Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
ML Ops Study 2
Search
ARIYAMA Keiji
May 29, 2018
Technology
0
130
ML Ops Study 2
5月29日にクックパッド株式会社で開催されたML Ops Study #2の発表資料です。
ARIYAMA Keiji
May 29, 2018
Tweet
Share
More Decks by ARIYAMA Keiji
See All by ARIYAMA Keiji
Build with AI
keiji
0
140
DroidKaigi 2023
keiji
0
1.5k
TechFeed Conference 2022
keiji
0
220
Android Bazaar and Conference Diverse 2021 Winter
keiji
0
830
ci-cd-conference-2021
keiji
1
1.2k
Android Bazaar and Conference 2021 Spring
keiji
3
750
TFUG KANSAI 20190928
keiji
0
92
Softpia Japan Seminar 20190724
keiji
1
150
pixiv App Night 20190611
keiji
1
560
Other Decks in Technology
See All in Technology
AGIについてChatGPTに聞いてみた
blueb
0
130
iOSチームとAndroidチームでブランチ運用が違ったので整理してます
sansantech
PRO
0
130
rootlessコンテナのすゝめ - 研究室サーバーでもできる安全なコンテナ管理
kitsuya0828
3
380
OCI 運用監視サービス 概要
oracle4engineer
PRO
0
4.8k
DMARC 対応の話 - MIXI CTO オフィスアワー #04
bbqallstars
1
160
フルカイテン株式会社 採用資料
fullkaiten
0
40k
テストコード品質を高めるためにMutation Testingライブラリ・Strykerを実戦導入してみた話
ysknsid25
7
2.6k
IBC 2024 動画技術関連レポート / IBC 2024 Report
cyberagentdevelopers
PRO
0
110
100 名超が参加した日経グループ横断の競技型 AWS 学習イベント「Nikkei Group AWS GameDay」の紹介/mediajaws202411
nikkei_engineer_recruiting
1
170
元旅行会社の情シス部員が教えるおすすめなre:Inventへの行き方 / What is the most efficient way to re:Invent
naospon
2
340
Terraform未経験の御様に対してどの ように導⼊を進めていったか
tkikuchi
2
430
Amplify Gen2 Deep Dive / バックエンドの型をいかにしてフロントエンドへ伝えるか #TSKaigi #TSKaigiKansai #AWSAmplifyJP
tacck
PRO
0
370
Featured
See All Featured
Java REST API Framework Comparison - PWX 2021
mraible
PRO
28
8.2k
How STYLIGHT went responsive
nonsquared
95
5.2k
ReactJS: Keep Simple. Everything can be a component!
pedronauck
665
120k
Save Time (by Creating Custom Rails Generators)
garrettdimon
PRO
27
840
Large-scale JavaScript Application Architecture
addyosmani
510
110k
Optimizing for Happiness
mojombo
376
70k
Music & Morning Musume
bryan
46
6.2k
Become a Pro
speakerdeck
PRO
25
5k
Rails Girls Zürich Keynote
gr2m
94
13k
Build The Right Thing And Hit Your Dates
maggiecrowley
33
2.4k
Optimising Largest Contentful Paint
csswizardry
33
2.9k
Design and Strategy: How to Deal with People Who Don’t "Get" Design
morganepeng
126
18k
Transcript
C-LIS CO., LTD.
.-0QT4UVEZ BUΫοΫύουגࣜձࣾ 5FOTPS'MPXͷ܇࿅ࡁΈϞσϧΛ "OESPJEΞϓϦʹࡌͤΔͱ͖ʹ ۤ࿑ͨ͠ 5FOTPS'MPXͰझຯͷը૾ऩूαʔόʔΛ࡞Δ݄߸
C-LIS CO., LTD. ༗ࢁܓೋʢ,FJKJ"3*:"."ʣ $-*4$0 -5% Photo :
Koji MORIGUCHI (MORIGCHOWDER) "OESPJEΞϓϦ։ൃνϣοτσΩϧ ػցֶशͪΐͬͱͬͨ͜ͱ͋Γ·͢
C-LIS CO., LTD.
