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
20171209 Sakura ML Night
Search
ARIYAMA Keiji
December 09, 2017
Technology
0
140
20171209 Sakura ML Night
2017年12月9日に大阪で開催された「さくらの機械学習ナイト」の発表資料です。
「TensorFlowによるNSFW(職場で不適切な)画像検出」について。
ARIYAMA Keiji
December 09, 2017
Tweet
Share
More Decks by ARIYAMA Keiji
See All by ARIYAMA Keiji
Build with AI
keiji
0
120
DroidKaigi 2023
keiji
0
1.4k
TechFeed Conference 2022
keiji
0
200
Android Bazaar and Conference Diverse 2021 Winter
keiji
0
820
ci-cd-conference-2021
keiji
1
1.2k
Android Bazaar and Conference 2021 Spring
keiji
3
740
TFUG KANSAI 20190928
keiji
0
88
Softpia Japan Seminar 20190724
keiji
1
140
pixiv App Night 20190611
keiji
1
540
Other Decks in Technology
See All in Technology
不動産 x AIことはじめ~データの真価を拓くために
estie
0
110
Oracle Autonomous Database:サービス概要のご紹介
oracle4engineer
PRO
1
7k
Mocking in Rust Applications
taiki45
1
410
プログラム検証入門
riru
6
860
チームビルディングは"感性"で向き合おう / Team Building with Awareness
kohzas
0
210
Javaにおける関数型プログラミンへの取り組み
skrb
7
320
突撃! 隣のAmazon Bedrockユーザー 〜YouはどうしてAWSで?〜
minorun365
PRO
3
380
OCI で始める!! Red Hat OpenShift / Get Started OpenShift on OCI
oracle4engineer
PRO
1
170
自社サービスのための独自リリース版Redmine「RedMica」の取り組み
vividtone
0
1.3k
技術的負債解消の取り組みと専門チームのお話
bengo4com
0
330
Creative UIs with Compose: DroidKaigi 2024
chrishorner
1
480
Privacy Sandbox on Android / DroidKaigi 2024
7pairs
1
240
Featured
See All Featured
Principles of Awesome APIs and How to Build Them.
keavy
125
16k
Git: the NoSQL Database
bkeepers
PRO
425
64k
The Pragmatic Product Professional
lauravandoore
31
6.2k
Build your cross-platform service in a week with App Engine
jlugia
228
18k
Adopting Sorbet at Scale
ufuk
73
8.9k
Fashionably flexible responsive web design (full day workshop)
malarkey
401
65k
jQuery: Nuts, Bolts and Bling
dougneiner
61
7.4k
Writing Fast Ruby
sferik
623
60k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
26
1.3k
Clear Off the Table
cherdarchuk
91
320k
I Don’t Have Time: Getting Over the Fear to Launch Your Podcast
jcasabona
26
1.9k
ReactJS: Keep Simple. Everything can be a component!
pedronauck
663
120k
Transcript
C-LIS CO., LTD.
C-LIS CO., LTD. ༗ࢁܓೋʢ,FJKJ"3*:"."ʣ $-*4$0 -5% "OESPJEΞϓϦ։ൃνϣοτσΩϧ Photo by
Koji MORIGUCHI (MORIGCHOWDER) ػցֶशͪΐͬͱͬͨ͜ͱ͋Γ·͢ Twitterͬͯ·ͤΜ
͘͞ΒͷػցֶशφΠτ 5FOTPS'MPXͰ /4'8ը૾ݕग़
5FOTPS'MPXʢ݄ൃදʣ ػցೳ͚ܭࢉϑϨʔϜϫʔΫ ࠷৽όʔδϣϯʢ݄ʣ
ษڧձΖ͏ͥ
(PPHMF%FWFMPQFS(SPVQ
IUUQTHEHLPCFEPPSLFFQFSKQFWFOUT
Πϯλʔωοτ͔Β Έͷը૾ΛࣗಈͰऩू͍ͨ͠
© ࠜઇΕ͍ ؟ ڸ ͬ ່
؟ڸ່ͬఆ 1 0
σʔληοτʢ݄࣌ʣ ؟ڸ່ͬɹຕ ඇ؟ڸ່ͬຕ ؟ڸ່ͬ ඇ؟ڸ່ͬ ޡݕग़ ؟ڸ່ͬ ඇ؟ڸ່ͬ
{ "generator": "Region Cropper", "file_name": "haruki_g17.png", "regions": [ { "probability":
