto var = scope_->FindVar("feed");
// auto &feed_list = *(var->GetMutable<std::vector<lite::Tensor>>());
// feed_list.resize(tensors.size());
for (size_t i = 0; i < tensors.size(); ++i)
feed_list[i].ShareDataWith(tensors[i]);
}
#endif
// for (size_t i = 0; i < tensors.size(); ++i)
// feed_list[i].ShareDataWith(tensors[i]);
// }
// #endif
} // namespace lite
} // namespace paddle
lite/api/cxx_api.h
浏览文件 @ 5dd85c2c
......@@ -101,14 +101,14 @@ class LITE_API Predictor {
bool record_info = false);
void SaveOpKernelInfo(const std::string& model_dir);
#ifdef LITE_WITH_TRAIN
void Run(const std::vector<framework::Tensor>& tensors) {
FeedVars(tensors);
program_->Run();
}
void FeedVars(const std::vector<framework::Tensor>& tensors);
#endif
// #ifdef LITE_WITH_TRAIN
// void Run(const std::vector<framework::Tensor>& tensors) {
// FeedVars(tensors);
// program_->Run();
// }
// void FeedVars(const std::vector<framework::Tensor>& tensors);
// #endif
private:
Optimizer optimizer_;
......
lite/api/cxx_api_bin.cc
浏览文件 @ 5dd85c2c
......@@ -67,7 +67,7 @@ void Run(const char* model_dir, int repeat) {
int main(int argc, char** argv) {
CHECK_EQ(argc, 3) << "usage: ./cmd <model_dir> <repeat>";
paddle::lite::Run(argv[1], std::stoi(argv[2]));
paddle::lite::Run(argv[1], atoi(argv[2]));
return 0;
}
......
lite/api/cxx_api_impl.cc
浏览文件 @ 5dd85c2c
......@@ -38,11 +38,13 @@ void CxxPaddleApiImpl::Init(const lite_api::CxxConfig &config) {
std::vector<std::string> passes{};
auto use_layout_preprocess_pass =
config.model_dir().find("OPENCL_PRE_PRECESS");
if (use_layout_preprocess_pass != std::string::npos) {
VLOG(1) << "use_layout_preprocess_pass:" << use_layout_preprocess_pass;
if (places[0].target == TARGET(kOpenCL) &&
use_layout_preprocess_pass != std::string::npos) {
passes = {"type_layout_cast_preprocess_pass"};
VLOG(1) << "add pass:" << passes[0];
}
raw_predictor_.Build(config, places, passes);
mode_ = config.power_mode();
threads_ = config.threads();
......
lite/api/light_api.cc
浏览文件 @ 5dd85c2c
......@@ -13,6 +13,12 @@
// limitations under the License.
#include "lite/api/light_api.h"
#include "paddle_use_kernels.h" // NOLINT
#include "paddle_use_ops.h" // NOLINT
#ifndef LITE_ON_TINY_PUBLISH
#include "lite/api/paddle_use_passes.h"
#endif
#include <algorithm>
namespace paddle {
......
lite/api/light_api_impl.cc
浏览文件 @ 5dd85c2c
......@@ -58,6 +58,7 @@ void LightPredictorImpl::Run() {
std::shared_ptr<lite_api::PaddlePredictor> LightPredictorImpl::Clone() {
LOG(FATAL) << "The Clone API is not supported in LigthPredictor";
return nullptr;
}
std::string LightPredictorImpl::GetVersion() const { return lite::version(); }
......
lite/api/light_api_shared.cc
浏览文件 @ 5dd85c2c
......@@ -12,11 +12,6 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "lite/api/paddle_api.h"
#include "lite/api/paddle_use_kernels.h"
#include "lite/api/paddle_use_ops.h"
#ifndef LITE_ON_TINY_PUBLISH
#include "lite/api/paddle_use_passes.h"
#endif
namespace paddle {
namespace lite_api {
......
lite/api/light_api_test.cc
浏览文件 @ 5dd85c2c
......@@ -15,9 +15,6 @@
#include "lite/api/light_api.h"
#include <gflags/gflags.h>
#include <gtest/gtest.h>
#include "lite/api/paddle_use_kernels.h"
#include "lite/api/paddle_use_ops.h"
#include "lite/api/paddle_use_passes.h"
DEFINE_string(optimized_model, "", "");
......
lite/api/lite_multithread_test.cc
浏览文件 @ 5dd85c2c
......@@ -16,9 +16,6 @@
#include <string>
#include <vector>
#include "lite/api/paddle_api.h"
#include "lite/api/paddle_use_kernels.h"
#include "lite/api/paddle_use_ops.h"
#include "lite/api/paddle_use_passes.h"
#include "lite/api/test_helper.h"
#include "lite/core/device_info.h"
#include "lite/core/profile/timer.h"
......
lite/api/mobilenetv1_test.cc
浏览文件 @ 5dd85c2c
......@@ -53,9 +53,13 @@ void TestModel(const std::vector<Place>& valid_places,
predictor.Run();
}
auto start = GetCurrentUS();
double sum_duration = 0.0; // millisecond;
for (int i = 0; i < FLAGS_repeats; ++i) {
auto start = GetCurrentUS();
predictor.Run();
auto duration = (GetCurrentUS() - start) / 1000.0;
sum_duration += duration;
VLOG(1) << "run_idx:" << i << " " << duration << " ms";
}
if (save_model) {
......@@ -68,8 +72,7 @@ void TestModel(const std::vector<Place>& valid_places,
LOG(INFO) << "================== Speed Report ===================";
LOG(INFO) << "Model: " << model_dir << ", threads num " << FLAGS_threads
<< ", warmup: " << FLAGS_warmup << ", repeats: " << FLAGS_repeats
<< ", spend " << (GetCurrentUS() - start) / FLAGS_repeats / 1000.0
<< " ms in average.";
<< ", spend " << sum_duration / FLAGS_repeats << " ms in average.";
std::vector<std::vector<float>> ref;
ref.emplace_back(std::vector<float>(
......@@ -92,7 +95,7 @@ void TestModel(const std::vector<Place>& valid_places,
if (first_target == TARGET(kOpenCL) || first_target == TARGET(kNPU)) {
ASSERT_EQ(out->dims().production(), 1000);
double eps = 0.1;
double eps = first_target == TARGET(kOpenCL) ? 0.12 : 0.1;
for (int i = 0; i < ref.size(); ++i) {
for (int j = 0; j < ref[i].size(); ++j) {
auto result = pdata[j * step + (out->dims()[1] * i)];
......@@ -115,24 +118,22 @@ void TestModel(const std::vector<Place>& valid_places,
}
// Get detailed result
auto* pred = &predictor;
size_t output_tensor_num = pred->GetOutputNames().size();
VLOG(1) << "output tesnor num:" << output_tensor_num;
size_t output_tensor_num = predictor.GetOutputNames().size();
VLOG(1) << "output tensor num:" << output_tensor_num;
for (size_t tidx = 0; tidx < output_tensor_num; ++tidx) {
std::unique_ptr<const Tensor> output_tensor(
std::move(pred->GetOutput(tidx)));
auto* output_tensor = predictor.GetOutput(tidx);
VLOG(1) << "============= output tensor " << tidx << " =============\n";
auto out_dims = output_tensor->dims();
VLOG(1) << "out_dims:" << out_dims;
float sum = 0.f;
for (int i = 0; i < out_dims.production(); ++i) {
sum += output_tensor->data<float>()[i];
}
VLOG(1) << "out_dims.production():" << out_dims.production();
VLOG(1) << "output tensor sum value:" << sum;
VLOG(1) << "output tensor mean value:" << sum / out_dims.production();
auto out_data = output_tensor->data<float>();
auto out_mean = compute_mean<float>(out_data, out_dims.production());
auto out_std_dev = compute_standard_deviation<float>(
out_data, out_dims.production(), true, out_mean);
VLOG(1) << "output tensor dims:" << out_dims;
VLOG(1) << "output tensor elements num:" << out_dims.production();
VLOG(1) << "output tensor standard deviation:" << out_std_dev;
VLOG(1) << "output tensor mean value:" << out_mean;
// print result
for (int i = 0; i < out_dims.production(); ++i) {
......
