ML Ops Study 2 - Speaker Deck

Slide 1

Slide 1 text

C-LIS CO., LTD.

Slide 2

Slide 2 text

.-0QT4UVEZ BUΫοΫύουגࣜձࣾ 5FOTPS'MPXͷ܇࿅ࡁΈϞσϧΛ  "OESPJEΞϓϦʹࡌͤΔͱ͖ʹ  ۤ࿑ͨ͠࿩ 5FOTPS'MPXͰझຯͷը૾ऩूαʔόʔΛ࡞Δ೥݄߸

Slide 3

Slide 3 text

C-LIS CO., LTD. ༗ࢁܓೋʢ,FJKJ"3*:"."ʣ $-*4$0 -5% Photo : Koji MORIGUCHI (MORIGCHOWDER) "OESPJEΞϓϦ։ൃνϣοτσΩϧ ػցֶश͸ͪΐͬͱ΍ͬͨ͜ͱ͋Γ·͢

Slide 4

Slide 4 text

C-LIS CO., LTD.

Slide 5

Slide 5 text

C-LIS CO., LTD. ΍ͬ΂

Slide 6

Slide 6 text

ຊ൪؀ڥ΁ͷ౤ೖܦݧ͕ඞཁͳΒ ౤ೖͯ͠͠·͑͹͍͍͡Όͳ͍

Slide 7

Slide 7 text

؟ ڸ ͬ ່ ࠜઇΕ͍

Slide 8

Slide 8 text

޷Έͷ؟ڸ່ͬը૾Λ  ࣗಈͰऩू͍ͨ͠

Slide 9

Slide 9 text

̎Ϋϥε෼ྨ 1 0

Slide 10

Slide 10 text

ݱࡏͷγεςϜ

Slide 11

Slide 11 text

ධՁ༻ αʔόʔ ܇࿅ɾֶश༻αʔόʔ σʔληοτసૹ ʢTFRecordʣ ֶशࡁΈ ύϥϝʔλʔऔಘ ը૾औಘ ը૾औಘ ϥϕϧ ෇͚ σʔληοτ؅ཧ αʔόʔ σʔλऩूݩ αʔϏε ը૾औಘ ϥϕϧ ෇͚ Android  ΞϓϦ

Slide 12

Slide 12 text

σʔληοτ؅ཧαʔόʔ $16"UISPO/FP()[ .FNPSZ(# 4UPSBHF44%(#  )%%5# 3"*%

Slide 13

Slide 13 text

σʔληοτ؅ཧαʔόʔͷ໾ׂ ը૾σʔλͷऩू ϝλσʔλʢΞϊςʔγϣϯɾϥϕϧʣͷ؅ཧ "1*ͷఏڙ ֶश༻σʔλʢ5'3FDPSEʣͷੜ੒

Slide 14

Slide 14 text

ը૾σʔλͷऩू

Slide 15

Slide 15 text

ϝλσʔλͷ؅ཧ

Slide 16

Slide 16 text

ϝλσʔλͷ؅ཧ label: 2 left: 283 top: 190 right: 435 bottom:301 = 1.0

Slide 17

Slide 17 text

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

Slide 18

Slide 18 text

"1*ͷఏڙ ը૾Ϧετͷऔಘ ը૾ͷऔಘ ը૾ͷݕࡧʢϥϕϧʣ ϥϕϧͷઃఆ ϥϕϧະઃఆը૾ͷϦετΛऔಘ σʔλऔಘݩ5XJUUFS*%ͷ௥Ճɾ࡟আ

Slide 19

Slide 19 text

ֶश༻σʔλͷੜ੒ 5'3FDPSEܗࣜPS+1&($47ܗࣜ ը૾ͷϦαΠζ͸͜ͷஈ֊Ͱߦ͏ʢτϥϑΟοΫΛ௿ݮ $ python ./create_dataset.py \ --base_dir /dataset/source/ \ --output_dir ~/tfrecords_classifier \ --image_size 256 \ --tag_names megane,nsfw,favorite,illust

Slide 20

Slide 20 text

σʔλͷάϧʔϓԽ 0 1 2 3 4 5 6 7 8 9 σʔληοτ ςετσʔληοτ

Slide 21

Slide 21 text

Slide 22

Slide 22 text

ֶश༻αʔόʔ

Slide 23

Slide 23 text

{ "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

Slide 24

Slide 24 text

ֶ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/

Slide 25

Slide 25 text

Slide 26

Slide 26 text

Slide 27

Slide 27 text

ධՁ༻αʔόʔ

Slide 28

Slide 28 text

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

Slide 29

Slide 29 text

Slide 30

Slide 30 text

"OESPJEΞϓϦ

Slide 31

Slide 31 text

Slide 32

Slide 32 text

Slide 33

Slide 33 text

Slide 34

Slide 34 text

Slide 35

Slide 35 text

௚ۙͷ՝୊

Slide 36

Slide 36 text

ෆద੾ը૾͕ଟ͗͢Δ

Slide 37

Slide 37 text

ෆద੾ʢ/4'8ʣը૾ͱ؟ڸը૾ /4'8 positive: 36,083 → 7.17%  negative: 466,738 ؟ڸ positive: 23,559 → 2.44%  negative: 938,563