C-LIS CO., LTD. ͬ
ຊ൪ڥͷೖܦݧ͕ඞཁͳΒ ೖͯ͠͠·͍͍͑͡Όͳ͍
؟ ڸ ͬ ່ ࠜઇΕ͍
Έͷ؟ڸ່ͬը૾Λ ࣗಈͰऩू͍ͨ͠
̎Ϋϥεྨ 1 0
ݱࡏͷγεςϜ
ධՁ༻ αʔόʔ ܇࿅ɾֶश༻αʔόʔ σʔληοτసૹ ʢTFRecordʣ ֶशࡁΈ ύϥϝʔλʔऔಘ ը૾औಘ
ը૾औಘ ϥϕϧ ͚ σʔληοτཧ αʔόʔ σʔλऩूݩ αʔϏε ը૾औಘ ϥϕϧ ͚ Android ΞϓϦ
σʔληοτཧαʔόʔ $16"UISPO/FP()[ .FNPSZ(# 4UPSBHF44%(# )%%5# 3"*%
σʔληοτཧαʔόʔͷׂ ը૾σʔλͷऩू ϝλσʔλʢΞϊςʔγϣϯɾϥϕϧʣͷཧ "1*ͷఏڙ ֶश༻σʔλʢ5'3FDPSEʣͷੜ
ը૾σʔλͷऩू
ϝλσʔλͷཧ
ϝλσʔλͷཧ label: 2 left: 283 top: 190 right:
435 bottom:301 = 1.0
ϥϕϧͷछྨ PSJHJOBM@BSU OTGX GBWPSJUF QIPUP JMMVTU DPNJD GBDF
GFNBMF NFHBOF TDISPPM@VOJGPSN CMB[FS@VOJGPSN TBJMPS@VOJGPSN HM LFNPOP NBMF CM DBU EPH GPPE EJTMJLF
"1*ͷఏڙ ը૾Ϧετͷऔಘ ը૾ͷऔಘ ը૾ͷݕࡧʢϥϕϧʣ ϥϕϧͷઃఆ ϥϕϧະઃఆը૾ͷϦετΛऔಘ σʔλऔಘݩ5XJUUFS*%ͷՃɾআ
ֶश༻σʔλͷੜ 5'3FDPSEܗࣜPS+1&( $47ܗࣜ ը૾ͷϦαΠζ͜ͷஈ֊Ͱߦ͏ʢτϥϑΟοΫΛݮ $ python ./create_dataset.py \
--base_dir /dataset/source/ \ --output_dir ~/tfrecords_classifier \ --image_size 256 \ --tag_names megane,nsfw,favorite,illust
σʔλͷάϧʔϓԽ 0 1 2 3 4 5 6
7 8 9 σʔληοτ ςετσʔληοτ
ධՁ༻ αʔόʔ ܇࿅ɾֶश༻αʔόʔ σʔληοτసૹ ʢTFRecordʣ ֶशࡁΈ ύϥϝʔλʔऔಘ ը૾औಘ
ը૾औಘ ϥϕϧ ͚ σʔληοτཧ αʔόʔ σʔλऩूݩ αʔϏε ը૾औಘ ϥϕϧ ͚ Android ΞϓϦ
ֶश༻αʔόʔ
{ "tag_name": "megane", "train_catalog_numbers": "0,1,2,3,4,5,6,7,8", "eval_catalog_numbers": "9", "data_augmentation": { "random_crop":
false, "random_colorize": true } } ֶश༻ͷઃఆϑΝΠϧ
ֶशͱݕূͷ࣮ߦ $ CUDA_VISIBLE_DEVICES=0,1 python ./train.py \ --learning_config config_megane.json
\ --tfrecords_dir ~/tfrecords_classifier \ --train_dir ~/train_single_discriminator \ --summary_dir ~/summary_single_discriminator \ --batch_size 64 \ --learning_rate 0.0001 \ --num_gpus 2 \ --max_step 100000 $ CUDA_VISIBLE_DEVICES=2 python ./eval.py \ --learning_config config_megane.json \ --tfrecords_dir ~/tfrecords_classifier \ --train_dir ~/train_single_discriminator \ --summary_dir ~/summary_single_discriminator $ tensorboard \ --logdir ~/summary_single_discriminator/megane/
