Perform Inference on Mini and Small Systems

Overview

Compared with mobile devices, IOT devices are equipped with MicroControllerUnits(MCUs), and are very resource-constrained due to low RAM and computation power. Therefore, AI application on IOT devices have strict restrictions on RAM and power consumption of AI model Inference.

MindSpore Lite provides a light-weight Micro solution for deploying AI models to IOT devices: converter AI models to source code of target HW, and don’t need parsing model from flatterbuf to database and compiling graph anymore. The generated source codes are very intuitive and with very small footprint and code size. It is easy to use MindSpore Lite converter tool to generates source codes for x86/ARM64/ARM32A/ARM32M platforms. For x86/ARM64/ARM32A, generated source codes call NNACL NN lib to do inference, and call CMSIS-NN lib on ARM32M instead.

Use the MindSpore Lite Converter to convert the pre-trained model into target device codes with specifying configuration.

Obtaining codegen

You can obtain codegen by any of the following ways:

Download pre-compiled Release Package from MindSpore.
Build from the source.

With below command, it converters MNIST model into codes of x86 target.

./converter_lite --fmk=TFLITE --modelFile=${model_dir}/mnist.tflite --outputFile=${SOURCE_CODE_DIR} --configFile=${COFIG_FILE}

The explicit form of configuration file please see below:

[micro_param]

# enable code-generation for MCU HW

enable_micro=true

# specify HW target, support x86,ARM32M, AMR32A, ARM64 only

target=x86

# code generation for Inference or Train

codegen_mode=Inference

# enable parallel inference or not

support_parallel=false

# enable debug

debug_mode=false

Here is the detailed description of parameters:

Parameter	Mandatory or Not	Parameter Description	Value Range	Default value
enable_micro	Yes	enable code generation or not	true, false	false
target	Yes	target platform for the generated code	x86, ARM32M, ARM32A, ARM64	x86
codegen_mode	No	generate inference or training codes	Inference, Train	Inference
supportParallel	No	generate parallel codes or not	true, false	false
debugMode	No	generate debug codes or not	true, false	false

debugMode is not available when the filesystem is not supported by os.

Please check the API Document to get the detailed API description.

The following 3 interfaces are currently not supported：

virtual std::unordered_map<String, mindspore::tensor::MSTensor *> GetOutputs() const = 0;

virtual Vector<tensor::MSTensor *> GetOutputsByNodeName(const String &node_name) const = 0;

virtual int Resize(const Vector<tensor::MSTensor *> &inputs, const Vector<Vector<int>> &dims) = 0;

Currently the code generator is only available on Linux x86_64.

After successful execution, codegen would generate a folder named mnist at the specified path. The structure of the project file is shown as follows:

mnist
├── benchmark                  # integrate debugging-related routines
│   ├── benchmark.c
│   ├── calib_output.c
│   ├── calib_output.h
│   ├── load_input.c
│   └── load_input.h
├── CMakeLists.txt
└── src                        # source files
    ├── CMakeLists.txt
    ├── net.bin                # binary model weights
    ├── net.c
    ├── net.cmake
    ├── net.h
    ├── model.c
    ├── context.c
    ├── context.h
    ├── tensor.c
    ├── tensor.h
    ├── weight.c
    └── weight.h

Codegen Directory Structure

mindspore-lite-{version}-linux-x64
└── tools
    └── codegen # dependency header files and library
        ├── include          # Header files of inference framework
        │   ├── nnacl        # nnacl operator header file
        │   └── wrapper
        ├── lib
        │   └── libwrapper.a # MindSpore Lite codegen generates code-dependent operator static library
        └── third_party
            ├── include
            │   └── CMSIS    # ARM CMSIS NN operator header files
            └── lib
                └── libcmsis_nn.a # ARM CMSIS NN operator static library

Performing Inference on STM Boards

This guide takes the deployment on STM32F746 as an example to show how the pre-complied model is built and deployed on Cortex-M platform. More information about Arm Cortex-M could be found in their Official Web Site.

STM32F746 Compile Dependencies

The generated program compilation and deployment need to install the following tools on Windows: J-Link, STM32CubeMX and GNU Arm Embedded Toolchain to perform Cross-compilation.

STM32CubeMX Windows Version >= 6.0.1
GNU Arm Embedded Toolchain >= 9-2019-q4-major-win32
J-Link Windows Version >= 6.56
GCC >= 7.3.0
CMake >= 3.18.3

STM32F746 Project Construction

The structure of the project files that needs to be managed as follows:

├── mnist              # generated inference code by codegen
├── include            # API header files (needs to be managed)
└── operator_library   # operator source code (needs to be managed)

API header files could be found in the Release Package provided by the MindSpore team.