lite/api/mobilenetv2_test.cc
浏览文件 @ 5dd85c2c
......@@ -54,9 +54,13 @@ void TestModel(const std::vector<Place>& valid_places,
predictor.Run();
}
auto start = GetCurrentUS();
double sum_duration = 0.0; // millisecond;
for (int i = 0; i < FLAGS_repeats; ++i) {
auto start = GetCurrentUS();
predictor.Run();
auto duration = (GetCurrentUS() - start) / 1000.0;
sum_duration += duration;
VLOG(1) << "run_idx:" << i << " " << duration << " ms";
}
if (save_model) {
......@@ -69,8 +73,7 @@ void TestModel(const std::vector<Place>& valid_places,
LOG(INFO) << "================== Speed Report ===================";
LOG(INFO) << "Model: " << model_dir << ", threads num " << FLAGS_threads
<< ", warmup: " << FLAGS_warmup << ", repeats: " << FLAGS_repeats
<< ", spend " << (GetCurrentUS() - start) / FLAGS_repeats / 1000.0
<< " ms in average.";
<< ", spend " << sum_duration / FLAGS_repeats << " ms in average.";
std::vector<std::vector<float>> ref;
// i = 1
......@@ -94,7 +97,7 @@ void TestModel(const std::vector<Place>& valid_places,
if (first_target == TARGET(kOpenCL) || first_target == TARGET(kNPU)) {
ASSERT_EQ(out->dims().production(), 1000);
double eps = 0.1;
double eps = first_target == TARGET(kOpenCL) ? 0.15 : 0.1;
for (int i = 0; i < ref.size(); ++i) {
for (int j = 0; j < ref[i].size(); ++j) {
auto result = pdata[j * step + (out->dims()[1] * i)];
......@@ -117,24 +120,22 @@ void TestModel(const std::vector<Place>& valid_places,
}
// Get detailed result
auto* pred = &predictor;
size_t output_tensor_num = pred->GetOutputNames().size();
VLOG(1) << "output tesnor num:" << output_tensor_num;
size_t output_tensor_num = predictor.GetOutputNames().size();
VLOG(1) << "output tensor num:" << output_tensor_num;
for (size_t tidx = 0; tidx < output_tensor_num; ++tidx) {
std::unique_ptr<const Tensor> output_tensor(
std::move(pred->GetOutput(tidx)));
auto* output_tensor = predictor.GetOutput(tidx);
VLOG(1) << "============= output tensor " << tidx << " =============\n";
auto out_dims = output_tensor->dims();
VLOG(1) << "out_dims:" << out_dims;
float sum = 0.f;
for (int i = 0; i < out_dims.production(); ++i) {
sum += output_tensor->data<float>()[i];
}
VLOG(1) << "out_dims.production():" << out_dims.production();
VLOG(1) << "output tensor sum value:" << sum;
VLOG(1) << "output tensor mean value:" << sum / out_dims.production();
auto out_data = output_tensor->data<float>();
auto out_mean = compute_mean<float>(out_data, out_dims.production());
auto out_std_dev = compute_standard_deviation<float>(
out_data, out_dims.production(), true, out_mean);
VLOG(1) << "output tensor dims:" << out_dims;
VLOG(1) << "output tensor elements num:" << out_dims.production();
VLOG(1) << "output tensor standard deviation:" << out_std_dev;
VLOG(1) << "output tensor mean value:" << out_mean;
// print result
for (int i = 0; i < out_dims.production(); ++i) {
......
lite/api/model_test.cc
浏览文件 @ 5dd85c2c
......@@ -17,9 +17,6 @@
#include <string>
#include <vector>
#include "lite/api/paddle_api.h"
#include "lite/api/paddle_use_kernels.h"
#include "lite/api/paddle_use_ops.h"
#include "lite/api/paddle_use_passes.h"
#include "lite/api/test_helper.h"
#include "lite/core/device_info.h"
#include "lite/core/profile/timer.h"
......@@ -141,7 +138,7 @@ void Run(const std::vector<std::vector<int64_t>>& input_shapes,
std::ofstream out(FLAGS_arg_name + ".txt");
for (size_t i = 0; i < arg_num; ++i) {
sum += arg_tensor->data<float>()[i];
out << std::to_string(arg_tensor->data<float>()[i]) << "\n";
out << paddle::lite::to_string(arg_tensor->data<float>()[i]) << "\n";
}
LOG(INFO) << FLAGS_arg_name << " shape is " << os.str()
<< ", mean value is " << sum * 1. / arg_num;
......
lite/api/model_test_classify.cc
浏览文件 @ 5dd85c2c
......@@ -17,9 +17,6 @@
#include <string>
#include <vector>
#include "lite/api/paddle_api.h"
#include "lite/api/paddle_use_kernels.h"
#include "lite/api/paddle_use_ops.h"
#include "lite/api/paddle_use_passes.h"
#include "lite/api/test_helper.h"
#include "lite/core/device_info.h"
#include "lite/core/profile/timer.h"
......@@ -253,7 +250,7 @@ void Run(const std::vector<std::vector<int64_t>>& input_shapes,
std::ofstream out(FLAGS_arg_name + ".txt");
for (size_t i = 0; i < arg_num; ++i) {
sum += arg_tensor->data<float>()[i];
out << std::to_string(arg_tensor->data<float>()[i]) << "\n";
out << paddle::lite::to_string(arg_tensor->data<float>()[i]) << "\n";
}
LOG(INFO) << FLAGS_arg_name << " shape is " << os.str()
<< ", mean value is " << sum * 1. / arg_num;
......
lite/api/model_test_detection.cc
浏览文件 @ 5dd85c2c
......@@ -17,9 +17,6 @@
#include <string>
#include <vector>
#include "lite/api/paddle_api.h"
#include "lite/api/paddle_use_kernels.h"
#include "lite/api/paddle_use_ops.h"
#include "lite/api/paddle_use_passes.h"
#include "lite/api/test_helper.h"
#include "lite/core/device_info.h"
#include "lite/core/profile/timer.h"
......@@ -267,7 +264,7 @@ void Run(const std::vector<std::vector<int64_t>>& input_shapes,
std::ofstream out(FLAGS_arg_name + ".txt");
for (size_t i = 0; i < arg_num; ++i) {
sum += arg_tensor->data<float>()[i];
out << std::to_string(arg_tensor->data<float>()[i]) << "\n";
out << paddle::lite::to_string(arg_tensor->data<float>()[i]) << "\n";
}
LOG(INFO) << FLAGS_arg_name << " shape is " << os.str()
<< ", mean value is " << sum * 1. / arg_num;
......
lite/api/ocr_attention_test.cc
浏览文件 @ 5dd85c2c
......@@ -32,18 +32,10 @@ void TestModel(const std::vector<Place>& valid_places, bool use_npu = false) {
predictor.Build(FLAGS_model_dir, "", "", valid_places);
auto* input_tensor = predictor.GetInput(0);
input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 1, 48, 512})));
auto* data = input_tensor->mutable_data<float>();
auto item_size = input_tensor->dims().production();
for (int i = 0; i < item_size; i++) {
data[i] = 1;
}
auto* init_scores = predictor.GetInput(2);
init_scores->Resize(DDim(std::vector<DDim::value_type>({1, 1})));
auto* data_scores = init_scores->mutable_data<float>();
auto scores_size = input_tensor->dims().production();
auto scores_size = init_scores->dims().production();
for (int i = 0; i < scores_size; i++) {
data_scores[i] = 0;
}
......@@ -53,7 +45,7 @@ void TestModel(const std::vector<Place>& valid_places, bool use_npu = false) {
auto* init_ids = predictor.GetInput(1);
init_ids->Resize(DDim(std::vector<DDim::value_type>({1, 1})));
auto* data_ids = init_ids->mutable_data<float>();
auto* data_ids = init_ids->mutable_data<int64_t>();
auto ids_size = init_ids->dims().production();
for (int i = 0; i < ids_size; i++) {
data_ids[i] = 0;
......@@ -62,6 +54,13 @@ void TestModel(const std::vector<Place>& valid_places, bool use_npu = false) {
std::vector<std::vector<uint64_t>> lod_i{{0, 1}, {0, 1}};
*lod_ids = lod_i;
auto* input_tensor = predictor.GetInput(0);
input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 1, 48, 512})));
auto* data = input_tensor->mutable_data<float>();
auto item_size = input_tensor->dims().production();
for (int i = 0; i < item_size; i++) {
data[i] = 1;
}
for (int i = 0; i < FLAGS_warmup; ++i) {
predictor.Run();
}
......@@ -102,6 +101,7 @@ void TestModel(const std::vector<Place>& valid_places, bool use_npu = false) {
TEST(OcrAttention, test_arm) {
std::vector<Place> valid_places({
Place{TARGET(kARM), PRECISION(kInt64)},
Place{TARGET(kARM), PRECISION(kFloat)},
});
......
lite/api/opt.cc
浏览文件 @ 5dd85c2c
......@@ -67,7 +67,6 @@ DEFINE_string(valid_targets,
"arm",
"The targets this model optimized for, should be one of (arm, "
"opencl, x86), splitted by space");
DEFINE_bool(prefer_int8_kernel, false, "Prefer to run model with int8 kernels");
DEFINE_bool(print_supported_ops,
false,
"Print supported operators on the inputed target");
......@@ -88,7 +87,10 @@ std::vector<Place> ParserValidPlaces() {
auto target_reprs = lite::Split(FLAGS_valid_targets, ",");
for (auto& target_repr : target_reprs) {
if (target_repr == "arm") {
valid_places.emplace_back(TARGET(kARM));
valid_places.emplace_back(
Place{TARGET(kARM), PRECISION(kFloat), DATALAYOUT(kNCHW)});
valid_places.emplace_back(
Place{TARGET(kARM), PRECISION(kInt32), DATALAYOUT(kNCHW)});
} else if (target_repr == "opencl") {
valid_places.emplace_back(
Place{TARGET(kOpenCL), PRECISION(kFP16), DATALAYOUT(kImageDefault)});
......@@ -118,11 +120,6 @@ std::vector<Place> ParserValidPlaces() {
<< "At least one target should be set, should set the "
"command argument 'valid_targets'";
if (FLAGS_prefer_int8_kernel) {
LOG(WARNING) << "Int8 mode is only support by ARM target";
valid_places.insert(valid_places.begin(),
Place{TARGET(kARM), PRECISION(kInt8)});
}
return valid_places;
}
......@@ -252,7 +249,6 @@ void PrintHelpInfo() {
" `--optimize_out_type=(protobuf|naive_buffer)`\n"
" `--optimize_out=<output_optimize_model_dir>`\n"
" `--valid_targets=(arm|opencl|x86|npu|xpu)`\n"
" `--prefer_int8_kernel=(true|false)`\n"
" `--record_tailoring_info=(true|false)`\n"
" Arguments of model checking and ops information:\n"
" `--print_all_ops=true` Display all the valid operators of "
......
