TensorFlow Lite实现移植到ART-Pi Smart

TensorFlow Lite 官方仓库：

https://github.com/iysheng/tflite-micro

我的ART-Pi-smart仓库：

https://gitee.com/iysheng/ART-Pi-smart

经过了多天的摸索学习，今天将 TensorFlow Lite 的框架移植到了ART-Pi Smart上，但是还是存在一点问题，主要是 C++ 和 C 语言混合编译的问题，根据目前我的总结，具体体现在：

• TensorFlow Lite 在函数中定义的静态对象无法正常构造，必须放在全局变量的位置

• 全局变量定义的对象只能使用默认的构造函数，无发使用带有参数的构造函数初始对象

针对上述两个现象，我在运行测试 Hello world 例程时，对 TensorFlow Lite 做出来的主要改动是：

 1diff--gita/kernel/bsp/imx6ull-artpi-smart/applications/tflm_tree/examples/hello_world/main_functions.ccb/kernel/bsp/imx6ull-artpi-smart/applications/tflm_tree/examples/hello_world/main_functions.cc
 2indexa585ba7..a318c32100644
 3---a/kernel/bsp/imx6ull-artpi-smart/applications/tflm_tree/examples/hello_world/main_functions.cc
 4+++b/kernel/bsp/imx6ull-artpi-smart/applications/tflm_tree/examples/hello_world/main_functions.cc
 5@@-37,18+37,30@@constexprintkTensorArenaSize=2000;
 6uint8_ttensor_arena[kTensorArenaSize];
 7}//namespace
 8
 9+statictflite::MicroErrorReportermicro_error_reporter;
10+//Thispullsinalltheoperationimplementationsweneed.
11+//NOLINTNEXTLINE(runtime-global-variables)
12+statictflite::AllOpsResolverresolver;
13+statictflite::MicroInterpreter*static_interpreter;
14+
15//ThenameofthisfunctionisimportantforArduinocompatibility.
16voidsetup(){
17+
18+DebugLog("setupbegin.
");
19tflite::InitializeTarget();
20
21+DebugLog("initialok.
");
22+
23//Setuplogging.Googlestyleistoavoidglobalsorstaticsbecauseof
24//lifetimeuncertainty,butsincethishasatrivialdestructorit'sokay.
25//NOLINTNEXTLINE(runtime-global-variables)
26-statictflite::MicroErrorReportermicro_error_reporter;
27+/*这个地方有问题*/
28error_reporter=µ_error_reporter;
29
30//Mapthemodelintoausabledatastructure.Thisdoesn'tinvolveany
31//copyingorparsing,it'saverylightweightoperation.
32+//printf("value:%x%u
",g_hello_world_model_data[0],g_hello_world_model_data_size);
33+
34model=tflite::GetModel(g_hello_world_model_data);
35if(model->version()!=TFLITE_SCHEMA_VERSION){
36TF_LITE_REPORT_ERROR(error_reporter,
37@@-57,15+69,10@@voidsetup(){
38model->version(),TFLITE_SCHEMA_VERSION);
39return;
40}
41-
42-//Thispullsinalltheoperationimplementationsweneed.
43-//NOLINTNEXTLINE(runtime-global-variables)
44-statictflite::AllOpsResolverresolver;
45-
46//Buildaninterpretertorunthemodelwith.
47-statictflite::MicroInterpreterstatic_interpreter(
48+static_interpreter=newtflite::MicroInterpreter(
49model,resolver,tensor_arena,kTensorArenaSize,error_reporter);
50-interpreter=&static_interpreter;
51+interpreter=static_interpreter;
52
53//Allocatememoryfromthetensor_arenaforthemodel'stensors.
54TfLiteStatusallocate_status=interpreter->AllocateTensors();
55@@-102,6+109,7@@intloop(){
56if(invoke_status!=kTfLiteOk){
57TF_LITE_REPORT_ERROR(error_reporter,"Invokefailedonx:%f
",
58static_cast<double>(x));
59+printf("invokestatus:%x
",invoke_status);
60return-1;
61}