1.0, "label": 2, "rect": { "left": 97.0, "top": 251.0, "right": 285.0, "bottom": 383.0 } }, { "probability": 1.0, "label": 2, "rect": { "left": 536.0, "top": 175.0, "right": 730.0, "bottom": 321.0 } } ] } Region Cropper: https://github.com/keiji/region_cropper
ߏ Downloader σʔληοτ Region + Label ઃఆ rsync
ཧͷߏ Downloader Face Detection Megane Detection ֬ೝɾमਖ਼ ೝࣝ݁Ռ ֶशʢ܇࿅ʣ
λΠϜϥΠϯ ϝσΟΞ σʔληοτ ֶशʢ܇࿅ʣ TensorFlow rsync
ઓͷաఔΛಉਓࢽʹ
͞·͟·ͳ՝ σʔληοτ͕(#Λ͑ͨ͋ͨΓ͔ΒϩʔΧϧͷಉظ͕ࠔʹɻ ྖҬʢ3FHJPOʣͷઃఆͱϥϕϧͷ༩૾Ҏ্ʹෛՙ͕ߴ͍ɻ
ը૾͕ສຕΛಥഁ σʔλཧ͕ࢸٸͷ՝ʹ
ඪΛ࠶֬ೝ
Πϯλʔωοτ͔Β Έͷ؟ڸ່ͬը૾ΛࣗಈͰऩू͍ͨ͠
Ҏલͷߏ Downloader σʔληοτ Region + Label ઃఆ rsync
ྖҬʴϥϕϧ
৽͍͠ߏ Downloader σʔληοτ Tagઃఆ
λά megane girl
؟ڸ່ͬผϞσϧ Ϟσϧ 1.00 0.00
%BUBTFU.BOBHFSGPS"OESPJE
σϞ
https://twitter.com/35s_00/status/930366666973757441
https://twitter.com/_meganeco
/4'8ʢ/PU4BGF'PS8PSLʣ
/4'8ը૾
͞·͟·ͳϦεΫ ࡞ۀͷϊΠζ ਫ਼ਆతͳෛՙ ๏తϦεΫ
/4'8ը૾ͷݕग़
ֶश༻σʔληοτʢ/4'8ʣ ਖ਼ྫɿ ෛྫɿ ← NSFWը૾
܇࿅ɾֶश
ڭࢣ༗Γֶश ◦ × Ϟσϧ 1.00 0.00
Ϟσϧͷߏ conv 3x3x64 stride 1 conv 3x3x64 stride 1
ReLU ReLU conv 3x3x128 stride 1 conv 3x3x128 stride 1 ReLU conv 3x3x256 stride 1 conv 3x3x256 stride 1 ReLU output 1 256x256x1 max_pool 2x2 stride 2 max_pool 2x2 stride 2 ReLU ReLU Sigmoid max_pool 2x2 stride 2 conv 3x3x64 stride 1 ReLU fc 768 ReLU bn bn bn
Sigmoid
# モデル定義 NUM_CLASSES = 1 NAME = 'model3' IMAGE_SIZE =
256 CHANNELS = 3 def prepare_layers(image, training=False): with tf.variable_scope('inference'): conv1 = tf.layers.conv2d(image, 64, [3, 3], [1, 1], padding='SAME', activation=tf.nn.relu, use_bias=False, trainable=training, name='conv1_1') conv1 = tf.layers.conv2d(conv1, 64, [3, 3], [1, 1], padding='VALID', activation=tf.nn.relu, use_bias=False, trainable=training, name='conv1_2') conv1 = tf.layers.batch_normalization(conv1, trainable=training, name='bn_1')
conv2 = tf.layers.conv2d(pool1, 128, [3, 3], [1, 1], padding='VALID', activation=tf.nn.relu,
use_bias=False, trainable=training, name='conv2_1') conv2 = tf.layers.conv2d(conv2, 128, [3, 3], [1, 1], padding='VALID', activation=tf.nn.relu, use_bias=False, trainable=training, name='conv2_2') conv2 = tf.layers.batch_normalization(conv2, trainable=training, name='bn_2') pool2 = tf.layers.max_pooling2d(conv2, [2, 2], [2, 2])
conv3 = tf.layers.conv2d(pool2, 256, [3, 3], [1, 1], padding='VALID', activation=tf.nn.relu,
use_bias=False, trainable=training, name='conv4_1') conv3 = tf.layers.conv2d(conv3, 256, [3, 3], [1, 1], padding='VALID', activation=tf.nn.relu, use_bias=False, trainable=training, name='conv4_2') conv3 = tf.layers.batch_normalization(conv3, trainable=training, name='bn_4') pool3 = tf.layers.max_pooling2d(conv3, [2, 2], [2, 2]) conv = tf.layers.conv2d(pool3, 64, [1, 1], [1, 1], padding='VALID', activation=tf.nn.relu, use_bias=True, trainable=training, name='conv') return conv
