Qualcomm TensorFlow Lite SDK Jagorar Mai Amfani
Tarihin bita
| Bita | Kwanan wata | Bayani |
| AA | Satumba 2023 | Sakin farko |
| AB | Oktoba 2023 |
|
Gabatarwa zuwa kayan aikin Qualcomm TFLite SDK
Kayan aikin haɓaka software na Qualcomm TensorFlow Lite (Qualcomm TFLite SDK) yana ba da tsarin TensorFlow Lite don ƙaddamar da hankali na wucin gadi (AI), wanda ke sauƙaƙe masu haɓaka aikace-aikacen don haɓaka ko gudanar da aikace-aikacen AI masu dacewa.
Wannan daftarin aiki yana ba da umarnin mataki-mataki don haɗa Qualcomm TFLite SDK na tsaye da kafa yanayin ci gaba. Wannan yana ba da damar aikin haɓakawa, wanda ya haɗa da:
- kafa yanayin ginin inda mai haɓakawa zai iya tattara Qualcomm TFLite SDK
- haɓaka aikace-aikacen Qualcomm TFLite SDK na tsaye
Don tallafi, dubahttps://www.qualcomm.com/tallafi. Hoto mai zuwa yana ba da taƙaitaccen aikin Qualcomm TFLite SDK:
Hoto 1-1 Qualcomm TFLite SDK aikin aiki
Kayan aikin yana buƙatar dandamali SDK da tsari file (Tsarin JSON) don samar da kayan tarihi na Qualcomm TFLite SDK.
Don gina aikace-aikacen ƙarshen-zuwa-ƙarshe ta amfani da multimedia, AI, da na'urorin hangen nesa na kwamfuta (CV), duba Qualcomm Intelligent Multimedia SDK (QIM SDK) Jagoran Fara Sauri (80-50450-51).
Tebur yana nuna taswirar sigar SDK ta Qualcomm TFLite tare da sakin CodeLinaro tag:
Tebur 1-1 Bayanin fitarwa
| Qualcomm TFlite SDK version | Sakin CodeLinaro tag |
| V1.0 | Qualcomm TFLITE.SDK.1.0.r1-00200-TFLITE.0 |
Tebur 1-2 Masu Goyan bayan Qualcomm TFLite SDK iri
| Qualcomm Sigar TFlite SDK | Samfurin software mai goyan baya | Sigar TFlite mai goyan baya |
| V1.0 | QCS8550.LE.1.0 |
|
|
||
|
||
|
||
|
||
|
Magana
Table 1-3 Takardu masu alaƙa
| Take | Lamba |
| Qualcomm | |
| 00067.1 Bayanin Sakin don QCS8550.LE.1.0 | RNO-230830225415 |
| SDK Multimedia Intelligent Qualcomm (QIM SDK) Jagoran Fara Mai Sauri | 80-50450-51 |
| Bayanin Qualcomm Multimedia SDK (QIM SDK) Magana | 80-50450-50 |
| Albarkatu | |
| https://source.android.com/docs/setup/start/initializing | – |
Tebur 1-4 Gagaru da ma'anoni
| Acronym ko kalma | Ma'anarsa |
| AI | Hankali na wucin gadi |
| BIOS | Tsarin shigarwa / fitarwa na asali |
| CV | Kwamfuta hangen nesa |
| IPK | Kunshin Itsy file |
| QIM SDK | Kit ɗin haɓaka software na Qualcomm Intelligent multimedia |
| SDK | Kayan haɓaka software |
| TFlite | TensorFlow Lite |
| XNN | Xth makwabci mafi kusa |
Sanya yanayin gini don kayan aikin Qualcomm TFLite SDK
Ana fitar da kayan aikin Qualcomm TFLite SDK a cikin sigar tushe; don haka, kafa yanayin gini don haɗa shi ya zama dole amma saitin lokaci ɗaya.
Abubuwan da ake bukata
- Tabbatar cewa kuna da sudoaccess zuwa na'ura mai masaukin Linux.
- Tabbatar cewa sigar uwar garken Linux ita ce Ubuntu 18.04 ko Ubuntu 20.04.
- Ƙara iyakar agogon mai amfani da matsakaicin misalan mai amfani akan tsarin runduna.
- Ƙara waɗannan layin umarni zuwa /etc/sysctl.confand sake kunna mai watsa shiri: fs.inotify.max_user_instances=8192 fs.inotify.max_user_watches=542288
Shigar da fakitin masaukin da ake buƙata
An shigar da fakitin rundunar a kan na'ura mai watsa shiri na Linux.
Gudun umarni don shigar da fakitin mai watsa shiri: $ sudo dace shigar -y jq $ sudo dace shigar -y texinfo chrpath libxml-simple-perl openjdk-8-jdkheadless
Don Ubuntu 18.04 da sama:
$ sudo dace-samun shigar git-core gnupg flex bison gina-mahimmanci zip curl zlib1g-dev gcc-multilib g++-multilib libc6-dev-i386 libncurses5 lib32ncurses5- dev x11proto-core-dev libx11-dev lib32z1-dev libgl1-mesa-dev libxml2-utils xslt
Don ƙarin bayani, duba https://smu.android.com/docs/setup/farawa/farawa.
Saita mahallin docker
Docker dandamali ne da ake amfani da shi don ginawa, haɓakawa, gwadawa, da isar da software. Don haɗa SDK, dole ne a saita docker akan na'ura mai ɗaukar hoto na Linux.