You need to obtain the source code corresponding to the target platform because the pre-compiled static library is not provided since the Cross compilation on Cortex-M platform is complicated. The corresponding project file structure is provided in the example and you could follow the instructions shown below to copy the source code and finish the compilation.

Use codegen to compile MNIST handwriting number identification model, generate corresponding inference codes for STM32F46. The command is as follows:

./converter_lite --fmk=TFLITE --modelFile=mnist.tflite --outputFile=${SOURCE_CODE_DIR} --configFile=${COFIG_FILE}

where target is specified to ARM32M in configure file.

The generated project file structure is shown below:

├── mnist               # root of the generated code
    ├── benchmark       # generated benchmark code
    └── src             # generated model inference code

The file structure of the prepared static operator library is shown below:

├── operator_library    # operator library
    ├── include         # header files of operator library
    └── nnacl           # operator source code provided by MindSpore team
    └── wrapper         # operator source code provided by MindSpore team
    └── CMSIS           # CMSIS source code provided by Arm

arm_nnfunctions.h needs to be added when using CMSIS v5.7.0 Softmax operator.

Project Compiling

Environment testing

When programs needed for Cross-compilation are installed, add them to the Windows PATH one by one, and test them with the following instructions:
```
gcc -v               # Check GCC
arm-none-eabi-gdb -v # Check Cross compiler
jlink -v             # Check J-Link
make -v              # Check Make
```
If all success, the environment preparation is done.

Generate the initialization codes run on the STM32F746 board. (detailed code example)

start STM32CubeMX, new project and choose STM32F746IG.
Choose Makefile and generator code.
Launch cmd on the generated project root, execute make to test whether the initialization code compilation is successful.

# make success result
arm-none-eabi-size build/test_stm32f746.elf
  text    data     bss     dec     hex filename
  3660      20    1572    5252    1484 build/test_stm32f746.elf
arm-none-eabi-objcopy -O ihex build/test_stm32f746.elf build/test_stm32f746.hex
arm-none-eabi-objcopy -O binary -S build/test_stm32f746.elf build/test_stm32f746.bin

Compiling Model

Copy operator library source code and header files provided by MindSpore team to the project folder generated by STM32CubeMX.

Copy model inference code generated by codegen to the project folder generated by STM32CubeMX.

├── .mxproject
├── build             # compile output folder
├── Core
├── Drivers
├── mnist             # cortex-m7 model inference code generated by codegen
├── Makefile          # modify makefile to organize mnist && operator_library source code
├── startup_stm32f746xx.s
├── STM32F746IGKx_FLASH.ld
└── test_stm32f746.ioc

Modify makefile, organize operator library source code and generated inference code, check example to get detailed information about makefile.

# C includes
C_INCLUDES =  \
-ICore/Inc \
-IDrivers/STM32F7xx_HAL_Driver/Inc \
-IDrivers/STM32F7xx_HAL_Driver/Inc/Legacy \
-IDrivers/CMSIS/Device/ST/STM32F7xx/Include \
-Imnist/operator_library/include \                # Added, header files for operator library
-Imnist/include \                                 # Added, header files of model inference code
-Imnist/src                                       # Added, source code of model inference code
......

Add code in Core/Src/main.c to call inference API. The code can be referenced is shown below:

while (1) {
    /* USER CODE END WHILE */
    SEGGER_RTT_printf(0, "***********mnist test start***********\n");
    MSContextHandle ms_context_handle = NULL;
    ms_context_handle = MSContextCreate();
    if (ms_context_handle) {
      MSContextSetThreadNum(ms_context_handle, 1);
      MSContextSetThreadAffinityMode(ms_context_handle, 0);
    }
    int model_size = 0;
    // read net.bin
    void *model_buffer = ReadInputData("net.bin", &model_size);
    MSModelHandle model_handle = MSModelCreate();
    int ret = MSModelBuild(model_handle, model_buffer, model_size, kMSModelTypeMindIR, ms_context_handle);
    MSContextDestroy(&ms_context_handle);
    if (model_buffer) {
      free(model_buffer);
      model_buffer = NULL;
    }
    // read input_data.bin
    MSTensorHandleArray inputs_handle = MSModelGetInputs(model_handle);
    size_t inputs_num = inputs_handle.handle_num;
    void *inputs_binbuf[inputs_num];
    int inputs_size[inputs_num];
    for (size_t i = 0; i < inputs_num; ++i) {
      MSTensorHandle tensor = inputs_handle.handle_list[i];
      inputs_size[i] = (int)MSTensorGetDataSize(tensor);
    }
    ret = ReadInputsFile("input.bin" inputs_binbuf, inputs_size, (int)inputs_num);
    for (size_t i = 0; i < inputs_num; ++i) {
      void *input_data = MSTensorGetMutableData(inputs_handle.handle_list[i]);
      memcpy(input_data, inputs_binbuf[i], inputs_size[i]);
      free(inputs_binbuf[i]);
      inputs_binbuf[i] = NULL;
    }
    MSTensorHandleArray outputs_handle = MSModelGetOutputs(model_handle);
    ret = MSModelPredict(model_handle, inputs_handle, &outputs_handle, NULL, NULL);
    if (ret != kMSStatusSuccess) {
      MSModelDestroy(&model_handle);
      SEGGER_RTT_printf("MSModelPredict failed, ret: %d", kMSStatusSuccess);
      return ret;
    }
    for (size_t i = 0; i < outputs_handle.handle_num; i++) {
      MSTensorHandle output = outputs_handle.handle_list[i];
      PrintTensorHandle(output);
    }
    SEGGER_RTT_printf(0, "***********mnist test end***********\n");

Launch cmd as admin and run make to compile.
```
make
```

STM32F746 Project Deployment

Deploy executable files to the board using J-Link and perform inference.

jlinkgdbserver           # start jlinkgdbserver set target device as STM32F746IG
jlinkRTTViewer           # start jlinkRTTViewer set target devices as STM32F746IG
arm-none-eabi-gdb        # start arm-gcc gdb service
file build/target.elf    # open debugging file
target remote 127.0.0.1  # connect jlink server
monitor reset            # reset board
monitor halt             # halt board
load                     # load executable to board
c                        # perform model inference

Performing Inference on HarmonyOS Lite

Installing build environment

For the environment preparation, please refer to HarmonyOS quick start, including gn/ninja/llvm.

Connecting to the board

For Hardware environment preparation, please refer to the HarmonyOS quick start How to Develop of board Hi3516 as example.

Compiling the model

Compile MNIST model for HarmonyOS lite by using codegen:

./converter_lite --fmk=TFLITE --modelFile=mnist.tflite --outputFile=${SOURCE_CODE_DIR} --configFile=${COFIG_FILE}

Writing build scripts

For the HarmonyOS application development, please refer to demo. Copy the mnist directory generated in the previous step to any HarmonyOS source code path, assuming it is applications/sample/, and then create a new BUILD.gn file

 <harmony-source-path>/applications/sample/mnist
 ├── benchmark
 ├── CMakeLists.txt
 ├── BUILD.gn
 └── src

Download the precompile runtime component for openharmony in Download page. This is a BUILD.gn example:

import("//build/lite/config/component/lite_component.gni")
   import("//build/lite/ndk/ndk.gni")

   lite_component("mnist_benchmark") {
       target_type = "executable"
       sources = [
            "benchmark/benchmark.cc",
            "benchmark/calib_output.cc",
            "benchmark/load_input.c",
            "src/net.c",
            "src/weight.c",
            "src/session.cc",
            "src/tensor.cc",
       ]
       features = []
       include_dirs = [
            "<YOUR MINDSPORE LITE RUNTIME PATH>/runtime",
            "<YOUR MINDSPORE LITE RUNTIME PATH>/tools/codegen/include",
            "//applications/sample/mnist/benchmark",
            "//applications/sample/mnist/src",
       ]
       ldflags = [
            "-fno-strict-aliasing",
            "-Wall",
            "-pedantic",
            "-std=gnu99",
       ]
       libs = [
            "<YOUR MINDSPORE LITE RUNTIME PATH>/runtime/lib/libmindspore-lite.a",
            "<YOUR MINDSPORE LITE RUNTIME PATH>/tools/codegen/lib/libwrapper.a",
       ]
       defines = [
           "NOT_USE_STL",
           "ENABLE_NEON",
           "ENABLE_ARM",
           "ENABLE_ARM32"
       ]
       cflags = [
            "-fno-strict-aliasing",
            "-Wall",
            "-pedantic",
            "-std=gnu99",
       ]
       cflags_cc = [
           "-fno-strict-aliasing",
           "-Wall",
           "-pedantic",
           "-std=c++17",
       ]
   }

<YOUR MINDSPORE LITE RUNTIME PATH> is the path where the runtime was unzipped, e.g. “//applications/sample/mnist/mindspore-lite-1.3.0-ohos-aarch32”. Add the configuration of the mnist_benchmark component to the build/lite/components/applications.json file.