lite/api/opt_base.cc 0 → 100644
浏览文件 @ 5dd85c2c
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "lite/api/opt_base.h"
#include "all_kernel_faked.cc" // NOLINT
namespace paddle {
namespace lite_api {
void OptBase::SetModelDir(const std::string& model_path) {
opt_config_.set_model_dir(model_path);
}
void OptBase::SetModelFile(const std::string& model_path) {
opt_config_.set_model_file(model_path);
}
void OptBase::SetParamFile(const std::string& param_path) {
opt_config_.set_param_file(param_path);
}
void OptBase::SetModelType(std::string optimize_out_type) {
if (optimize_out_type == "protobuf") {
model_type_ = LiteModelType::kProtobuf;
} else if (optimize_out_type == "naive_buffer") {
model_type_ = LiteModelType::kNaiveBuffer;
} else {
LOG(FATAL) << "Unsupported Model type :" << optimize_out_type;
}
}
void OptBase::SetValidPlaces(const std::string& valid_places) {
valid_places_.clear();
auto target_reprs = lite::Split(valid_places, ",");
for (auto& target_repr : target_reprs) {
if (target_repr == "arm") {
valid_places_.emplace_back(TARGET(kARM));
} else if (target_repr == "opencl") {
valid_places_.emplace_back(
Place{TARGET(kOpenCL), PRECISION(kFP16), DATALAYOUT(kImageDefault)});
valid_places_.emplace_back(
Place{TARGET(kOpenCL), PRECISION(kFloat), DATALAYOUT(kNCHW)});
valid_places_.emplace_back(
Place{TARGET(kOpenCL), PRECISION(kAny), DATALAYOUT(kImageDefault)});
valid_places_.emplace_back(
Place{TARGET(kOpenCL), PRECISION(kAny), DATALAYOUT(kNCHW)});
valid_places_.emplace_back(
TARGET(kARM)); // enable kARM CPU kernel when no opencl kernel
} else if (target_repr == "x86") {
valid_places_.emplace_back(TARGET(kX86));
} else if (target_repr == "npu") {
valid_places_.emplace_back(TARGET(kNPU));
} else if (target_repr == "xpu") {
valid_places_.emplace_back(TARGET(kXPU));
} else {
LOG(FATAL) << lite::string_format(
"Wrong target '%s' found, please check the command flag "
"'valid_targets'",
target_repr.c_str());
}
}
CHECK(!valid_places_.empty())
<< "At least one target should be set, should set the "
"command argument 'valid_targets'";
}
void OptBase::SetOptimizeOut(const std::string& optimized_out_path) {
optimize_out_path_ = optimized_out_path;
}
void OptBase::RunOptimize(bool record_strip_info) {
CheckIfModelSupported(false);
OpKernelInfoCollector::Global().SetKernel2path(kernel2path_map);
opt_config_.set_valid_places(valid_places_);
if (model_set_dir_ != "") {
RunOptimizeFromModelSet(record_strip_info);
} else {
auto opt_predictor = lite_api::CreatePaddlePredictor(opt_config_);
opt_predictor->SaveOptimizedModel(
optimize_out_path_, model_type_, record_strip_info);
auto resulted_model_name =
record_strip_info ? "information of striped model" : "optimized model";
std::cout << "Save the " << resulted_model_name
<< " into :" << optimize_out_path_ << "successfully";
}
}
// collect ops info of modelset
void CollectModelMetaInfo(const std::string& output_dir,
const std::vector<std::string>& models,
const std::string& filename) {
std::set<std::string> total;
for (const auto& name : models) {
std::string model_path =
lite::Join<std::string>({output_dir, name, filename}, "/");
auto lines = lite::ReadLines(model_path);
total.insert(lines.begin(), lines.end());
}
std::string output_path =
lite::Join<std::string>({output_dir, filename}, "/");
lite::WriteLines(std::vector<std::string>(total.begin(), total.end()),
output_path);
}
void OptBase::SetModelSetDir(const std::string& model_set_path) {
model_set_dir_ = model_set_path;
}
void OptBase::RunOptimizeFromModelSet(bool record_strip_info) {
// 1. mkdir of outputed optimized model set.
lite::MkDirRecur(optimize_out_path_);
auto model_dirs = lite::ListDir(model_set_dir_, true);
if (model_dirs.size() == 0) {
LOG(FATAL) << "[" << model_set_dir_ << "] does not contain any model";
}
// 2. optimize each model in inputed model set dir.
std::string model_file = opt_config_.model_file();
std::string param_file = opt_config_.param_file();
for (const auto& name : model_dirs) {
std::string input_model_dir =
lite::Join<std::string>({model_set_dir_, name}, "/");
std::string output_model_dir =
lite::Join<std::string>({optimize_out_path_, name}, "/");
if (opt_config_.model_file() != "" && opt_config_.param_file() != "") {
auto model_file_path =
lite::Join<std::string>({input_model_dir, model_file}, "/");
auto param_file_path =
lite::Join<std::string>({input_model_dir, param_file}, "/");
}
std::cout << "Start optimize model: " << input_model_dir;
opt_config_.set_model_dir(input_model_dir);
opt_config_.set_model_file(model_file);
opt_config_.set_param_file(param_file);
auto opt_predictor = lite_api::CreatePaddlePredictor(opt_config_);
opt_predictor->SaveOptimizedModel(
optimize_out_path_, model_type_, record_strip_info);
std::cout << "Optimize done. ";
}
// 3. if record_strip_info = true, we will record striping info
if (record_strip_info) {
// Collect all models information
CollectModelMetaInfo(
optimize_out_path_, model_dirs, lite::TAILORD_OPS_SOURCE_LIST_FILENAME);
CollectModelMetaInfo(
optimize_out_path_, model_dirs, lite::TAILORD_OPS_LIST_NAME);
CollectModelMetaInfo(optimize_out_path_,
model_dirs,
lite::TAILORD_KERNELS_SOURCE_LIST_FILENAME);
CollectModelMetaInfo(
optimize_out_path_, model_dirs, lite::TAILORD_KERNELS_LIST_NAME);
std::cout << "Record the information of stripped models into :"
<< optimize_out_path_ << "successfully";
}
}
void OptBase::PrintHelpInfo() {
const std::string opt_version = lite::version();
const char help_info[] =
"At least one argument should be inputed. Valid arguments are listed "
"below:\n"
" Arguments of help information:\n"
" `help()` Print help infomation\n"
" Arguments of model optimization:\n"
" `set_model_dir(model_dir)`\n"
" `set_model_file(model_file_path)`\n"
" `set_param_file(param_file_path)`\n"
" `set_model_type(protobuf|naive_buffer)`\n"
" `set_optimize_out(output_optimize_model_dir)`\n"
" `set_valid_places(arm|opencl|x86|npu|xpu)`\n"
" `run_optimize(false|true)`\n"
" ` ----fasle&true refer to whether to record ops info for "
"tailoring lib, false by default`\n"
" Arguments of model checking and ops information:\n"
" `print_all_ops()` Display all the valid operators of "
"Paddle-Lite\n"
" `print_supported_ops` Display supported operators of valid "
"places\n"
" `check_if_model_supported()` Check if the input model is "
"supported\n";
std::cout << "opt version:" << opt_version << std::endl
<< help_info << std::endl;
}
// 2. Print supported info of inputed ops
void OptBase::PrintOpsInfo(const std::set<std::string>& valid_ops) {
std::vector<std::string> lite_supported_targets = {"kHost",
"kX86",
"kCUDA",
"kARM",
"kOpenCL",
"kFPGA",
"kNPU",
"kXPU",
"kAny",
"kUnk"};
// Get the lengh of the first column: maximum length of the op_type
size_t maximum_optype_length = 0;
for (auto it = supported_ops.begin(); it != supported_ops.end(); it++) {
maximum_optype_length = it->first.size() > maximum_optype_length
? it->first.size()
: maximum_optype_length;
}
std::cout << std::setiosflags(std::ios::internal);
// Print the first row: OP_nam taget1 target2 ...
std::cout << std::setw(maximum_optype_length) << "OP_name";
for (size_t i = 0; i < lite_supported_targets.size(); i++) {
std::cout << std::setw(10) << lite_supported_targets[i].substr(1);
}
std::cout << std::endl;
// Print the name of supported ops and mark if it's supported by each target
// print the support info of inputed ops: valid_ops
for (auto op = valid_ops.begin(); op != valid_ops.end(); op++) {
std::cout << std::setw(maximum_optype_length) << *op;
// Check: If this kernel doesn't match any operator, we will skip it.
if (supported_ops.find(*op) == supported_ops.end()) {
continue;
}
// Print OP info.