Tabbatar cewa an kunna aikin haɓakawa na CPU akan na'ura mai ɗaukar hoto na Linux. Idan ba a kunna shi ba, yi waɗannan don kunna shi daga ainihin tsarin shigarwa/fitarwa (BIOS) saitunan daidaitawa:
- Kunna hangen nesa daga BIOS:
a. Latsa F1 ko F2 lokacin da tsarin ke tashi don shiga BIOS. Ana nuna taga BIOS.
b. Canja zuwa Babba shafin.
c. A cikin sashin Kanfigareshan CPU, saita Fasahar Mahimmanci don Kunnawa.
a. Latsa F12 don ajiyewa da fita, sannan kuma sake kunna tsarin.
Idan waɗannan matakan ba su yi aiki ba, bi takamaiman umarnin daga mai samar da tsarin don kunna haɓakawa - Cire duk wani tsohon misali na docker:
$ sudo dace cire docker-desktop
$ rm -r $HOME/.docker/desktop
$ sudo rm /usr/local/bin/com.docker.cli
$ sudo apt share docker-desktop - Saita ma'ajiyar nesa ta docker:
$ sudo dace-samun sabuntawa $ sudo dace-samu shigar ca-certificates curl gnupg lsb-saki $ sudo mkdir -p /etc/apt/keyrings $ curl -fsSL https://download.docker.com/linux/ubuntu/gpg | sudo gpg - dearmor -o /etc/apt/keyrings/docker.gpg $ echo "deb [arch=$(dpkg -print-architecture) sanya hannu-by=/etc/apt/ keyrings/ docker.gpg] https:// download.docker.com/linux/ubuntu $ (lsb_release -cs) barga" | sudo tee /etc/apt/sources.list.d/ docker.list > /dev/null - Sanya injin docker:
$ sudo dace-samun sabuntawa $ sudo dace-samu shigar docker-ce docker-ce-cli - Ƙara mai amfani zuwa rukunin docker:
$ sudo groupadd docker $ sudo usermod -aG docker $ USER - Sake kunna tsarin.
Ƙirƙirar dandalin SDK
SDK dandali wajibi ne don haɗa kayan aikin Qualcomm TFLite SDK. Yana ba da duk abubuwan dogaro da dandamali da ake buƙata ta Qualcomm TFLite SDK.
Yi abubuwan da ke biyowa don samar da dandalin SDK:
- Ƙirƙiri gini don samfurin software da aka fi so.
An bayar da umarnin gina sakin QCS8550.LE.1.0 a cikin bayanan saki. Don samun damar bayanin kula, duba References.
Idan hotunan an gina su a baya, aiwatar da mataki na 2, sannan ƙirƙirar gini mai tsabta. - Gudun umarni mai zuwa don gina hotunan sararin mai amfani da dandamalin SDK:
Don QCS8550.LE.1.0, ƙara fasalin injin qti-tflite-delegate a cikin MACHINE_FEATURES a cikin kalama.conf file kuma tushen yanayin ginin bisa ga umarnin daga bayanan sakin.
Bayan samar da hotunan sarari mai amfani daga ginin, gudanar da umarni mai zuwa don samar da dandalin SDK.
$ bitbake -fc populate_sdk qti-robotics-hoton
Gina kayan aikin Qualcomm TFLite SDK - haɓaka aikin haɓakawa
Gudun aikin kayan aikin Qualcomm TFlite SDK yana buƙatar mai haɓakawa don samar da tsari file tare da ingantaccen shigarwar shigarwa. Rubutun harsashi masu taimako daga aikin tflite-kayan aikin (wanda yake a cikin bishiyar tushen Qualcomm TFLite SDK) yana ba da ayyuka masu amfani don saita yanayin harsashi, wanda za'a iya amfani dashi don aikin Qualcomm TFLite SDK.
Mai haɓakawa yana gina ayyukan Qualcomm TFLite SDK a cikin akwati kuma yana samar da kayan tarihi ta amfani da kayan aikin da tflite-kayan aikin ke bayarwa.
Bayan an gina akwati na Qualcomm TFLite SDK, mai haɓakawa zai iya haɗawa da akwati kuma yayi amfani da kayan aikin taimako a cikin yanayin harsashi na kwantena don ci gaba da haɓakawa.
- Akwai tanadi don shigar da kayan tarihi na Qualcomm TFLite SDK zuwa na'urar Qualcomm da aka haɗa da rundunar Linux ta USB/adb.
- Hakanan akwai tanadi don kwafin kayan tarihi na Qualcomm TFLite SDK daga akwati zuwa na'ura mai masaukin baki daban-daban inda aka haɗa na'urar Qualcomm.
Hoton da ke gaba yana lissafin saitin abubuwan amfani da ke akwai bayan kafa yanayin ginin kwantena ta amfani da rubutun taimako don gina Qualcomm TFLite SDK.