def output_layers(prev, batch_size, keep_prob=0.8, training=False): flatten = tf.reshape(prev, [batch_size, -1])
fc1 = tf.layers.dense(flatten, 768, trainable=training, activation=tf.nn.relu, name='fc1') fc1 = tf.layers.dropout(fc1, rate=keep_prob, training=training) output = tf.layers.dense(fc1, NUM_CLASSES, trainable=training, activation=None, name='output') return output
def _loss(logits, labels, batch_size, positive_ratio): cross_entropy = tf.nn.sigmoid_cross_entropy_with_logits( labels=labels, logits=logits)
loss = tf.reduce_mean(cross_entropy) return loss def _init_optimizer(learning_rate): return tf.train.AdamOptimizer(learning_rate=learning_rate) ޡࠩؔͱ࠷దԽΞϧΰϦζϜ
ֶशΛ্ख͘ਐΊΔ
ਖ਼ྫɾෛྫͷൺ ਖ਼ྫɿ ෛྫɿ ← NSFWը૾ NSFW
def _hard_negative_mining(loss, labels, batch_size): positive_count = tf.reduce_sum(labels) positive_count = tf.reduce_max((positive_count,
1)) negative_count = positive_count * HARD_SAMPLE_MINING_RATIO negative_count = tf.reduce_max((negative_count, 1)) negative_count = tf.reduce_min((negative_count, batch_size)) positive_losses = loss * labels negative_losses = loss - positive_losses top_negative_losses, _ = tf.nn.top_k(negative_losses, k=tf.cast(negative_count, tf.int32)) loss = (tf.reduce_sum(positive_losses / positive_count) + tf.reduce_sum(top_negative_losses / negative_count)) return loss )BSE/FHBUJWF.JOJOH
ֶशڥʢ͘͞ΒͷߴՐྗίϯϐϡʔςΟϯάʣ $169FPO$PSFʷ .FNPSZ(# 44%(# (F'PSDF(595*5"/9ʢ1BTDBMΞʔΩςΫνϟʣ(#ʷ (F'PSDF(595Jʢ1BTDBMΞʔΩςΫνϟʣ(#ʷ
ֶश݅ ޡࠩؔަࠩΤϯτϩϐʔ ࠷దԽΞϧΰϦζϜ"EBN ֶश όοναΠζ
طଘͷσʔληοτʹਪʢJOGFSFODFʣΛ࣮ߦ Downloader σʔληοτ Tagઃఆ inference trainer ֶशࡁΈϞσϧ ֶश༻σʔληοτ
ਪ݁Ռ /4'8 Ұൠը૾ NSFW 8.6%
ֶश༻σʔληοτʢ/4'8ʣ ਖ਼ྫɿ ɹˠɹ ෛྫɿ ɹˠɹ
܇࿅ɾֶशʹ͔͔Δܭࢉ࣌ؒ
σϞ (16ɾ$16ͷൺֱ
$16ɾ(16ͷൺֱʢCBUDI4J[Fʣ 5*5"/9 TFDTUFQ 9FPO$PSF TFDTUFQ ࠓճͷϞσϧͷֶशʹ͍ͭͯ 5*5"/9ͷํ͕ഒ͍ʂ
$16ɾ(16ͷൺֱʢCBUDI4J[F ʣ 5*5"/9 (595J TFDTUFQ 9FPO$PSF TFDTUFQ
ࠓճͷϞσϧͷֶशʹ͍ͭͯ (16ʷͷํ͕ഒ͍ʂ
ࠓޙͷ՝
σʔληοταʔόʔͷ৴པੑ্
JOGFSFODFʢਪʣͷͨΊͷܭࢉࢿݯͷ֬อ Downloader σʔληοτ Tagઃఆ inference trainer ֶशࡁΈϞσϧ ֶश༻σʔληοτ
TAGS = [ 'original_art', 'nsfw', 'like', 'photo', 'illust', 'comic', 'face',
'girl', 'megane', ϥϕϧʢλάʣ 'school_uniform', 'blazer_uniform', 'sailor_uniform', 'gl', 'kemono', 'boy', 'bl', 'cat', 'dog', 'food', 'dislike', ]
.PWJEJVT
ਪΛ.PWJEJVTҠߦ Downloader σʔληοτ Tagઃఆ trainer ֶशࡁΈϞσϧ ֶश༻σʔληοτ inference
ΫϥεఆϞσϧ conv 3x3x64 stride 1 conv 3x3x64 stride 1
ReLU ReLU conv 3x3x128 stride 1 conv 3x3x128 stride 1 ReLU conv 3x3x256 stride 1 conv 3x3x256 stride 1 ReLU output 20 256x256x1 max_pool 2x2 stride 2 max_pool 2x2 stride 2 ReLU ReLU Sigmoid max_pool 2x2 stride 2 conv 3x3x64 stride 1 ReLU fc 768 ReLU bn bn bn
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 https://speakerdeck.com/keiji/20171209-sakura-ml-night