ධՁ༻ αʔόʔ ܇࿅ɾֶश༻αʔόʔ σʔληοτసૹ ʢTFRecordʣ ֶशࡁΈ ύϥϝʔλʔऔಘ ը૾औಘ
ը૾औಘ ϥϕϧ ͚ σʔληοτཧ αʔόʔ σʔλऩूݩ αʔϏε ը૾औಘ ϥϕϧ ͚ Android ΞϓϦ
ධՁ༻αʔόʔ
ධՁͷ࣮ߦ $ python3 client.py \ --tag_name megane \
--train_base_path ~/train_single_discriminator \ --train_file_name precision-0.956463/megane.ckpt-294000 \ --batch_size 16 \ --limit_batch 100
ධՁ༻ αʔόʔ ܇࿅ɾֶश༻αʔόʔ σʔληοτసૹ ʢTFRecordʣ ֶशࡁΈ ύϥϝʔλʔऔಘ ը૾औಘ
ը૾औಘ ϥϕϧ ͚ σʔληοτཧ αʔόʔ σʔλऩूݩ αʔϏε ը૾औಘ ϥϕϧ ͚ Android ΞϓϦ
"OESPJEΞϓϦ
ۙͷ՝
ෆదը૾͕ଟ͗͢Δ
ෆదʢ/4'8ʣը૾ͱ؟ڸը૾ /4'8 positive: 36,083 → 7.17% negative:
466,738 ؟ڸ positive: 23,559 → 2.44% negative: 938,563
ෆదը૾ϑΟϧλʔΛ ΞϓϦʹΈࠐΉ
NPEFM NSFW positive: 5,628 negative: 17,253
NPEFM ֶशࡁΈύϥϝʔλʔϑΝΠϧ ֶशࡁΈϞσϧ .pb 170MB
NPEFM@MJUF ֶशࡁΈύϥϝʔλʔϑΝΠϧ ֶशࡁΈϞσϧ .pb 10.7MB
NPEFM@MJUF
Ϟσϧͷߏ .pb input result 128x128x3 1
private val IMAGE_WIDTH = 128 private val IMAGE_HEIGHT = 128
private val IMAGE_CHANNEL = 3 private val IMAGE_BYTES_LENGTH = IMAGE_WIDTH * IMAGE_HEIGHT * IMAGE_CHANNEL val imageByteBuffer: ByteBuffer = ByteBuffer.allocate(IMAGE_BYTES_LENGTH) val scaledBitmap = Bitmap.createScaledBitmap(bitmap, IMAGE_WIDTH, IMAGE_HEIGHT, false) scaledBitmap.copyPixelsToBuffer(imageByteBuffer) ը૾ΛόοϑΝʹ֨ೲ
val resultArray = FloatArray(1) fun recognize(imageByteArray: ByteArray): Float { val
start = Debug.threadCpuTimeNanos() tfInference.feed("input", imageByteArray, imageByteArray.size.toLong()) tfInference.run(arrayOf("result")) tfInference.fetch("result", resultArray) val elapsed = Debug.threadCpuTimeNanos() - start Log.d(TAG, "Elapsed: %d ns".format(elapsed)) return resultArray[0] } GFFESVOGFUDI