auto ops_valid_places = supported_ops.at(*op);
for (size_t i = 0; i < lite_supported_targets.size(); i++) {
if (std::find(ops_valid_places.begin(),
ops_valid_places.end(),
lite_supported_targets[i]) != ops_valid_places.end()) {
std::cout << std::setw(10) << "Y";
} else {
std::cout << std::setw(10) << " ";
}
}
std::cout << std::endl;
}
}
void OptBase::DisplayKernelsInfo() { // Display kernel information
std::cout << ::paddle::lite::KernelRegistry::Global().DebugString();
}
void OptBase::PrintAllOps() {
// 1. Get supported ops on these targets
std::set<std::string> valid_ops;
for (size_t i = 0; i < supported_ops_target.size(); i++) {
auto ops = supported_ops_target[i];
valid_ops.insert(ops.begin(), ops.end());
}
// 2. Print support info of these ops
PrintOpsInfo(valid_ops);
}
void OptBase::PrintSupportedOps() {
// 1. Get the valid hardware targets
std::vector<TargetType> target_types = {};
for (size_t i = 0; i < valid_places_.size(); i++) {
target_types.push_back(valid_places_[i].target);
}
std::string targets_str = TargetToStr(target_types[0]);
for (size_t i = 1; i < target_types.size(); i++) {
targets_str = targets_str + TargetToStr(target_types[i]);
}
std::cout << "Supported OPs on '" << targets_str << "': " << std::endl;
target_types.push_back(TARGET(kHost));
target_types.push_back(TARGET(kUnk));
// 2. Get supported ops on these targets
std::set<std::string> valid_ops;
for (size_t i = 0; i < target_types.size(); i++) {
auto ops = supported_ops_target[static_cast<int>(target_types[i])];
valid_ops.insert(ops.begin(), ops.end());
}
// 3. Print support info of these ops
PrintOpsInfo(valid_ops);
}
// test whether this model is supported
void OptBase::CheckIfModelSupported(bool print_ops_info) {
// 1. parse valid places and valid targets
auto valid_ops = supported_ops_target[static_cast<int>(TARGET(kHost))];
auto valid_unktype_ops = supported_ops_target[static_cast<int>(TARGET(kUnk))];
valid_ops.insert(
valid_ops.end(), valid_unktype_ops.begin(), valid_unktype_ops.end());
for (size_t i = 0; i < valid_places_.size(); i++) {
auto target = valid_places_[i].target;
auto ops = supported_ops_target[static_cast<int>(target)];
valid_ops.insert(valid_ops.end(), ops.begin(), ops.end());
}
// get valid ops
std::set<std::string> valid_ops_set(valid_ops.begin(), valid_ops.end());
// 2.Load model into program to get ops in model
std::string prog_path = opt_config_.model_dir() + "/__model__";
if (!(opt_config_.model_file()).empty() &&
!(opt_config_.param_file()).empty()) {
prog_path = opt_config_.model_file();
}
lite::cpp::ProgramDesc cpp_prog;
framework::proto::ProgramDesc pb_proto_prog =
*lite::LoadProgram(prog_path, false);
lite::pb::ProgramDesc pb_prog(&pb_proto_prog);
// Transform to cpp::ProgramDesc
lite::TransformProgramDescAnyToCpp(pb_prog, &cpp_prog);
std::set<std::string> unsupported_ops;
std::set<std::string> input_model_ops;
for (size_t index = 0; index < cpp_prog.BlocksSize(); index++) {
auto current_block = cpp_prog.GetBlock<lite::cpp::BlockDesc>(index);
for (size_t i = 0; i < current_block->OpsSize(); ++i) {
auto& op_desc = *current_block->GetOp<lite::cpp::OpDesc>(i);
auto op_type = op_desc.Type();
input_model_ops.insert(op_type);
if (valid_ops_set.count(op_type) == 0) {
unsupported_ops.insert(op_type);
}
}
}
// 3. Print ops_info of input model and check if this model is supported
if (print_ops_info) {
std::cout << "OPs in the input model include:\n";
PrintOpsInfo(input_model_ops);
}
if (!unsupported_ops.empty()) {
std::string unsupported_ops_str = *unsupported_ops.begin();
for (auto op_str = ++unsupported_ops.begin();
op_str != unsupported_ops.end();
op_str++) {
unsupported_ops_str = unsupported_ops_str + ", " + *op_str;
}
std::vector<TargetType> targets = {};
for (size_t i = 0; i < valid_places_.size(); i++) {
targets.push_back(valid_places_[i].target);
}
std::sort(targets.begin(), targets.end());
targets.erase(unique(targets.begin(), targets.end()), targets.end());
std::string targets_str = TargetToStr(targets[0]);
for (size_t i = 1; i < targets.size(); i++) {
targets_str = targets_str + "," + TargetToStr(targets[i]);
}
LOG(ERROR) << "Error: This model is not supported, because "
<< unsupported_ops.size() << " ops are not supported on '"
<< targets_str << "'. These unsupported ops are: '"
<< unsupported_ops_str << "'.";
exit(1);
}
if (print_ops_info) {
std::cout << "Paddle-Lite supports this model!" << std::endl;
exit(1);
}
}
} // namespace lite_api
} // namespace paddle
lite/api/opt_base.h 0 → 100644
浏览文件 @ 5dd85c2c
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
/*
* This file defines Opt and basic functions about model transformation.
*/
#ifndef PADDLE_LITE_OPT_H_ // NOLINT
#define PADDLE_LITE_OPT_H_
#include <algorithm>
#include <iomanip>
#include <set>
#include <string>
#include <vector>
// stores the map that records the source_file path of each kernel.
#include "kernel_src_map.h" // NOLINT
#include "lite/api/cxx_api.h"
// version of Paddle-lite
#include "lite/core/version.h"
// model parser functions to pre-load model to verify if this model is supported
#include "lite/model_parser/compatible_pb.h"
#include "lite/model_parser/pb/program_desc.h"
#include "lite/utils/string.h"
// recorded all the ops supported by paddle-lite
#include "supported_kernel_op_info.h" // NOLINT
namespace paddle {
namespace lite_api {
/// The PaddlePredictor defines the basic interfaces for different kinds of
/// predictors.
class LITE_API OptBase {
public:
OptBase() = default;
void SetModelSetDir(const std::string &model_set_path);
void SetModelDir(const std::string &model_path);
void SetModelFile(const std::string &model_path);
void SetParamFile(const std::string &param_path);
void SetValidPlaces(const std::string &valid_places);
void SetOptimizeOut(const std::string &optimized_out_path);
// set optimized_model type
void SetModelType(std::string model_type);
// transform and save the optimized model
void RunOptimize(bool record_strip_info = false);
// fuctions of printing info
// 1. help info
void PrintHelpInfo();
// 2. PrintOpsInfo
void PrintOpsInfo(const std::set<std::string> &valid_ops =
{}); // print supported ops on target_types
void PrintAllOps(); // print all ops
void PrintSupportedOps(); // print ops supported on valid_places_
void DisplayKernelsInfo(); // Display kernel information
// 3. Check if this model is supported
void CheckIfModelSupported(bool print_ops_info = true);
private:
CxxConfig opt_config_;
// valid places for the optimized_model
std::vector<Place> valid_places_;
// filename of the optimized_model
std::string optimize_out_path_;
// type of the optimized_model, kNaiveBuffer default.
LiteModelType model_type_{LiteModelType::kNaiveBuffer};
// Dir path of a set of models, this should be combined with model
std::string model_set_dir_;
void RunOptimizeFromModelSet(bool record_strip_info = false);
};
} // namespace lite_api
} // namespace paddle
#endif // NOLINT
lite/api/paddle_api.cc
浏览文件 @ 5dd85c2c
......@@ -38,6 +38,7 @@ void Tensor::Resize(const shape_t &shape) {
tensor(raw_tensor_)->Resize(shape);
}
// Tensor::data
template <>
const float *Tensor::data() const {
return ctensor(raw_tensor_)->data<float>();
......@@ -47,15 +48,19 @@ const int8_t *Tensor::data() const {
return ctensor(raw_tensor_)->data<int8_t>();
}
template <>
const uint8_t *Tensor::data() const {
return ctensor(raw_tensor_)->data<uint8_t>();
}
template <>
const int64_t *Tensor::data() const {
return ctensor(raw_tensor_)->data<int64_t>();
}
template <>
const int32_t *Tensor::data() const {
return ctensor(raw_tensor_)->data<int32_t>();
}
// Tensor::mutable_data
template <>
int *Tensor::mutable_data(TargetType type) const {
return tensor(raw_tensor_)->mutable_data<int>(type);
......@@ -69,6 +74,10 @@ int8_t *Tensor::mutable_data(TargetType type) const {
return tensor(raw_tensor_)->mutable_data<int8_t>(type);
}
template <>
uint8_t *Tensor::mutable_data(TargetType type) const {
return tensor(raw_tensor_)->mutable_data<uint8_t>(type);
}
template <>
int64_t *Tensor::mutable_data(TargetType type) const {
return tensor(raw_tensor_)->mutable_data<int64_t>(type);
}
......@@ -116,18 +125,22 @@ void Tensor::CopyToCpu(T *data) const {
template void Tensor::CopyFromCpu<int, TargetType::kHost>(const int *);
template void Tensor::CopyFromCpu<float, TargetType::kHost>(const float *);
template void Tensor::CopyFromCpu<int8_t, TargetType::kHost>(const int8_t *);
template void Tensor::CopyFromCpu<uint8_t, TargetType::kHost>(const uint8_t *);
template void Tensor::CopyFromCpu<int, TargetType::kARM>(const int *);
template void Tensor::CopyFromCpu<float, TargetType::kARM>(const float *);
template void Tensor::CopyFromCpu<int8_t, TargetType::kARM>(const int8_t *);
template void Tensor::CopyFromCpu<uint8_t, TargetType::kARM>(const uint8_t *);
template void Tensor::CopyFromCpu<int, TargetType::kCUDA>(const int *);
template void Tensor::CopyFromCpu<int64_t, TargetType::kCUDA>(const int64_t *);
template void Tensor::CopyFromCpu<float, TargetType::kCUDA>(const float *);
template void Tensor::CopyFromCpu<int8_t, TargetType::kCUDA>(const int8_t *);
template void Tensor::CopyToCpu(int8_t *) const;
template void Tensor::CopyToCpu(float *) const;
template void Tensor::CopyToCpu(int *) const;
template void Tensor::CopyToCpu(int8_t *) const;
template void Tensor::CopyToCpu(uint8_t *) const;
shape_t Tensor::shape() const {
return ctensor(raw_tensor_)->dims().Vectorize();
......
lite/api/paddle_api.h
浏览文件 @ 5dd85c2c
......@@ -206,7 +206,7 @@ class LITE_API MobileConfig : public ConfigBase {
};
template <typename ConfigT>
std::shared_ptr<PaddlePredictor> CreatePaddlePredictor(const ConfigT&);
LITE_API std::shared_ptr<PaddlePredictor> CreatePaddlePredictor(const ConfigT&);
} // namespace lite_api
} // namespace paddle
......
lite/api/paddle_api_test.cc
浏览文件 @ 5dd85c2c
......@@ -15,9 +15,6 @@
#include "lite/api/paddle_api.h"
#include <gflags/gflags.h>
#include <gtest/gtest.h>
#include "lite/api/paddle_use_kernels.h"
#include "lite/api/paddle_use_ops.h"
#include "lite/api/paddle_use_passes.h"
#include "lite/utils/cp_logging.h"
#include "lite/utils/io.h"
DEFINE_string(model_dir, "", "");
......