Hoton yana nuna jerin aiwatar da kayan aikin:
Hoto 4-3 Jerin abubuwan amfani akan mai masaukin baki
Daidaita kuma gina Qualcomm TFLite SDK
An haɗa Qualcomm TFLite SDK lokacin da aka ƙirƙiri hoton docker. Don daidaitawa da gina Qualcomm TFLite SDK, yi masu zuwa:
- Ƙirƙiri adireshi akan mai watsa shiri file tsarin aiki tare da Qualcomm TFLite SDK wurin aiki. Domin
exampda: $mkdir $cd - Nemo lambar tushe na Qualcomm TFLite SDK daga CodeLinaro:
$ repo init -u https://git.codelinaro.org/clo/le/sdktflite/tflite/ manifest.git -repo-branch=qc/stable -repo-url=git://git.quicinc.com/ tools/repo.git -m TFLITE.SDK.1.0.r1-00200-TFLITE.0.xml -b saki && repo sync -qc -no-tags -j - Ƙirƙiri adireshi akan mai watsa shiri file tsarin da za a iya sakawa cikin docker. Don misaliample: mkdir-p / Ana iya ƙirƙirar wannan jagorar a ko'ina akan na'ura mai watsa shiri na Linux, kuma bai dogara da inda aka daidaita aikin Qualcomm TFLite SDK ba. Bayan an kammala aikin aiki a cikin akwati, ana iya samun kayan aikin Qualcomm TFLite SDK a cikin littafin da aka ƙirƙira a wannan matakin.
- Gyara tsarin JSON file samuwa a /tflite-tools/ targets/le-tflite-tools-builder.json tare da shigarwar masu zuwa:
"Hoto": "tflite-tools-builder", "Device_OS": "le", "Ƙari_tag": "", "TFLite_Version": "2.11.1", "Delegates": {"Hexagon_delegate": "KASHE", "Gpu_delegate": "ON", "Xnnpack_delegate": "ON"}, "TFLite_rsync_destination": " /", "SDK_path": "/build-qti-distro-fullstack-perf/tmpglibc/deploy/sdk>", "SDK_shell_file": "", "Base_Dir_Location": "" }
Don ƙarin bayani kan shigarwar da aka ambata a cikin tsarin json file, duba Docker.md readme file a /tflite-tools/.
NOTE Don QCS8550, ba a tallafawa Qualcomm® Hexagon™ DSP wakilin. - Tushen rubutun don saita yanayin:
$ cd /tflite-tools $ tushen ./scripts/host/docker_env_setup.sh - Gina hoton docker na Qualcomm TFLite SDK: $ tflite-tools-host-build-image ./targets/le-tflite-tools-builder.json Idan saitin ginin ya gaza, duba saitin docker na matsala. Bayan an gama nasara, ana nuna saƙo mai zuwa: “Status:Gina hoton da aka kammala cikin nasara!!” Gudun wannan matakin yana gina Qualcomm TFLite SDK shima.
- Gudanar da kwandon docker na Qualcomm TFLite SDK. Wannan yana farawa akwati da tags an bayar a cikin tsarin JSON file. $tflite-tools-host-run-container ./targets/le-tflite-tools-builder.json
- Haɗe da kwandon da aka fara daga mataki na baya.
$ docker abin da aka makala
An harhada Qualcomm TFLite SDK, kuma kayan tarihi a shirye suke don turawa ko kuma ana iya ƙarawa.
da aka yi amfani da shi don samar da filogi na QIM SDK TFLite.
Haɗa na'urar zuwa ɗaukar nauyi da tura kayan tarihi]
Bayan haɗawa, akwai hanyoyi guda biyu don haɗa na'urar zuwa mai watsa shiri da tura ta
Qualcomm TFLite SDK kayan tarihi.
- Na'urar da aka haɗa zuwa mai masaukin Linux na gida:
Mai haɓakawa yana haɗa na'urar zuwa wurin aiki kuma yana shigar da kayan tarihi na Qualcomm TFLite SDK daga akwati kai tsaye akan na'urar (QCS8550). - Na'urar da aka haɗa zuwa mai watsa shiri mai nisa:
Mai haɓakawa yana haɗa na'urar zuwa wurin aiki mai nisa, kuma suna iya amfani da umarnin mai sakawa mai sarrafa fakiti akan dandamali na Windows da Linux don shigar da kayan tarihi na Qualcomm TFLite SDK zuwa na'urar (QCS8550)
Hoto 4-4 Haɗin allon na'ura zuwa mai haɓakawa da wurin aiki mai nisa
Haɗa na'urar zuwa wurin aiki
An haɗa na'urar zuwa wurin aiki kuma kwandon haɓaka na iya samun damar na'urar akan USB/adb.
Adadin ya nuna stages a cikin jerin Qualcomm TFLite SDK aiki:
- Gudun waɗannan umarni don shigar da kayan tarihi a na'urar:
$ tflite-kayan aiki-na'urar-shirya
$ tflite-kayan aiki-na'urar-aikawa - Don cire kayan tarihi, gudanar da umarni mai zuwa:
$ tflite-kayan aiki-na'urar- fakitin-cire
Haɗa na'urar zuwa na'ura mai nisa
An haɗa na'urar zuwa na'ura mai nisa, kuma akwati na Qualcomm TFLite SDK ba zai iya shiga na'urar akan USB/ad b.
Adadin ya nuna stages a cikin jerin Qualcomm TFLite SDK aiki:
Gudanar da waɗannan umarni a cikin kwandon tflite-kayan aikin don kwafi kayan tarihi zuwa na'ura mai nisa
dangane da mai sarrafa kunshin akan na'urar:
$ tflite-kayan aikin-nesa-sync-ipk-rel-pkg
NOTE Ana ba da bayanin injin nesa a cikin tsarin JSON file.