private val IMAGE_WIDTH = 128 private val IMAGE_HEIGHT = 128
private val IMAGE_CHANNEL = 3 private val IMAGE_BYTES_LENGTH = IMAGE_WIDTH * IMAGE_HEIGHT * IMAGE_CHANNEL val imageByteBuffer: ByteBuffer = ByteBuffer.allocate(IMAGE_BYTES_LENGTH) val scaledBitmap = Bitmap.createScaledBitmap(bitmap, IMAGE_WIDTH, IMAGE_HEIGHT, false) scaledBitmap.copyPixelsToBuffer(imageByteBuffer) scaledBitmap.recycle() ݪҼΒ͖͠ͷ ˢ"MQIBνϟϯωϧ͕ೖ͍ͬͯΔ
Ϟσϧͷߏ .pb input result 128x128x4 1
with tf.Graph().as_default() as g: image_ph = tf.placeholder( tf.uint8, [model.IMAGE_SIZE *
model.IMAGE_SIZE * 4], name='input') image = tf.cast(image_ph, tf.float32) image = tf.reshape( image, [model.IMAGE_SIZE, model.IMAGE_SIZE, 4]) image = image[:, :, :3] QCग़ྗ࣌ʹDIΛड͚ೖΕΔΑ͏ʹάϥϑΛมߋ
σϞ
'PPE(BMMFSZ https://github.com/keiji/food_gallery_with_tensorflow ΪϟϥϦʔʹอଘ͞Ε͍ͯΔ৯ͷը૾Λදࣔ http://techlife.cookpad.com/entry/2017/09/14/161756 ΫοΫύου։ൃऀϒϩά ྉཧ͖Ζ͘ʹ͓͚ΔྉཧʗඇྉཧผϞσϧͷৄࡉ
ΈࠐΜͰΈ͚ͨΕͲɺ ਫ਼͋·Γߴ͘ͳ͍ʜʜ
ࠓޙͷ՝ 5FOTPS'MPX-JUFͷҠߦ ߴਫ਼ͷϞσϧͷѹॖʢল༰ྔԽʣ 1SVOJOH 2VBOUJ[BUJPO %JTUJMMBUJPO ML Kit:
Machine Learning SDK for mobile developers (Google I/O '18) https://youtu.be/Z-dqGRSsaBs?t=32m10s
ࠓޙͷ՝ ֶशαΠΫϧΛࣗಈԽ͍͖͍ͯͨ͠ɻ ʢఆظతʹֶशσʔλͷੜͱసૹΛߦ͍ɺ࠶ֶश͢ΔͳͲʣ σʔλϕʔεͷߴԽ ɹϥϕϧݕࡧ͕ඇৗʹ͍ͷ͕՝ɻઃܭΛݟ͢ඞཁ͋Γ ɹݕࡧΠϯσοΫεͷ(PPHMF$MPVE4UPSFҠߦΛݕ౼
ࠓޙͷ՝ σʔληοτཧαʔόʔͷߋ৽ 5FOTPS'MPXΑΓɺ$16ʹ*OUFM"79͕ඞਢʹͳͬͨʢQJQ൛ʣ
C-LIS CO., LTD. ຊࢿྉɺ༗ݶձࣾγʔϦεͷஶ࡞Ͱ͢ɻຊࢿྉͷશ෦ɺ·ͨҰ෦ʹ͍ͭͯɺஶ࡞ऀ͔ΒจॻʹΑΔڐΛಘͣʹෳ͢Δ͜ͱې͡ΒΕ͍ͯ·͢ɻ 5IF"OESPJE4UVEJPJDPOJTSFQSPEVDFEPSNPEJpFEGSPNXPSLDSFBUFEBOETIBSFECZ(PPHMFBOEVTFEBDDPSEJOHUPUFSNTEFTDSJCFEJOUIF$SFBUJWF$PNNPOT"UUSJCVUJPO-JDFOTF ໊֤ɾϒϥϯυ໊ɺձ໊ࣾͳͲɺҰൠʹ֤ࣾͷඪ·ͨొඪͰ͢ɻຊࢿྉதͰɺɺɺäΛׂѪ͍ͯ͠·͢ɻ 5IF"OESPJESPCPUJTSFQSPEVDFEPSNPEJpFEGSPNXPSLDSFBUFEBOETIBSFECZ(PPHMFBOEVTFEBDDPSEJOHUPUFSNTEFTDSJCFEJOUIF$SFBUJWF$PNNPOT"UUSJCVUJPO-JDFOTF