lite/api/paddle_place.cc
浏览文件 @ 5dd85c2c
......@@ -45,6 +45,21 @@ std::string Place::DebugString() const {
return os.str();
}
const std::string& ActivationTypeToStr(ActivationType act) {
static const std::string act2string[] = {"unk",
"Relu",
"Relu6",
"PRelu",
"LeakyRelu",
"Sigmoid",
"Tanh",
"Swish",
"Exp"};
auto x = static_cast<int>(act);
CHECK_LT(x, static_cast<int>(ActivationType::NUM));
return act2string[x];
}
const std::string& TargetToStr(TargetType target) {
static const std::string target2string[] = {"unk",
"host",
......
lite/api/paddle_place.h
浏览文件 @ 5dd85c2c
......@@ -96,7 +96,9 @@ enum class ActivationType : int {
kLeakyRelu = 4,
kSigmoid = 5,
kTanh = 6,
kSwish = 7
kSwish = 7,
kExp = 8,
NUM = 9,
};
static size_t PrecisionTypeLength(PrecisionType type) {
......@@ -148,6 +150,8 @@ _ForEachPrecisionType(DefinePrecisionTypeTrait);
#define PRECISION(item__) paddle::lite_api::PrecisionType::item__
#define DATALAYOUT(item__) paddle::lite_api::DataLayoutType::item__
const std::string& ActivationTypeToStr(ActivationType act);
const std::string& TargetToStr(TargetType target);
const std::string& PrecisionToStr(PrecisionType precision);
......
lite/api/paddle_use_passes.h
浏览文件 @ 5dd85c2c
......@@ -24,7 +24,7 @@ USE_MIR_PASS(generate_program_pass);
USE_MIR_PASS(io_copy_kernel_pick_pass);
USE_MIR_PASS(argument_type_display_pass);
USE_MIR_PASS(runtime_context_assign_pass);
USE_MIR_PASS(graph_visualze);
USE_MIR_PASS(graph_visualize_pass);
USE_MIR_PASS(lite_conv_bn_fuse_pass);
USE_MIR_PASS(lite_fc_fuse_pass);
......@@ -46,3 +46,4 @@ USE_MIR_PASS(elementwise_mul_constant_eliminate_pass)
USE_MIR_PASS(npu_subgraph_pass);
USE_MIR_PASS(xpu_subgraph_pass);
USE_MIR_PASS(weight_quantization_preprocess_pass);
USE_MIR_PASS(quantized_op_attributes_inference_pass);
lite/api/python/CMakeLists.txt
浏览文件 @ 5dd85c2c
......@@ -2,6 +2,23 @@ if (NOT LITE_WITH_PYTHON)
return()
endif()
# to create setup.py for packeting whl for Paddle-Lite and opt
execute_process(
COMMAND git describe --tags --exact-match
WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
OUTPUT_VARIABLE PADDLE_LITE_TAG
OUTPUT_STRIP_TRAILING_WHITESPACE
)
execute_process(
COMMAND git log -1 --format=%h
WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
OUTPUT_VARIABLE PADDLE_LITE_COMMIT
OUTPUT_STRIP_TRAILING_WHITESPACE
)
configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in
${CMAKE_CURRENT_BINARY_DIR}/setup.py)
add_subdirectory(pybind)
#add_subdirectory(interface)
lite/api/python/__init__.py 0 → 100644
浏览文件 @ 5dd85c2c
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
lite/api/python/pybind/CMakeLists.txt
浏览文件 @ 5dd85c2c
set(PYBIND_DEPS pybind python paddle_api_light paddle_api)
if (NOT LITE_ON_TINY_PUBLISH)
set(PYBIND_DEPS ${PYBIND_DEPS} paddle_api_full)
set(PYBIND_DEPS ${PYBIND_DEPS} paddle_api_full opt_base)
endif()
lite_cc_library(lite_pybind SHARED SRCS pybind.cc DEPS ${PYBIND_DEPS})
......
lite/api/python/pybind/pybind.cc
浏览文件 @ 5dd85c2c
......@@ -26,13 +26,11 @@
#ifndef LITE_ON_TINY_PUBLISH
#include "lite/api/cxx_api.h"
#include "lite/api/paddle_use_passes.h"
#include "lite/api/opt_base.h"
#endif
#include "lite/api/light_api.h"
#include "lite/api/paddle_api.h"
#include "lite/api/paddle_use_kernels.h"
#include "lite/api/paddle_use_ops.h"
#include "lite/core/tensor.h"
namespace py = pybind11;
......@@ -50,10 +48,27 @@ using lite_api::PrecisionType;
using lite_api::DataLayoutType;
using lite_api::Place;
using lite::LightPredictorImpl;
using lite_api::OptBase;
#ifndef LITE_ON_TINY_PUBLISH
using lite::CxxPaddleApiImpl;
static void BindLiteCxxPredictor(py::module *m);
void BindLiteOpt(py::module *m) {
py::class_<OptBase> opt_base(*m, "Opt");
opt_base.def(py::init<>())
.def("set_model_dir", &OptBase::SetModelDir)
.def("set_modelset_dir", &OptBase::SetModelSetDir)
.def("set_model_file", &OptBase::SetModelFile)
.def("set_param_file", &OptBase::SetParamFile)
.def("set_valid_places", &OptBase::SetValidPlaces)
.def("set_optimize_out", &OptBase::SetOptimizeOut)
.def("set_model_type", &OptBase::SetModelType)
.def("run_optimize", &OptBase::RunOptimize)
.def("help", &OptBase::PrintHelpInfo)
.def("print_supported_ops", &OptBase::PrintSupportedOps)
.def("display_kernels_info", &OptBase::DisplayKernelsInfo)
.def("print_all_ops", &OptBase::PrintAllOps);
}
#endif
static void BindLiteLightPredictor(py::module *m);
static void BindLiteCxxConfig(py::module *m);
......
lite/api/python/pybind/pybind.h
浏览文件 @ 5dd85c2c
......@@ -22,11 +22,15 @@ namespace lite {
namespace pybind {
void BindLiteApi(pybind11::module *m);
void BindLiteOpt(pybind11::module *m);
PYBIND11_MODULE(lite_core, m) {
PYBIND11_MODULE(lite, m) {
m.doc() = "C++ core of Paddle-Lite";
BindLiteApi(&m);
#ifndef LITE_ON_TINY_PUBLISH
BindLiteOpt(&m);
#endif
}
} // namespace pybind
......
lite/api/python/setup.py.in 0 → 100644
浏览文件 @ 5dd85c2c
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# module of pack whl installer for Paddle-lite
import shutil
import os
from setuptools import setup, Distribution
class BinaryDistribution(Distribution):
'binary distribution'
def has_ext_modules(foo):
return True
# get paddle-lite version, if it's not based on a release tag, we use commit id instead
PADDLELITE_COMMITE = "@PADDLE_LITE_COMMIT@"
PADDLELITE_TAG = "@PADDLE_LITE_TAG@"
if PADDLELITE_TAG == "":
PADDLELITE_VERSION = PADDLELITE_COMMITE
else:
PADDLELITE_VERSION = PADDLELITE_TAG
# core lib of paddlelite is stored as lite.so
LITE_PATH = '${PADDLE_BINARY_DIR}/inference_lite_lib/python/install/lite'
PACKAGE_DATA = {'paddlelite': ['lite.so']}
# put all thirdparty libraries in paddlelite.libs
PACKAGE_DATA['paddlelite.libs'] = []
LIB_PATH = '${PADDLE_BINARY_DIR}/inference_lite_lib/python/install/libs'
if '${WITH_MKL}' == 'ON':
shutil.copy('${MKLML_SHARED_IOMP_LIB}', LIB_PATH)
shutil.copy('${MKLML_SHARED_LIB}', LIB_PATH)
PACKAGE_DATA['paddlelite.libs'] += ['libmklml_intel.so', 'libiomp5.so']
# link lite.so to paddlelite.libs
COMMAND = "patchelf --set-rpath '$ORIGIN/../libs/' ${PADDLE_BINARY_DIR}\
/inference_lite_lib/python/install/lite/lite.so"
if os.system(COMMAND) != 0:
raise Exception("patch third_party libs failed, command: %s" % COMMAND)
# remove unused paddle/libs/__init__.py
if os.path.isfile(LIB_PATH+'/__init__.py'):
os.remove(LIB_PATH+'/__init__.py')
# set dir path of each package
PACKAGE_DIR = {
# The paddle.fluid.proto will be generated while compiling.
# So that package points to other directory.
'paddlelite.libs': LIB_PATH,
'paddlelite': LITE_PATH
}
setup(
name='paddlelite',
version=PADDLELITE_VERSION,
description='Paddle-Lite Library',
packages=['paddlelite', 'paddlelite.libs'],
package_dir=PACKAGE_DIR,
package_data=PACKAGE_DATA,
distclass=BinaryDistribution
)
lite/api/test_resnet50_lite_bm.cc → lite/api/test_classify_lite_bm.cc
浏览文件 @ 5dd85c2c
......@@ -80,7 +80,7 @@ void TestModel(const std::vector<Place>& valid_places) {
fclose(fp);
}
TEST(ResNet50, test_bm) {
TEST(Classify, test_bm) {
std::vector<Place> valid_places({Place{TARGET(kBM), PRECISION(kFloat)},
Place{TARGET(kX86), PRECISION(kFloat)}});
......