Shigar da kayan tarihi don dandalin Windows
Za a iya shigar da kayan tarihi na Qualcomm TFLite SDK akan na'urar bisa tsarin aiki na na'ura mai nisa.
Don dandamali na Windows, yi masu zuwa:
A kan PowerShell, yi amfani da rubutun mai zuwa: PS C:
Adb tushen PS C:> adb disable-verity PS C:> adb sake yi PS C:> adb jira-for-na'urar PS C:> adb tushen PS C:> adb remount PS C:> adb harsashi Dutsen -o remount, rw / PS C:> adb harsashi “mkdir -p / tmp” PS C:> adb push /tmp Idan kunshin ipk ne (don QCS8550.LE.1.0), yi amfani da umarni masu zuwa: PS C:> adb harsashi “ opkg –force-depends –force-reinstall –force-overwrite install /tmp/”
Shigar da kayan tarihi don dandalin Linux
Yi amfani da umarni masu zuwa:
$ adb tushen $ adb kashe-gaskiya $ adb sake yi $ adb jira-na'urar $ adb tushen $ adb remount $ adb harsashi Dutsen -o remount,rw / $ adb harsashi "mkdir -p / tmp" $ adb tura / tmp Idan Kunshin IPk ne (na QCS8550.LE.1.0): $ adb harsashi “opkg –force-depends –force-reinstall –force-overwrite install /tmp/”
Share hoton docker
Bayan kammala aikin mai haɓakawa, ya kamata a tsaftace mahallin docker don 'yantar da ajiya akan faifai. Share docker yana cire kwantena da hotuna da ba a yi amfani da su ba, don haka yantar da sararin diski.
Yi amfani da waɗannan umarni don tsaftace hoton docker:
- Gudun umarni mai zuwa akan ma'ajin Linux:
$ cd /tflite-kayan aiki - Tsaya kwandon:
$ tflite-tools-host-stop-container ./targets/le-tflite-tools-builder.json - Cire akwati:
$ tflite-tools-host-rm-container ./targets/ le-tflite-tools-builder.json - Cire tsoffin hotunan docker:
$ tflite-kayan aiki-bautar-hotuna-tsaftacewa
Shirya matsala saitin docker
Idan umarnin tflite-tools-host-build-image ya dawo da Nospace da aka bari akan saƙon na'ura, sannan matsar da directory ɗin docker zuwa/local/mnt. Yi waɗannan abubuwan don magance saitin:
- Ajiye mai docker da ke akwai files:
$ tar -zcC /var/lib docker> /mnt/pd0/var_lib_docker-backup-$(kwana +%s).tar.gz - Dakatar da docker:
tashar tashar sabis $ - Tabbatar cewa babu aikin docker da ke gudana:
$ps faux | grep docker - Duba tsarin kundin adireshi:
$ sudo ls /var/lib/docker/ - Matsar da adireshin docker zuwa sabon bangare:
$ mv /var/lib/docker /local/mnt/docker - Yi alamar haɗin kai zuwa kundin adireshin docker a cikin sabon bangare:
$ ln -s /local/mnt/docker /var/lib/docker - Tabbatar cewa tsarin adireshin docker ya kasance baya canzawa:
$ sudo ls /var/lib/docker/ - Fara docker:
$ fara docker sabis - Sake kunna duk kwantena bayan matsar da adiresoshin docker.
Ƙirƙirar TFlite SDK tare da aikin Linux
Ana iya kunna aikin TFLite SDK ba tare da kwantena ta amfani da aikin Linux ba. Wannan hanya madadin yin amfani da kwantena.
Don daidaitawa da gina Qualcomm TFLite SDK, yi masu zuwa:
- Ƙirƙiri adireshi akan mai watsa shiri file tsarin aiki tare da Qualcomm TFLite SDK wurin aiki. Don misaliampda:
$mkdir
$cd - Nemo lambar tushe na Qualcomm TFLite SDK daga CodeLinaro:
$ repo init -u https://git.codelinaro.org/clo/le/sdktflite/tflite/ manifest.git -repo-branch=qc/stable -repo-url=git://git.quicinc.com/ tools/repo.git -m TFLITE.SDK.1.0.r1-00200-TFLITE.0.xml -b saki && repo sync -qc -no-tags -j8 && repo sync -qc –no-tags -j8 - 3. Gyara tsarin JSON file ba a /tflite-tools/ targets/le-tflite-tools-builder.json tare da shigarwar masu zuwa
"Hoto": "tflite-tools-builder", "Device_OS": "le", "Ƙari_tag": "", "TFLite_Version": "2.11.1", "Delegates": {"Hexagon_delegate": "KASHE", "Gpu_delegate": "ON", "Xnnpack_delegate": "ON"}, "TFLite_rsync_destination": " ", "SDK_path": "/build-qti-distro-fullstack-perf/tmpglibc/deploy/sdk>", "SDK_shell_file": "", "Base_Dir_Location": ""
Don ƙarin bayani kan shigarwar da aka ambata a cikin tsarin json file, duba Docker.md readme file a /tflite-kayan aiki/.
NOTE Don QCS8550, ba a tallafawa Hexagon DSP wakilin - Tushen rubutun don saita yanayin:
$ cd /tflite-kayan aiki
tushen $ ./scripts/host/host_env_setup.sh - Gina Qualcomm TFLite SDK.