Slide 38

Slide 38 text

ෆద੾ը૾ϑΟϧλʔΛ ΞϓϦʹ૊ΈࠐΉ

Slide 39

Slide 39 text

NPEFM NSFW positive: 5,628 negative: 17,253

Slide 40

Slide 40 text

NPEFM ֶशࡁΈύϥϝʔλʔϑΝΠϧ ֶशࡁΈϞσϧ .pb 170MB

Slide 41

Slide 41 text

NPEFM@MJUF ֶशࡁΈύϥϝʔλʔϑΝΠϧ ֶशࡁΈϞσϧ .pb 10.7MB

Slide 42

Slide 42 text

NPEFM@MJUF

Slide 43

Slide 43 text

Ϟσϧͷߏ଄ .pb input result 128x128x3 1

Slide 44

Slide 44 text

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

Slide 45

Slide 45 text

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

Slide 46

Slide 46 text

Slide 47

Slide 47 text

Ϟσϧͷߏ଄ .pb input result 128x128x4 1

Slide 48

Slide 48 text

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

Slide 49

Slide 49 text

σϞ

Slide 50

Slide 50 text

'PPE(BMMFSZ https://github.com/keiji/food_gallery_with_tensorﬂow ΪϟϥϦʔʹอଘ͞Ε͍ͯΔ৯෺ͷը૾Λදࣔ http://techlife.cookpad.com/entry/2017/09/14/161756 ΫοΫύου։ൃऀϒϩά ྉཧ͖Ζ͘ʹ͓͚Δྉཧʗඇྉཧ൑ผϞσϧͷৄࡉ

Slide 51

Slide 51 text

૊ΈࠐΜͰΈ͚ͨΕͲɺ ਫ਼౓͸͋·Γߴ͘ͳ͍ʜʜ

Slide 52

Slide 52 text

ࠓޙͷ՝୊ 5FOTPS'MPX-JUF΁ͷҠߦ ߴਫ਼౓ͷϞσϧͷѹॖʢল༰ྔԽʣ 1SVOJOH 2VBOUJ[BUJPO %JTUJMMBUJPO ML Kit: Machine Learning SDK for mobile developers (Google I/O '18) https://youtu.be/Z-dqGRSsaBs?t=32m10s

Slide 53

Slide 53 text

ࠓޙͷ՝୊ ֶशαΠΫϧΛࣗಈԽ͍͖͍ͯͨ͠ɻ ʢఆظతʹֶशσʔλͷੜ੒ͱసૹΛߦ͍ɺ࠶ֶश͢ΔͳͲʣ σʔλϕʔεͷߴ଎Խ  ɹϥϕϧݕࡧ͕ඇৗʹ஗͍ͷ͕՝୊ɻઃܭΛݟ௚͢ඞཁ͋Γ  ɹݕࡧΠϯσοΫεͷ(PPHMF$MPVE4UPSFҠߦΛݕ౼ 

Slide 54

Slide 54 text

ࠓޙͷ՝୊ σʔληοτ؅ཧαʔόʔͷߋ৽ 5FOTPS'MPXΑΓɺ$16ʹ*OUFM"79͕ඞਢʹͳͬͨʢQJQ൛ʣ

Slide 55

Slide 55 text

C-LIS CO., LTD. ຊࢿྉ͸ɺ༗ݶձࣾγʔϦεͷஶ࡞෺Ͱ͢ɻຊࢿྉͷશ෦ɺ·ͨ͸Ұ෦ʹ͍ͭͯɺஶ࡞ऀ͔ΒจॻʹΑΔڐ୚Λಘͣʹෳ੡͢Δ͜ͱ͸ې͡ΒΕ͍ͯ·͢ɻ 5IF"OESPJE4UVEJPJDPOJTSFQSPEVDFEPSNPEJpFEGSPNXPSLDSFBUFEBOETIBSFECZ(PPHMFBOEVTFEBDDPSEJOHUPUFSNTEFTDSJCFEJOUIF$SFBUJWF$PNNPOT"UUSJCVUJPO-JDFOTF ֤੡඼໊ɾϒϥϯυ໊ɺձ໊ࣾͳͲ͸ɺҰൠʹ֤ࣾͷ঎ඪ·ͨ͸ొ࿥঎ඪͰ͢ɻຊࢿྉதͰ͸ɺɺɺäΛׂѪ͍ͯ͠·͢ɻ 5IF"OESPJESPCPUJTSFQSPEVDFEPSNPEJpFEGSPNXPSLDSFBUFEBOETIBSFECZ(PPHMFBOEVTFEBDDPSEJOHUPUFSNTEFTDSJCFEJOUIF$SFBUJWF$PNNPOT"UUSJCVUJPO-JDFOTF