lite/api/test_helper.h
浏览文件 @ 5dd85c2c
......@@ -17,6 +17,7 @@
#include <gflags/gflags.h>
#include <sys/time.h>
#include <time.h>
#include <cmath>
// for eval
DEFINE_string(model_dir, "", "model dir");
......@@ -43,5 +44,31 @@ inline double GetCurrentUS() {
return 1e+6 * time.tv_sec + time.tv_usec;
}
template <typename T>
double compute_mean(const T* in, const size_t length) {
double sum = 0.;
for (size_t i = 0; i < length; ++i) {
sum += in[i];
}
return sum / length;
}
template <typename T>
double compute_standard_deviation(const T* in,
const size_t length,
bool has_mean = false,
double mean = 10000) {
if (!has_mean) {
mean = compute_mean<T>(in, length);
}
double variance = 0.;
for (size_t i = 0; i < length; ++i) {
variance += pow((in[i] - mean), 2);
}
variance /= length;
return sqrt(variance);
}
} // namespace lite
} // namespace paddle
lite/api/transform_test.cc
浏览文件 @ 5dd85c2c
......@@ -28,11 +28,10 @@ DEFINE_int32(batch, 1, "batch");
namespace paddle {
namespace lite {
namespace test_transformer {
namespace test_transformer {
std::vector<std::string> inputed_lines;
void LoadInputLines(const char* filename) {
void load_input_lines(const char* filename) {
static const int max_line_buf_size = 100 * 1024 * 1024;
char* line_buffer = (char*)calloc(max_line_buf_size, sizeof(char)); // NOLINT
FILE* input_file = fopen(filename, "r");
......@@ -49,7 +48,7 @@ void LoadInputLines(const char* filename) {
line_buffer = NULL;
fclose(input_file);
}
void Split2(const std::string& main_str,
void split2(const std::string& main_str,
std::vector<std::string>& str_list, // NOLINT
const std::string& delimiter) {
size_t pre_pos = 0;
......@@ -75,19 +74,19 @@ void Split2(const std::string& main_str,
}
} // NOLINT
void PadBatchInput(std::vector<std::string>& input_lines, // NOLINT
int pad_idx,
int n_head,
Tensor* src_word,
Tensor* src_pos,
Tensor* src_attn_bias,
Tensor* trg_word,
Tensor* init_scores,
Tensor* init_idx,
Tensor* trg_bias,
int line_start,
int batch_size,
int bos_idx) {
void pad_batch_input(std::vector<std::string>& input_lines, // NOLINT
int pad_idx,
int n_head,
Tensor* src_word,
Tensor* src_pos,
Tensor* src_attn_bias,
Tensor* trg_word,
Tensor* init_scores,
Tensor* init_idx,
Tensor* trg_bias,
int line_start,
int batch_size,
int bos_idx) {
int max_len = 0;
int max_line = input_lines.size();
......@@ -98,27 +97,27 @@ void PadBatchInput(std::vector<std::string>& input_lines, // NOLINT
std::vector<std::string> split_str;
test_transformer::Split2(cur_line, split_str, " ");
test_transformer::split2(cur_line, split_str, " ");
batch_lines.push_back(split_str);
max_len = max_len >= split_str.size() ? max_len : split_str.size();
}
src_word->Resize(std::vector<DDim::value_type>({batch_size, max_len, 1}));
src_pos->Resize(std::vector<DDim::value_type>({batch_size, max_len, 1}));
src_word->Resize(std::vector<DDim::value_type>({batch_size, max_len}));
src_pos->Resize(std::vector<DDim::value_type>({batch_size, max_len}));
src_attn_bias->Resize(
std::vector<DDim::value_type>({batch_size, n_head, max_len, max_len}));
trg_bias->Resize(
std::vector<DDim::value_type>({batch_size, n_head, 1, max_len}));
float* src_word_data = src_word->mutable_data<float>();
float* src_pos_data = src_pos->mutable_data<float>();
std::vector<DDim::value_type>({batch_size, n_head, max_len, max_len}));
auto* src_word_data = src_word->mutable_data<int64_t>();
auto* src_pos_data = src_pos->mutable_data<int64_t>();
float* src_bias_data = src_attn_bias->mutable_data<float>();
float* trg_bias_data = trg_bias->mutable_data<float>();
for (int i = 0; i < batch_size; ++i) {
std::vector<std::string> cur_words = batch_lines[i];
int fill_len = cur_words.size();
int src_bias_start = i * n_head * max_len * max_len;
int trg_bias_start = i * n_head * max_len;
int trg_bias_start = i * n_head * max_len * max_len;
for (int j = 0; j < fill_len; ++j) {
src_word_data[i * max_len + j] = (atoi(cur_words[j].c_str()));
src_pos_data[i * max_len + j] = j;
......@@ -137,22 +136,24 @@ void PadBatchInput(std::vector<std::string>& input_lines, // NOLINT
int value_ind = j % max_len + src_bias_start;
src_bias_data[j] = src_bias_data[value_ind];
}
for (int j = trg_bias_start; j < trg_bias_start + n_head * max_len; ++j) {
for (int j = trg_bias_start;
j < trg_bias_start + n_head * max_len * max_len;
++j) {
int value_ind = j % max_len + trg_bias_start;
trg_bias_data[j] = trg_bias_data[value_ind];
}
}
trg_word->Resize(std::vector<DDim::value_type>({batch_size, 1, 1}));
auto* trg_word_data = trg_word->mutable_data<float>();
for (int i = 0; i < batch_size; ++i) {
trg_word->Resize(std::vector<DDim::value_type>({batch_size, max_len}));
auto* trg_word_data = trg_word->mutable_data<int64_t>();
for (int i = 0; i < batch_size * max_len; ++i) {
trg_word_data[i] = bos_idx;
}
init_scores->Resize(std::vector<DDim::value_type>({batch_size, 1}));
init_idx->Resize(std::vector<DDim::value_type>({batch_size}));
float* score_data = init_scores->mutable_data<float>();
float* idx_data = init_idx->mutable_data<float>();
auto* idx_data = init_idx->mutable_data<int32_t>();
for (int i = 0; i < init_scores->numel(); ++i) {
score_data[i] = 0;
}
......@@ -175,21 +176,25 @@ void PadBatchInput(std::vector<std::string>& input_lines, // NOLINT
void TestModel(const std::vector<Place>& valid_places,
const Place& preferred_place,
bool use_npu = false) {
#ifdef LITE_WITH_ARM
DeviceInfo::Init();
DeviceInfo::Global().SetRunMode(lite_api::LITE_POWER_HIGH, FLAGS_threads);
#endif
lite::Predictor predictor;
std::string test_data_path = FLAGS_input;
predictor.Build(FLAGS_model_dir, "", "", preferred_place, valid_places);
predictor.Build("",
FLAGS_model_dir + "/__model__",
FLAGS_model_dir + "/weights",
valid_places);
// predictor.Build(FLAGS_model_dir, "", "", valid_places);
int n_head = 8;
int batch_size = FLAGS_batch;
int bos_idx = 0;
int eos_idx = 1;
LOG(INFO) << "reading";
test_transformer::LoadInputLines(test_data_path.c_str());
LOG(INFO) << "reading finished";
test_transformer::load_input_lines(test_data_path.c_str());
auto* trg_bias = predictor.GetInput(6);
auto* src_word = predictor.GetInput(0);
......@@ -205,28 +210,31 @@ void TestModel(const std::vector<Place>& valid_places,
auto start = GetCurrentUS();
for (int i = 0; i < FLAGS_repeats; ++i) {
auto start_i = GetCurrentUS();
PadBatchInput(test_transformer::inputed_lines,
eos_idx,
n_head,
src_word, // src_word
src_pos, // src_pos
src_bias, // src_bias
trg_word, // trg_word
init_score, // init_score
init_idx, // init_idx
trg_bias, // trg_bias
i * batch_size,
batch_size,
bos_idx);
LOG(INFO) << "src_word:" << src_word->dims();
auto start_ii = GetCurrentUS();
LOG(INFO) << i << "->ii:" << (start_ii - start_i) / 1000.0;
pad_batch_input(test_transformer::inputed_lines,
eos_idx,
n_head,
src_word, // src_word
src_pos, // src_pos
src_bias, // src_bias
trg_word, // trg_word
init_score, // init_score
init_idx, // init_idx
trg_bias, // trg_bias
i * batch_size,
batch_size,
bos_idx);
predictor.Run();
auto start_iii = GetCurrentUS();
LOG(INFO) << i << "->iii:" << (start_iii - start_ii) / 1000.0;
auto* outs = predictor.GetOutputs();
LOG(INFO) << "out:" << (*outs)[0].dims();
auto* outs = predictor.GetOutput(0);
auto o_data = outs->data<int64_t>();
auto lod = outs->lod();
for (int i = 0; i < outs->numel(); ++i) {
LOG(INFO) << o_data[i];
}
for (int i = 0; i < lod.size(); ++i) {
for (int j = 0; j < lod[i].size(); ++j) {
LOG(INFO) << lod[i][j];
}
}
}
LOG(INFO) << "================== Speed Report ===================";
......@@ -234,25 +242,18 @@ void TestModel(const std::vector<Place>& valid_places,
<< ", warmup: " << FLAGS_warmup << ", repeats: " << FLAGS_repeats
<< ", spend " << (GetCurrentUS() - start) / FLAGS_repeats / 1000.0
<< " ms in average.";
auto* outs = predictor.GetOutputs();
for (auto out : *outs) {
LOG(INFO) << "======"
<< "here";
LOG(INFO) << out;
}
LOG(INFO) << "======"
<< "hereggg";
}
TEST(OcrAttention, test_arm) {
} // namespace lite
} // namespace paddle
using namespace paddle::lite; // NOLINT
int main(int argc, char** argv) {
gflags::ParseCommandLineFlags(&argc, &argv, true);
std::vector<Place> valid_places({
Place{TARGET(kHost), PRECISION(kFloat)},
Place{TARGET(kARM), PRECISION(kInt64)},
Place{TARGET(kARM), PRECISION(kFloat)},
Place{TARGET(kHost), PRECISION(kFloat)},
});
TestModel(valid_places, Place({TARGET(kARM), PRECISION(kFloat)}));
}
} // namespace lite
} // namespace paddle
lite/backends/arm/math/CMakeLists.txt
浏览文件 @ 5dd85c2c
......@@ -68,6 +68,8 @@ if (NOT HAS_ARM_MATH_LIB_DIR)
gemv_arm_int8.cc
conv3x3s1_direct_fp32.cc
conv3x3s2_direct_fp32.cc
conv3x3s1p01_depthwise_fp32_relu.cc
conv3x3s2p01_depthwise_fp32_relu.cc
conv3x3s1p01_depthwise_fp32.cc
conv3x3s2p01_depthwise_fp32.cc
conv3x3s1px_depthwise_fp32.cc
......