$ tflite-tools-setup targets/le-tflite-tools-builder.json - Gudun waɗannan umarni masu amfani a cikin harsashi na Linux iri ɗaya don tattara kayan tarihi na TFLite SDK daga
TFlite_rsync_makomawa.
$ tflite-tools-host-get-rel-package targets/le-tflite-tools-builder.json
$ tflite-tools-host-get-dev-package targets/le-tflite-tools-builder.json - Shigar da kayan tarihi bisa tsarin aiki
- Don dandalin Windows, akan PowerShell, yi amfani da rubutun mai zuwa
PS C:> adb tushen PS C:> adb disable-verity PS C:> adb sake yi PS C:> adb jira-for-na'urar PS C:> adb tushen PS C:> adb remount PS C:> adb harsashi Dutsen - o remount, rw / PS C:> adb harsashi "mkdir -p / tmp" PS C:> adb tura / tmp
Idan fakitin IPk ne (na QCS8550.LE.1.0), yi amfani da umarni masu zuwa:
PS C:> adb harsashi “opkg –force-depends –force-reinstall –forceoverwrite install /tmp/
Don dandalin Linux, yi amfani da rubutun mai zuwa:
$ adb tushen $ adb disable-gaskiya $ adb sake yi $ adb jira-na'urar $ adb tushen $ adb remount $ adb harsashi Dutsen -o remount,rw / $ adb harsashi "mkdir -p / tmp" $ adb tura / tmp Idan fakitin IPk ne (na QCS8550.LE.1.0):
$ adb harsashi “opkg –force-depends –force-reinstall –force-overwrite install /tmp/”
- Don dandalin Windows, akan PowerShell, yi amfani da rubutun mai zuwa
Ƙirƙirar Qualcomm TFLite SDK kayan tarihi don gina QIM SDK
Don amfani da kayan tarihi da aka ƙirƙira don kunna Qualcomm TFLite SDK GStreamer plug-in a cikin QIM SDK, yi haka:
- Kammala hanya a cikin Daidaitawa kuma gina Qualcomm TFLite SDK, sannan gudanar da umarni mai zuwa: $ tflite-tools-host-get-dev-tar-package ./targets/le-tflite-toolsbuilder.json
A tar file ake haifarwa. Ya ƙunshi Qualcomm TFLite SDK a hanyar da aka bayar a "TFLite_rsync_destination" - Don kunna Qualcomm TFLite SDK GStreamer plug-in, yi amfani da tar file a matsayin hujja a cikin tsarin JSON file don gina QIM SDK.
Don bayani kan haɗa QIM SDK, duba Qualcomm Intelligent Multimedia SDK (QIM SDK) Jagoran Farawa Mai sauri (80-50450-51).
Gina Qualcomm TFLite SDK na karuwa
Idan kuna gina Qualcomm TFLite SDK a karon farko, duba Gina kayan aikin Qualcomm TFLite SDK - masu haɓaka aikin haɓakawa. Ana iya sake amfani da yanayin gini iri ɗaya don haɓaka haɓakawa.
Kayan aikin taimako (a cikin akwati) da aka ambata a cikin adadi suna samuwa ga masu haɓakawa don haɗa aikace-aikacen da aka gyara da filogi.
Hoto 5-1 Gudun aiki a cikin akwati
Bayan an kammala canje-canjen lambar a cikin kundin adireshin, yi haka:
- Haɗa lambar da aka gyara:
$ tflite-kayan aikin-ƙara-build-shigar - Kunshin da aka haɗa lambar:
$ tflite-kayan aikin-ipk-rel-pkg ko $ tflite-kayan aikin-deb-rel-pkg - Daidaita fakitin saki tare da mai watsa shiri file tsarin:
$ tflite-kayan aikin-nesa-sync-ipk-rel-pkg
Or
$ tflite-kayan aikin-nesa-sync-deb-rel-pkg - Shirya fakitin dev:
$ tflite-kayan aiki-ipk-dev-pkg
Ana samun kayan tarihin da aka haɗa a cikin babban fayil ɗin TFLite_rsync_destination da aka ambata a cikin JSON file, wanda za a iya kwafi zuwa kowane kundin adireshi.
Yi aiki tare da Wakilin TFlite na waje na QNN
Wakilin Waje na TFLite yana ba ku damar gudanar da samfuran ku (ɓangare ko gabaɗaya) akan wani mai zartarwa ta amfani da ɗakunan karatu waɗanda amintaccen ɓangare na uku suka samar kamar QNN ta Qualcomm. Wannan tsarin na iya yin amfani da nau'ikan masu haɓaka kayan aiki iri-iri kamar GPU ko Hexagon Tensor Processor (HTP) don ƙaddamarwa. Wannan yana ba wa masu haɓaka hanyar sassauƙa da ɓata lokaci daga tsohowar TFLite don haɓaka ƙima.
Abubuwan da ake buƙata:
- Tabbatar cewa kun yi amfani da wurin aiki na Ubuntu don cire tarin QNN AI.
- Tabbatar cewa kuna amfani da sigar QNN 2.14 don kasancewa tare da Qualcomm TFLite SDK
An kunna Qualcomm TFLite SDK don gudanar da ra'ayoyi a kan ƙarshen bayan QNN da yawa ta hanyar TFLite Wakilin waje na QNN. Samfuran TFlite tare da wakilcin flatbuffer gama gari ana iya gudana akan GPU da HTP.