参考的资料主要有：

1、tinyML基TensorFlow Lite在Arduino

2、https://github.com/iysheng/tflite-micro/blob/main/tensorflow/lite/micro/docs/new_platform_support.md

特别关键的一点是提取基础框架的文件：python3 tensorflow/lite/micro/tools/project_generation/create_tflm_tree.py -e hello_world /tmp/tflm-tree
根据官方的文档，上述命令就可以将 hello_world 工程相关的文件提取到指定的 /tmp/tflm-tree 目录，下一步做的就是将这部分内容添加到 RT-Thread Smart 工程中，编译出来。相关的代码，我都上传到了我的ART-Pi Smart的仓库。如果有问题欢迎和我讨论。展示下，运行起来的效果，为了展示需要，删除了部分空行打印：

  1U-Boot2017.03-g002b758ac9-dirty(Jul082021-10:45:12+0000)
  2CPU:Freescalei.MX6ULLrev1.1696MHz(runningat396MHz)
  3CPU:Industrialtemperaturegrade(-40Cto105C)at44C
  4Resetcause:POR
  5Model:Freescalei.MX6ULL14x14EVKBoard
  6Board:MX6ULL14x14EVK
  7DRAM:512MiB
  8MMC:FSL_SDHC:0,FSL_SDHC:1
  9Display:TFT7016(1024x600)
 10Video:1024x600x24
 11In:serial
 12Out:serial
 13Err:serial
 14switchtopartitions#0,OK
 15mmc1(part0)iscurrentdevice
 16Net:eth0:ethernet@02188000[PRIME]
 17NormalBoot
 18Hitanykeytostopautoboot:3210
 19ethernet@02188000WaitingforPHYautonegotiationtocomplete....done
 20Usingethernet@02188000device
 21TFTPfromserver10.20.52.39;ourIPaddressis10.20.52.89
 22Filename'rtthread.bin'.
 23Loadaddress:0x80001000
 24Loading:*#################################################################
 25###############################################################
 26159.2KiB/s
 27done
 28Bytestransferred=1866656(1c7ba0hex)
 29##Startingapplicationat0x80001000...
 30|/
 31-RT-ThreadSmartOperatingSystem
 32/|5.0.0buildApr302022
 332006-2020Copyrightbyrt-threadteam
 34docomponentsinitialization.
 35initializerti_board_end:0done
 36initializedfs_init:0done
 37initializert_mmcsd_core_init:0done
 38initializert_usbd_winusb_class_register:0done
 39initializefutex_system_init:0done
 40initializepmutex_system_init:0done
 41initializelwip_system_initlwIP-2.1.2initialized!
 42:0done
 43initializeulog_console_backend_init:0done
 44initializeulog_init:0done
 45initializert_hw_adc_init:0done
 46initializert_hw_csi_init:0done
 47initializert_hw_dbg_mm_init:0done
 48initializeimx6ul_eth_initredfirstva=f00here0000
 49here1111mmu_l2=c107c000
 50[31m[63]E/drv.enet:emacdeviceinitsuccess
 51[0m:0done
 52initializert_hw_i2c_init[32m[70]I/I2C:I2Cbus[i2c1]registered
 53[0m[32m[75]I/I2C:I2Cbus[i2c3]registered
 54[0m[32m[79]I/I2C:I2Cbus[i2c4]registered
 55[0m:0done
 56initializert_hw_elcd_init[86]D/drv.lcd:fbaddress=>0x811c0000
 57:0done
 58initializert_hw_pwm_init:0done
 59initializert_hw_rtc_init:0done
 60initializeimxrt_mci_init:0done
 61initializert_hw_spi_init:0done
 62initializegt911_init[31m[158]E/drv.enet:PHYLinkdown,pleasecheckthecableconnectionandlinkpartnersetting.