{
   "component": "mnist_benchmark",
   "description": "Communication related samples.",
   "optional": "true",
   "dirs": [
     "applications/sample/mnist"
   ],
   "targets": [
     "//applications/sample/mnist:mnist_benchmark"
   ],
   "rom": "",
   "ram": "",
   "output": [],
   "adapted_kernel": [ "liteos_a" ],
   "features": [],
   "deps": {
     "components": [],
     "third_party": []
   }
 },

Add the configuration of the mnist_benchmark component to the vendor/hisilicon/hispark_taurus/config.json.

{ "component": "mnist_benchmark", "features":[] }

Building benchmark

cd <OPENHARMONY SOURCE PATH>
hb set
.
(select ipcamera_hispark_taurus@hisilicon)
hb build mnist_benchmark

The result file is generated in out/hispark_taurus/ipcamera_hispark_taurus directory.

Running benchmark

Copy mnist_benchmark, net.bin and test data(https://download.mindspore.cn/model_zoo/official/lite/quick_start/micro/mnist.tar.gz) to the board, and run:

 OHOS # ./mnist_benchmark mnist_input.bin net.bin 1
 OHOS # =======run benchmark======
 input 0: mnist_input.bin

 loop count: 1
 total time: 10.11800ms, per time: 10.11800ms

 outputs:
 name: int8toft32_Softmax-7_post0/output-0, DataType: 43, Elements: 10, Shape: [1 10 ], Data:
 0.000000, 0.000000, 0.003906, 0.000000, 0.000000, 0.992188, 0.000000, 0.000000, 0.000000, 0.000000,
 ========run success=======

Register Kernel

Currently, Users can only register their own kernels for custom operator. We will support registering the built-in operators’ kernels in the future. We use Hi3516D board as an example to show you how to use kernel register in codegen.

For how to register custom operators, please refer to Usage Description of the Integrated NNIE.

Run codegen

Codegen can generate custom kernel’s function declaration and reference code if the model has custom operators.

./converter_lite --fmk=TFLITE --modelFile=mnist.tflite --outputFile=${SOURCE_CODE_DIR} --configFile=${COFIG_FILE}

where target sets to be ARM32A.

Implement custom kernel by users

A header file named registered_kernel.h in the generated files. The custom kernel function is declared in this file:

int CustomKernel(TensorC *inputs, int input_num, TensorC *outputs, int output_num, CustomParameter *param);

Users need to implement this function then add their source files to the cmake project. For example, we provide a sample library named libmicro_nnie.so in the nnie runtime package, download. The library contains the implementation of custom kernel for NNIE. Users can download it and modify the CMakeLists.txt：

link_directories(<YOUR_PATH>/mindspore-lite-1.5.0-linux-aarch32/providers/Hi3516D)
link_directories(<HI3516D_SDK_PATH>)
target_link_libraries(benchmark net micro_nnie nnie mpi VoiceEngine upvqe securec -lm -pthread)

Finally, we build the benchmark:

cd nnie && mkdir buid && cd build
cmake -DCMAKE_TOOLCHAIN_FILE=<MS_SRC_PATH>/mindspore/lite/cmake/himix200.toolchain.cmake -DPLATFORM_ARM32=ON -DPKG_PATH=<RUNTIME_PKG_PATH> ..
make

Perform Inference on Mini and Small Systems

Overview

Obtaining codegen

Codegen Directory Structure

Performing Inference on STM Boards

STM32F746 Compile Dependencies

STM32F746 Project Construction

Project Compiling

Compiling Model

STM32F746 Project Deployment

Performing Inference on HarmonyOS Lite

Installing build environment

Connecting to the board

Compiling the model

Writing build scripts

Building benchmark

Running benchmark

Register Kernel

Run codegen

Implement custom kernel by users

More Details

Linux_x86_64 platform compile and deploy 

Android platform compile and deploy 

Perform Inference on Mini and Small Systems

Overview

Obtaining codegen

Codegen Directory Structure

Performing Inference on STM Boards

STM32F746 Compile Dependencies

STM32F746 Project Construction

Project Compiling

Compiling Model

STM32F746 Project Deployment

Performing Inference on HarmonyOS Lite

Installing build environment

Connecting to the board

Compiling the model

Writing build scripts

Building benchmark

Running benchmark

Register Kernel

Run codegen

Implement custom kernel by users

More Details

Linux_x86_64 platform compile and deploy

Android platform compile and deploy

Linux_x86_64 platform compile and deploy 

Android platform compile and deploy 