Bayan an shigar da fakitin Qualcomm TFLite SDK akan na'urar, yi waɗannan don shigar da ɗakunan karatu na QNN akan na'urar.
- Zazzage Manajan Package na Qualcomm 3 don Ubuntu.
a. Dannahttps://qpm.qualcomm.com/, kuma danna Kayan aiki.
b. A cikin sashin hagu, a cikin filin Kayan aikin Bincike, rubuta QPM. Daga cikin System OS list, zaɓi Linux.
Sakamakon binciken yana nuna jerin Manajan Fakitin Qualcomm.
c. Zaɓi Qualcomm Package Manager 3 kuma zazzage fakitin debian Linux. - Shigar da Qualcomm Package Manager 3 don Linux. Yi amfani da umarni mai zuwa:
$ dpkg -i – tilasta-sake rubutawa /hanyar/zuwa/
QualcommPackageManager 3.3.0.83.1.Linux-x86.deb - Zazzage Qualcomm®
AI Engine Direct SDK akan aikin Ubuntu.
a. Danna https://qpm.qualcomm.com/ kuma danna Kayan aiki.
b. A cikin sashin hagu, a cikin filin Kayan aikin Bincike, rubuta tarin AI. Daga cikin System OS list, zaɓi Linux.
A Jerin zaɓuka mai ɗauke da injunan tarin AI iri-iri an nuna.
c. Danna Qualcomm® AI Engine Direct SDK kuma zazzage fakitin Linux v2.14.0. - Shigar Qualcomm® AI Engine Direct SDK akan aikin Ubuntu.
a. Kunna lasisin:
qpm-cli -lasisi- kunna qualcomm_ai_engine_direct
b Shigar AI Engine Direct SDK:
$ qpm-cli – cire /hanya/to/ qualcomm_ai_engine_direct.2.14.0.230828.Linux-AnyCPU.qik - Tura dakunan karatu zuwa na'urar daga aikin Ubuntu tare da turawa adb.
$ cd /opt/qcom/aistack/qnn/2.14.0.230828 $ adb push ./lib/aarch64-oe-linux-gcc11.2/ libQnnDsp.so /usr/lib/ $ adb push ./lib/aarch64-oe- linux-gcc11.2/ libQnnDspV66Stub.so /usr/lib/ $ adb push ./lib/aarch64-oe-linux-gcc11.2/ libQnnGpu.so /usr/lib/ $ adb push ./lib/aarch64-oe- linux-gcc11.2/ libQnnHtpPrepare.so /usr/lib/ $ adb push ./lib/aarch64-oe-linux-gcc11.2/ libQnnHtp.so /usr/lib/ $ adb push ./lib/aarch64-oe- linux-gcc11.2/ libQnnHtpV68Stub.so /usr/lib/ $ adb push ./lib/aarch64-oe-linux-gcc11.2/ libQnnSaver.so /usr/lib/ $ adb push ./lib/aarch64-oe- linux-gcc11.2/ libQnnSystem.so /usr/lib/ $ adb push ./lib/aarch64-oe-linux-gcc11.2/ libQnnTFLiteDelegate.so /usr/lib/ $ adb push ./lib/hexagon-v65/ unsigned/ libQnnDspV65Skel.so /usr/lib/rfsa/adsp $ adb push ./lib/hexagon-v66/unsigned/ libQnnDspV66Skel.so /usr/lib/rfsa/adsp $ adb push ./lib/hexagon-v68/unsigned/ libQnnHtpV68Skel.so /usr/lib/rfsa/adsp $ adb push ./lib/hexagon-v69/unsigned/ libQnnHtpV69Skel.so /usr/lib/rfsa/adsp $ adb push ./lib/hexagon-v73/unsigned/tp. so /usr/lib/rfsa/adsp
Gwada Qualcomm TFlite SDK
Qualcomm TFLite SDK yana ba da takamaiman example aikace-aikacen, waɗanda za a iya amfani da su don ingantawa, ma'auni, da samun daidaiton samfuran da mai haɓaka ke son tantancewa.
Bayan an shigar da fakitin Qualcomm TFLite SDK akan na'urar, lokacin aiki yana samuwa akan na'urar don gudanar da waɗannan tsoffin.ampda aikace-aikace.
Abubuwan da ake bukata
Ƙirƙiri kundayen adireshi masu zuwa akan na'urar:
$ adb harsashi "mkdir / bayanai / Model"
$ adb harsashi "mkdir /data/Lables"
$ adb harsashi "mkdir /data/profiling"
Hoton lakabin
Hoton lakabin kayan aiki ne wanda Qualcomm TFLite SDK ke bayarwa wanda ke nuna yadda zaku iya loda ƙirar TensorFlow Lite da aka riga aka yi da ita kuma amfani da shi don gane abubuwa a cikin hotuna. Abubuwan da ake buƙata:
Zazzagewar sampmodel da kuma image:
Kuna iya amfani da kowane samfurin da ya dace, amma samfurin MobileNet v1 na gaba yana ba da kyakkyawar nuni na ƙirar da aka horar don gane abubuwa daban-daban 1000.