 63[0m[32m[192]I/SDIO:emmccardcapacity3817472KB.
 64[0mfoundpart[0],begin:2098176,size:500.0MB
 65foundpart[1],begin:526386176,size:1.0GB
 66foundpart[2],begin:1600128000,size:10.0MB
 67foundpart[3],begin:1610613760,size:200.0MB
 68[32m[328]I/touch:rt_touchinitsuccess
 69[0m[32m[332]I/gt911:touchdevicegt911initsuccess
 70[0m:0done
 71initializert_usbd_init:0done
 72initializert_hw_wdt_init:0done
 73initializenull_device_init:0done
 74initializerandom_device_init:0done
 75initializeurandom_device_init:0done
 76initializezero_device_init:0done
 77initializert_work_sys_workqueue_init:0done
 78initializeptmx_register:0done
 79initializecritical_init:0done
 80initializert_hw_ov5640_init:0done
 81initializecplusplus_system_init:0done
 82initializeelm_init:0done
 83initializedfs_romfs_init:0done
 84initializert_i2c_core_init:0done
 85initializelibc_system_init:0done
 86initializeclock_time_system_init:0done
 87initializesal_init[32m[390]I/sal.skt:SocketAbstractionLayerinitializesuccess.
 88[0m:0done
 89initializemnt_initemmcfilesysteminitializationdone!
 90:0done
 91initializeTF_lite_init[5309]D/drv.enet:enet1linkup
 92:0done
 93initializeem_initopenemfailed!
 94:-1done
 95initializeset_adc_init:0done
 96initializeset_pwm_default:0done
 97initializert_sysinfo_init:0done
 98initializefinsh_system_init:0done
 99Redsay:hellort-smart.
100openlcdokwatchdogstarting.....
101testtf.
102setupbegin.
103gohere
104initialok.
105getmodebegin.
106gethere?
107gethere000?
108constructtheclass.
109initok.
110x_value:1.0*2^-127,y_value:1.0297613*2^-7
111x_value:1.0053092*2^-5,y_value:1.8020826*2^-5
112x_value:1.0053092*2^-4,y_value:1.1584818*2^-4
113x_value:1.507964*2^-4,y_value:1.9308027*2^-4
114x_value:1.0053092*2^-3,y_value:1.0941217*2^-3
115x_value:1.2566366*2^-3,y_value:1.4802819*2^-3
116x_value:1.507964*2^-3,y_value:1.6733624*2^-3
117x_value:1.7592913*2^-3,y_value:1.8664425*2^-3
118x_value:1.0053092*2^-2,y_value:1.0619412*2^-2
119x_value:1.130973*2^-2,y_value:1.1584818*2^-2

特别地：

 1x_value:1.0*2^-127,y_value:1.0297613*2^-7
 2x_value:1.0053092*2^-5,y_value:1.8020826*2^-5
 3x_value:1.0053092*2^-4,y_value:1.1584818*2^-4
 4x_value:1.507964*2^-4,y_value:1.9308027*2^-4
 5x_value:1.0053092*2^-3,y_value:1.0941217*2^-3
 6x_value:1.2566366*2^-3,y_value:1.4802819*2^-3
 7x_value:1.507964*2^-3,y_value:1.6733624*2^-3
 8x_value:1.7592913*2^-3,y_value:1.8664425*2^-3
 9x_value:1.0053092*2^-2,y_value:1.0619412*2^-2
10x_value:1.130973*2^-2,y_value:1.1584818*2^-2
11...

上述部分是模型估算的正弦函数的数据，我绘制成图表的效果如下：
效果还是不错的，这使用的是官方例程 16 * 16 * 1 的三层模型。

原文标题：在 ART-Pi Smart 上运行 TensorFlow Lite

文章出处：【微信公众号：RTThread物联网操作系统】欢迎添加关注！文章转载请注明出处。

审核编辑：汤梓红