lite/backends/arm/math/activation.cc
浏览文件 @ 5dd85c2c
......@@ -700,6 +700,35 @@ void act_rsqrt<float>(const float* din, float* dout, int size, int threads) {
}
}
template <>
void act_square<float>(const float* din, float* dout, int size, int threads) {
const float* ptr_in = din;
float* ptr_out = dout;
for (int i = 0; i < size; ++i) {
ptr_out[0] = ptr_in[0] * ptr_in[0];
ptr_in++;
ptr_out++;
}
}
#ifdef LITE_WITH_TRAIN
template <>
void act_square_grad(const float* din,
const float* dout_grad,
float* din_grad,
int size,
int threads) {
const float* ptr_out_grad = dout_grad;
float* ptr_in_grad = din_grad;
for (int i = 0; i < size; ++i) {
ptr_in_grad[0] = ptr_out_grad[0] * 2.0 * din[0];
ptr_out_grad++;
ptr_in_grad++;
din++;
}
}
#endif
} // namespace math
} // namespace arm
} // namespace lite
......
lite/backends/arm/math/activation.h
浏览文件 @ 5dd85c2c
......@@ -69,6 +69,15 @@ void act_hard_sigmoid(const T* din,
template <typename T>
void act_rsqrt(const T* din, T* dout, int size, int threads);
template <typename T>
void act_square(const T* din, T* dout, int size, int threads);
#ifdef LITE_WITH_TRAIN
template <typename T>
void act_square_grad(
const T* din, const T* dout_grad, T* din_grad, int size, int threads);
#endif
} // namespace math
} // namespace arm
} // namespace lite
......
lite/backends/arm/math/beam_search.cc
浏览文件 @ 5dd85c2c
......@@ -70,7 +70,7 @@ void PruneEndBeams(const Tensor *pre_ids,
std::vector<std::vector<Item>> *items,
size_t lod_level,
int end_id) {
auto *pre_ids_data = pre_ids->data<float>();
auto *pre_ids_data = pre_ids->data<int64_t>();
auto &high_level = abs_lod[lod_level];
for (size_t src_idx = 0; src_idx < high_level.size() - 1; ++src_idx) {
size_t src_prefix_start = high_level[src_idx];
......@@ -152,10 +152,10 @@ std::vector<std::vector<Item>> SelectTopBeamSizeItems(const Tensor *pre_ids,
// find the current candidates
// auto abs_lod = framework::ToAbsOffset(scores->lod());
auto abs_lod = scores->lod();
auto *pre_ids_data = pre_ids->data<float>();
auto *pre_ids_data = pre_ids->data<int64_t>();
auto *pre_scores_data = pre_scores->data<float>();
auto *ids_data = ids ? ids->data<int>() : nullptr;
auto *ids_data = ids ? ids->data<int64_t>() : nullptr;
auto *scores_data = scores->data<float>();
size_t num_seqs = abs_lod[lod_level].size() - 1;
......@@ -236,7 +236,7 @@ void beam_search(const Tensor *pre_ids,
if (parent_idx) {
parent_idx->Resize(dims);
}
auto *selected_ids_data = selected_ids->mutable_data<float>();
auto *selected_ids_data = selected_ids->mutable_data<int64_t>();
auto *selected_scores_data = selected_scores->mutable_data<float>();
auto *parent_idx_data =
parent_idx ? parent_idx->mutable_data<int>() : nullptr;
......
lite/backends/arm/math/conv3x3s1px_depthwise_fp32.cc
浏览文件 @ 5dd85c2c
......@@ -508,6 +508,8 @@ void act_switch_3x3s1(const float* inr0,
"x5",
"x6",
"x7");
#else
#if 1 // def LITE_WITH_ARM_CLANG
#else
asm volatile(COMPUTE RELU STORE
: [r0] "+r"(inr0),
......@@ -541,6 +543,7 @@ void act_switch_3x3s1(const float* inr0,
"r3",
"r4",
"r5");
#endif
#endif
break;
case lite_api::ActivationType::kRelu6:
......@@ -593,6 +596,8 @@ void act_switch_3x3s1(const float* inr0,
"x5",
"x6",
"x7");
#else
#if 1 // def LITE_WITH_ARM_CLANG
#else
asm volatile(COMPUTE RELU RELU6 STORE
: [r0] "+r"(inr0),
......@@ -626,6 +631,7 @@ void act_switch_3x3s1(const float* inr0,
"r3",
"r4",
"r5");
#endif
#endif
break;
case lite_api::ActivationType::kLeakyRelu:
......@@ -678,6 +684,8 @@ void act_switch_3x3s1(const float* inr0,
"x5",
"x6",
"x7");
#else
#if 1 // def LITE_WITH_ARM_CLANG
#else
asm volatile(COMPUTE LEAKY_RELU STORE
: [r0] "+r"(inr0),
......@@ -711,6 +719,7 @@ void act_switch_3x3s1(const float* inr0,
"r3",
"r4",
"r5");
#endif
#endif
break;
default:
......@@ -768,6 +777,8 @@ void act_switch_3x3s1(const float* inr0,
"x5",
"x6",
"x7");
#else
#if 1 // def LITE_WITH_ARM_CLANG
#else
asm volatile(COMPUTE STORE
: [r0] "+r"(inr0),
......@@ -801,6 +812,7 @@ void act_switch_3x3s1(const float* inr0,
"r3",
"r4",
"r5");
#endif
#endif
}
}
......@@ -988,6 +1000,8 @@ void conv_3x3s1_depthwise_fp32(const float* i_data,
w8,
vbias,
act_param);
#else
#if 1 // def LITE_WITH_ARM_CLANG
#else
act_switch_3x3s1(inr0,
inr1,
......@@ -1008,6 +1022,7 @@ void conv_3x3s1_depthwise_fp32(const float* i_data,
vbias,
vbias,
act_param);
#endif
#endif
outl[0] += 4;
outl[1] += 4;
......
lite/backends/arm/math/conv5x5s2_depthwise_int8.cc
浏览文件 @ 5dd85c2c
......@@ -102,7 +102,7 @@ void conv_depthwise_5x5s2_int8(Dtype* dout,
if (h + hout_r_block > hout) {
h_kernel = hout - h;
}
int hs = h - padh;
int hs = h * 2 - padh;
int he = hs + h_kernel * 2 + 3;
#pragma omp parallel for num_threads(threads)
......
lite/backends/arm/math/conv_block_utils.h
浏览文件 @ 5dd85c2c
......@@ -2203,7 +2203,7 @@ inline void act_switch_c8_fp32(const float* din_ptr,
[cnt] "+r"(cnt_loop)
:
: "cc",
"meemory",
"memory",
"q0",
"q1",
"q2",
......
lite/backends/arm/math/conv_depthwise.h
浏览文件 @ 5dd85c2c
......@@ -207,6 +207,118 @@ void conv_depthwise_5x5s2_int8(Dtype* dout,
int padh,
ARMContext* ctx);
void conv_depthwise_3x3s1p0_bias_relu(float* dout,
const float* din,
const float* weights,
const float* bias,
bool flag_bias,
bool flag_relu,
const int num,
const int ch_in,
const int h_in,
const int w_in,
const int h_out,
const int w_out,
ARMContext* ctx);
void conv_depthwise_3x3s1p0_bias_s_relu(float* dout,
const float* din,
const float* weights,
const float* bias,
bool flag_bias,
bool flag_relu,
const int num,
const int ch_in,
const int h_in,
const int w_in,
const int h_out,
const int w_out,
ARMContext* ctx);
void conv_depthwise_3x3s1p1_bias_relu(float* dout,
const float* din,
const float* weights,
const float* bias,
bool flag_bias,
bool flag_relu,
const int num,
const int ch_in,
const int h_in,
const int w_in,
const int h_out,
const int w_out,
ARMContext* ctx);
void conv_depthwise_3x3s1p1_bias_s_relu(float* dout,
const float* din,
const float* weights,
const float* bias,
bool flag_bias,
bool flag_relu,
const int num,
const int ch_in,
const int h_in,
const int w_in,
const int h_out,
const int w_out,
ARMContext* ctx);
void conv_depthwise_3x3s2p0_bias_relu(float* dout,
const float* din,
const float* weights,
const float* bias,
bool flag_bias,
bool flag_relu,
const int num,
const int ch_in,
const int h_in,
const int w_in,
const int h_out,
const int w_out,
ARMContext* ctx);
void conv_depthwise_3x3s2p0_bias_s_relu(float* dout,
const float* din,
const float* weights,
const float* bias,
bool flag_bias,
bool flag_relu,
const int num,
const int ch_in,
const int h_in,
const int w_in,
const int h_out,
const int w_out,
ARMContext* ctx);
void conv_depthwise_3x3s2p1_bias_relu(float* dout,
const float* din,
const float* weights,
const float* bias,
bool flag_bias,
bool flag_relu,
const int num,
const int ch_in,
const int h_in,
const int w_in,
const int h_out,
const int w_out,
ARMContext* ctx);
void conv_depthwise_3x3s2p1_bias_s_relu(float* dout,
const float* din,
const float* weights,
const float* bias,
bool flag_bias,
bool flag_relu,
const int num,
const int ch_in,
const int h_in,
const int w_in,
const int h_out,
const int w_out,
ARMContext* ctx);
} // namespace math
} // namespace arm
} // namespace lite
......