- Samu samfuri
$ curl https://ajiya.googleapis.com/download.tensorflow.org/models/ mobilenet_v1_2018_02_22/mobilenet_v1_1.0_224.tgz | tar xzv -C /data $ mv /data/mobilenet_v1_1.0_224.tflite /data/Models/ - Samu lakabi
$ curl https://ajiya.googleapis.com/download.tensorflow.org/models/ mobilenet_v1_1.0_224_frozen.tgz | tar xzv -C /data mobilenet_v1_1.0_224/ labels.txt
$ mv /data/mobilenet_v1_1.0_224/labels.txt /data/Labels/
Bayan kun haɗa zuwa kwandon docker na Qualcomm TFLite SDK, ana iya samun hoton a:
"/mnt/tflite/src/tensorflow/tensorflow/lite/examples/label_image/ testdata/grace_hopper.bmp”
a. Tura wannan file zuwa/data/Labels/
b. Gudanar da umarni:
$ adb harsashi "label_image -l /data/Labels/labels.txt -i /data/Labels/ grace_hopper.bmp -m /data/Models/mobilenet_v1_1.0_224.tflite -c 10 -j 1 -p 1"
Alamar alama
Qualcomm TFLite SDK yana ba da kayan aikin benchmarking don ƙididdige ayyukan lokutan gudu daban-daban.
Waɗannan kayan aikin ma'auni a halin yanzu suna aunawa da ƙididdige ƙididdiga don mahimman ma'aunin ayyuka masu zuwa:
- Lokacin farawa
- Lokacin ƙaddamarwa na yanayin zafi
- Lokacin ƙaddamarwa na tsayayyen yanayi
- Amfani da ƙwaƙwalwar ajiya yayin lokacin farawa
- Gabaɗaya amfani da ƙwaƙwalwar ajiya
Abubuwan da ake bukata
Tura samfuran da za a gwada daga TFLite Model Zoo (https://tfhub.dev/) zuwa/data/Model/. Gudu da rubutun masu zuwa:
- Kunshin XNN
$ adb harsashi "benchmark_model -graph = / bayanai / Models / - kunna_op_profiling = gaskiya -use_xnnpack = gaskiya -num_threads = 4 -max_secs = 300 -profiling_output_csv_file=/data/profiling/” - Wakilin GPU
$ adb harsashi "benchmark_model -graph = / bayanai / Models / - kunna_op_profiling = gaskiya -use_gpu = gaskiya -num_runs = 100 -warmup_runs = 10 - max_secs = 300 -profiling_output_csv_file=/data/profiling/” - Wakilin Waje
GPU na Wakilin Waje na QNN:
Gudanar da ƙima tare da ƙirar ƙira mai iyo:
$ adb shell-umurni “benchmark_model –graph=/data/Models/ .tflite –external_delegate_path=libQnnTFLiteDelegate.so — external_delegate_options='backend_type:gpu;library_path:/usr/lib/ libQnnGpu.so/brarrf:/uskel /adsp'"
Wakilin Waje na QNN HTP:
Shigar da ƙididdiga tare da ƙirar ƙira:
$ adb shell-umurni “benchmark_model –graph=/data/Models/ .tflite –external_delegate_path=libQnnTFLiteDelegate.so — external_delegate_options='backend_type:htp;library_path:/usr/lib/ libQnnHtp.so/ libQnnHtp.so/slib /adsp'"
Daidaitaccen kayan aiki
Qualcomm TFLite SDK yana ba da ingantaccen kayan aiki don ƙididdige daidaiton ƙira tare da lokutan gudu daban-daban.
- Rarraba tare da wakilan GPU
Matakan zazzage abin da ake bukata fileAna iya samun s don gwadawa a: "/mnt/tflite/src/tensorflow/tensorflow/lite/tools/evaluation/tasks/ imagenet_image_classificatio/README.md"
Binary don gudanar da wannan kayan aikin ya riga ya zama ɓangare na SDK, don haka mai haɓaka baya buƙatar sake gina shi.
$ adb harsashi "image_classify_run_eval - model_file=/data/Model/ -ground_truth_images_path =/data/ - ground_truth_labels=/data/ –model_output_labels=/ data/ –delegate=gpu” - Gano abu tare da fakitin XNN
$ adb harsashi “inf_diff_run_eval –model_file=/data/Model/ –delegate=xnnpac
BAYANIN SHARI'A
Samun damar ku zuwa da amfani da wannan takaddar, tare da kowane takamaiman bayani, allon tunani files, zane-zane, bincike-bincike da sauran bayanan da ke ƙunshe a ciki (gaɗin wannan “Takardun”), yana ƙarƙashin naku (ciki har da kamfani ko wani mahaluƙi na doka da kuke wakilta, tare "Kai" ko "Naku") yarda da sharuɗɗan ("Sharuɗɗan Amfani") saita a kasa. Idan ba ku yarda da waɗannan Sharuɗɗan Amfani ba, ba za ku iya amfani da wannan Takardun ba kuma za ku lalata kowane kwafinsa nan da nan.
- Sanarwa na Shari'a.