lite/backends/arm/math/conv_impl.cc
浏览文件 @ 5dd85c2c
......@@ -573,6 +573,22 @@ template void conv_im2col_gemm_int8<float>(const int8_t* i_data,
ARMContext* ctx,
const float* scale);
template void im2col<float>(const float* data_im,
int channels,
int height,
int width,
int kernel_h,
int kernel_w,
int pad_top,
int pad_bottom,
int pad_left,
int pad_right,
int stride_h,
int stride_w,
int dilation_h,
int dilation_w,
float* data_col);
void conv_depthwise_3x3_fp32(const void* din,
void* dout,
int num,
......@@ -613,6 +629,26 @@ void conv_depthwise_3x3_fp32(const void* din,
act_param,
ctx);
} else {
#ifdef __aarch64__
conv_3x3s1_depthwise_fp32(reinterpret_cast<const float*>(din),
reinterpret_cast<float*>(dout),
num,
ch_out,
h_out,
w_out,
ch_in,
h_in,
w_in,
reinterpret_cast<const float*>(weights),
bias,
param,
act_param,
ctx);
#else
#ifdef LITE_WITH_ARM_CLANG
LOG(FATAL) << "fp32 depthwise conv3x3s1px doesnot support in v7-clang, "
"this can run in basic";
#else
conv_3x3s1_depthwise_fp32(reinterpret_cast<const float*>(din),
reinterpret_cast<float*>(dout),
num,
......@@ -627,6 +663,8 @@ void conv_depthwise_3x3_fp32(const void* din,
param,
act_param,
ctx);
#endif
#endif
}
} else if (stride == 2) {
if (pads_less && pad_h == pad_w && (pad < 2)) { // support pad = [0, 1]
......
lite/backends/arm/math/conv_impl.h
浏览文件 @ 5dd85c2c
......@@ -359,6 +359,24 @@ void conv_compute_2x2_3x3_small(const float* input,
const float* bias,
const operators::ConvParam& param,
ARMContext* ctx);
template <typename Dtype>
void im2col(const Dtype* data_im,
int channels,
int height,
int width,
int kernel_h,
int kernel_w,
int pad_top,
int pad_bottom,
int pad_left,
int pad_right,
int stride_h,
int stride_w,
int dilation_h,
int dilation_w,
Dtype* data_col);
} // namespace math
} // namespace arm
} // namespace lite
......
lite/backends/arm/math/increment.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/arm/math/increment.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/arm/math/layout.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/arm/math/scale.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/arm/math/scale.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/arm/math/sgemv.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/arm/math/topk.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/arm/math/topk.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/fpga/KD/debugger.hpp
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/fpga/KD/tensor.hpp
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/mlu/CMakeLists.txt 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/mlu/mlu_utils.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/mlu/target_wrapper.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/mlu/target_wrapper.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/npu/device.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/npu/device.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/opencl/CMakeLists.txt
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/opencl/cl_caller.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/opencl/cl_caller.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/opencl/cl_context.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/opencl/cl_context.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/opencl/cl_image.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/opencl/cl_kernel/image/dropout_kernel.cl 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/opencl/cl_kernel/image/lrn_kernel.cl 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/opencl/cl_kernel/image/pad2d_kernel.cl 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/opencl/cl_kernel/image/slice_kernel.cl 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/opencl/cl_runtime.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/opencl/cl_runtime.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/x86/jit/gen/blas.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/x86/jit/gen/matmul.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/backends/x86/jit/gen/seqpool.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/CMakeLists.txt
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/arena/framework.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/arena/framework.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/context.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/context.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/lite.map
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/lite_tensor_test.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/memory.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/memory.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/mir/CMakeLists.txt
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/mir/mlu_postprocess_pass.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/mir/mlu_postprocess_pass.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/mir/node.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/mir/quantized_op_attributes_inference_pass.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/mir/quantized_op_attributes_inference_pass.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/mir/ssa_graph.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/mir/subgraph_cast_display_pass.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/op_lite.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/op_lite.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/op_registry.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/op_registry.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/optimizer.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/program.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/program_fake_utils.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/tensor.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/tensor.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/type_system.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/core/version.h.in
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/demo/cxx/README.md
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/demo/cxx/cuda_demo/CMakeLists.txt 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/demo/cxx/cuda_demo/demo.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/demo/cxx/test_cv/README.md
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/gen_code/gen_code.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/gen_code/gen_code.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/CMakeLists.txt
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/activation_grad_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/activation_grad_compute.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/assign_compute.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/assign_compute.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/calib_compute.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/calib_compute.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/cast_compute.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/cast_compute.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/compare_compute.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/compare_compute.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/conv_compute.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/conv_winograd.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/elementwise_grad_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/elementwise_grad_compute.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/fill_constant_compute.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/gather_compute.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/gather_compute.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/layout_compute.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/lookup_table_compute_test.cc 已删除 100644 → 0
浏览文件 @ 1a56bcba
此差异已折叠。
lite/kernels/arm/lookup_table_dequant_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/lookup_table_dequant_compute.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/mean_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/mean_compute.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/mean_grad_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/mean_grad_compute.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/mul_grad_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/mul_grad_compute.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/scale_compute.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/scale_compute.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/sequence_conv_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/sequence_conv_compute.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/sgd_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/sgd_compute.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/arm/topk_compute.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/bm/bridges/act_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/bm/bridges/cast_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/bm/bridges/conv_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/bm/bridges/conv_transpose_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/bm/bridges/graph.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/bm/bridges/interpolate_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/bm/bridges/reduce_full_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/bm/bridges/slice_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/bm/bridges/squeeze_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/bm/bridges/utility.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/bm/bridges/yolo_box_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/bm/subgraph_compute.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/bm/subgraph_compute.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/cuda/CMakeLists.txt
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/cuda/fetch_compute.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/fpga/CMakeLists.txt
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/host/CMakeLists.txt
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/host/crf_decoding_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/host/crf_decoding_compute.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/CMakeLists.txt 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/CMakeLists.txt 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/act_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/batch_norm_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/batch_norm_op_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/conv_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/conv_op_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/elementwise_ops.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/fc_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/fc_op_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/graph.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/graph.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/paddle_use_bridges.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/pool_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/softmax_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/tensor.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/tensor.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/test_helper.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/test_helper.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/utility.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/bridges/utility.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/calib_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/calib_compute.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/io_copy_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/subgraph_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/mlu/subgraph_compute.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/npu/bridges/act_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/npu/bridges/conv_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/npu/bridges/engine.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/npu/bridges/fill_constant_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/npu/bridges/graph.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/npu/bridges/increment_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/npu/bridges/pool_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/npu/bridges/square_op_test.cc 已删除 100644 → 0
浏览文件 @ 1a56bcba
此差异已折叠。
lite/kernels/npu/bridges/utility.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/npu/bridges/utility.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/npu/subgraph_compute.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/CMakeLists.txt
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/bilinear_interp_image_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/dropout_image_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/dropout_image_compute_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/elementwise_sub_image_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/elementwise_sub_image_compute.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/grid_sampler_image_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/grid_sampler_image_compute_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/instance_norm_image_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/instance_norm_image_compute_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/lrn_image_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/lrn_image_compute_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/pad2d_image_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/pad2d_image_compute_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/slice_image_compute.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/slice_image_compute_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/opencl/test_helper.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/xpu/bridges/graph.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/kernels/xpu/subgraph_compute.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/CMakeLists.txt
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/activation_grad_ops.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/activation_grad_ops.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/activation_ops.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/activation_ops.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/batch_norm_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/beam_search_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/conv_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/conv_op.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/crf_decoding_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/crf_decoding_op.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/elementwise_grad_ops.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/elementwise_grad_ops.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/elementwise_ops.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/elementwise_ops.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/fc_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/fc_op.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/fill_constant_batch_size_like_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/fill_constant_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/fill_constant_op.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/gather_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/increment_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/io_copy_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/lookup_table_dequant_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/lookup_table_dequant_op.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/lookup_table_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/mean_grad_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/mean_grad_op.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/mean_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/mean_op.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/mul_grad_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/mul_grad_op.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/mul_op.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/op_params.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/pad2d_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/sequence_conv_op.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/sequence_conv_op.h 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/sgd_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/softmax_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/softmax_op.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/topk_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/operators/write_to_array_op.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tests/cv/cv_basic.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tests/cv/image_convert_test.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tests/kernels/CMakeLists.txt
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tests/kernels/activation_grad_compute_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tests/kernels/crf_decoding_compute_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tests/kernels/elementwise_grad_compute_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tests/kernels/lookup_table_dequant_compute_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tests/kernels/mean_compute_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tests/kernels/mul_grad_compute_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tests/kernels/sequence_conv_compute_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tests/kernels/sgd_compute_test.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tools/build.sh
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tools/build_bm.sh
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tools/build_mlu.sh 0 → 100755
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tools/ci_build.sh
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/tools/cmake_tools/gen_opencl_code.py 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/utils/cv/CMakeLists.txt
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/utils/cv/bgr_rotate.cc 0 → 100644
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/utils/cv/image2tensor.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/utils/cv/image_convert.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/utils/cv/image_flip.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/utils/cv/image_resize.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/utils/cv/image_rotate.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/utils/env.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/utils/logging.h
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/utils/replace_stl/stream.cc
浏览文件 @ 5dd85c2c
此差异已折叠。
lite/utils/string.h
浏览文件 @ 5dd85c2c
此差异已折叠。
mobile/src/operators/op_param.h
浏览文件 @ 5dd85c2c
此差异已折叠。
mobile/tools/build.sh
浏览文件 @ 5dd85c2c
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或