Ana ba da wannan Takaddun zuwa gare ku kawai don amfanin cikin ku tare da waɗannan samfuran da hadayun sabis na Qualcomm Technologies, Inc. ("Qualcomm Technologies") da masu haɗin gwiwa da aka kwatanta a cikin wannan Takardun, kuma ba za a yi amfani da su don kowane dalilai ba. Ba za a iya canza wannan Takaddun ba, gyara, ko gyara ta kowace hanya ba tare da amincewar rubuce-rubucen Qualcomm Technologies ba. Amfani ko bayyana wannan ba da izini ba
Takaddun bayanai ko bayanan da ke ƙunshe a ciki an haramta su sosai, kuma Kun yarda da ba da Lamuni na Qualcomm Technologies, abokan haɗin gwiwa da masu lasisi don duk wani lahani ko asarar da Qualcomm Technologies suka fuskanta, masu alaƙa da masu ba da lasisi ga kowane irin amfani ko bayyana wannan takaddar ba tare da izini ba, gabaɗaya ko bangare. Qualcomm Technologies, masu haɗin gwiwa da masu lasisi suna riƙe duk haƙƙoƙi da mallaka a ciki da wannan Takardun. Babu lasisi ga kowane alamar kasuwanci, haƙƙin mallaka, haƙƙin mallaka, haƙƙin kare aikin rufe fuska ko kowane haƙƙin mallakar fasaha da aka bayar ko aka ba da shi ta wannan Takaddun ko duk wani bayanin da aka bayyana a ciki, gami da, amma ba'a iyakance shi ba, kowane lasisi don yin, amfani, shigo da ko sayar da kowane samfur, sabis ko fasaha wanda ke ba da duk wani bayanin da ke cikin wannan Takardun.
ANA BARKA DA WANNAN TAKARDUN “KAMAR YADDA” BA TARE DA WARRANTI KOWANE IRIN BA, KO BAYANI, BAYANI, DOKA KO SAURAN. ZUWA MATSALAR DOKA, FASSARAR KWAKWALWA, ALAMOMINSA DA MASU LASANCE NA MUSAMMAN DUK GARANTIN LAKARYA, KYAUTATAWA, RASHIN CI GABA, GAMSAR MATSALOLIN CY, DA DUKAN GARANTIN DA SUKE FARUWA DAGA AMFANIN CINIKI KO FITAR DA SHA'AWAR MU'AMALA KO SHA'AWAR AIKI. KARIN BAYANI, BABU FASSARAR QUALCOMM, KO WATA DAGA CIKIN ALAMOMINSA KO MASU LASANCE, BA ZAI IYA HANNU A GAREKU KO WATA JAM'IYYA NA UKU DON WANI KUDI, Asara, AMFANI, KO AIYUKA DUK HANYAR SANADIYYAR KU KO SANARWA.
Wasu na'urorin samfur, kayan aiki da kayan da aka ambata a cikin wannan Takardun na iya buƙatar ka karɓi ƙarin sharuɗɗa da sharuɗɗa kafin samun dama ko amfani da waɗannan abubuwan.
Bayanan fasaha da aka ƙayyade a cikin wannan Takardun na iya kasancewa ƙarƙashin Amurka da sauran dokokin sarrafa fitarwar da suka dace. An hana watsawa sabanin Amurka da kowace doka da ta dace.
Babu wani abu a cikin wannan Takardun da aka bayar don siyar da kowane kayan haɗin gwiwa ko na'urorin da aka ambata a nan.
Wannan Takardun na iya canzawa ba tare da ƙarin sanarwa ba. A yayin da aka sami sabani tsakanin waɗannan Sharuɗɗan Amfani da WebSharuɗɗan Amfani a kunne www.qualcomm.com ko Ka'idodin Sirri na Qualcomm da aka ambata www.qualcomm.com, waɗannan Sharuɗɗan Amfani za su sarrafa. A yayin da ake samun rikici tsakanin waɗannan Sharuɗɗan Amfani da duk wata yarjejeniya (rubuta ko danna-ta) waɗanda Kai da Qualcomm Technologies suka aiwatar ko haɗin gwiwar Qualcomm Technologies dangane da damar ku da amfani da wannan Takardun, ɗayan yarjejeniyar za ta sarrafa. .
Waɗannan Sharuɗɗan Amfani za a sarrafa su kuma a yi amfani da su da aiwatar da su daidai da dokokin Jihar California, ban da Yarjejeniyar Majalisar Dinkin Duniya kan Siyar da Kaya ta Duniya, ba tare da la'akari da rikice-rikice na ƙa'idodin doka ba. Duk wata gardama, da'awa ko gardama da ta taso daga ko ta shafi waɗannan Sharuɗɗan Amfani, ko keta ko ingancin wannan, wata kotun da ke da ikon yanke hukunci za ta yanke hukunci ne kawai a gundumar San Diego, Jihar California, kuma kun yarda da haka. hukunce-hukuncen irin wadannan kotuna don haka. - Alamar kasuwanci da Bayanin Haɗin Samfur.
Qualcomm alamar kasuwanci ce ko alamar kasuwanci mai rijista ta Qualcomm Incorporated. Arm alamar kasuwanci ce mai rijista ta Arm Limited (ko rassan sa) a cikin Amurka da/ko wani wuri. Alamar kalmar Bluetooth® alamar kasuwanci ce mai rijista ta Bluetooth SIG, Inc. Sauran samfura da sunayen iri da aka ambata a cikin wannan Takardun na iya zama alamun kasuwanci ko alamun kasuwanci masu rijista na masu su.
Samfuran samfuran Snapdragon da Qualcomm da aka ambata a cikin wannan Takaddun samfuran samfuran Qualcomm Technologies, Inc. da/ko rassan sa ne. Qualcomm Incorporated ta ba da lasisin fasaha na Qualcomm.
Takardu / Albarkatu
 |
Qualcomm TensorFlow Lite SDK Software [pdf] Jagorar mai amfani TensorFlow Lite SDK Software, Lite SDK Software, SDK Software, Software |
