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。
C-LIS CO., LTD.
View Slide
C-LIS CO., LTD.༗ࢁܓೋʢ,FJKJ"3*:"."ʣ$-*4$0 -5%"OESPJEΞϓϦ։ൃνϣοτσΩϧPhoto by Koji MORIGUCHI (MORIGCHOWDER)ػցֶशͪΐͬͱͬͨ͜ͱ͋Γ·͢Twitterͬͯ·ͤΜ
͘͞ΒͷػցֶशφΠτ 5FOTPS'MPXͰ /4'8ը૾ݕग़
5FOTPS'MPXʢ݄ൃදʣػցೳ͚ܭࢉϑϨʔϜϫʔΫ࠷৽όʔδϣϯʢ݄ʣ
ษڧձΖ͏ͥ
(PPHMF%FWFMPQFS(SPVQ
IUUQTHEHLPCFEPPSLFFQFSKQFWFOUT
Πϯλʔωοτ͔ΒΈͷը૾ΛࣗಈͰऩू͍ͨ͠
© ࠜઇΕ͍؟ڸ່ͬ
؟ڸ່ͬఆ1 0
σʔληοτʢ݄࣌ʣ؟ڸ່ͬɹຕඇ؟ڸ່ͬຕ؟ڸ່ͬ ඇ؟ڸ່ͬޡݕग़؟ڸ່ͬඇ؟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
ߏDownloaderσʔληοτRegion +Label ઃఆrsync
ཧͷߏDownloader Face Detection Megane Detection֬ೝɾमਖ਼ೝࣝ݁Ռֶशʢ܇࿅ʣλΠϜϥΠϯ ϝσΟΞσʔληοτֶशʢ܇࿅ʣTensorFlowrsync
ઓͷաఔΛಉਓࢽʹ
͞·͟·ͳ՝σʔληοτ͕(#Λ͑ͨ͋ͨΓ͔ΒϩʔΧϧͷಉظ͕ࠔʹɻྖҬʢ3FHJPOʣͷઃఆͱϥϕϧͷ༩૾Ҏ্ʹෛՙ͕ߴ͍ɻ
ը૾͕ສຕΛಥഁσʔλཧ͕ࢸٸͷ՝ʹ
ඪΛ࠶֬ೝ
Πϯλʔωοτ͔ΒΈͷ؟ڸ່ͬը૾ΛࣗಈͰऩू͍ͨ͠
ҎલͷߏDownloaderσʔληοτRegion +Label ઃఆrsync
ྖҬʴϥϕϧ
৽͍͠ߏDownloaderσʔληοτTagઃఆ
λάmeganegirl
؟ڸ່ͬผϞσϧϞσϧ1.000.00
%BUBTFU.BOBHFSGPS"OESPJE
σϞ
https://twitter.com/35s_00/status/930366666973757441
https://twitter.com/_meganeco
/4'8ʢ/PU4BGF'PS8PSLʣ
/4'8ը૾
͞·͟·ͳϦεΫ࡞ۀͷϊΠζਫ਼ਆతͳෛՙ๏తϦεΫ
/4'8ը૾ͷݕग़
ֶश༻σʔληοτʢ/4'8ʣਖ਼ྫɿ ෛྫɿ ← NSFWը૾
܇࿅ɾֶश
ڭࢣ༗Γֶश○×Ϟσϧ1.000.00
Ϟσϧͷߏconv3x3x64stride 1conv3x3x64 stride 1ReLU ReLUconv3x3x128 stride 1conv3x3x128 stride 1ReLUconv3x3x256 stride 1conv3x3x256 stride 1ReLUoutput1256x256x1max_pool2x2stride 2max_pool2x2stride 2ReLU ReLU Sigmoidmax_pool2x2stride 2conv3x3x64 stride 1ReLUfc768ReLUbnbnbn
Sigmoid
# モデル定義NUM_CLASSES = 1NAME = 'model3'IMAGE_SIZE = 256CHANNELS = 3def 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 lossdef _init_optimizer(learning_rate):return tf.train.AdamOptimizer(learning_rate=learning_rate)ޡࠩؔͱ࠷దԽΞϧΰϦζϜ
ֶशΛ্ख͘ਐΊΔ
ਖ਼ྫɾෛྫͷൺਖ਼ྫɿ ෛྫɿ ← NSFWը૾NSFW
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_RATIOnegative_count = tf.reduce_max((negative_count, 1))negative_count = tf.reduce_min((negative_count, batch_size))positive_losses = loss * labelsnegative_losses = loss - positive_lossestop_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
ֶश݅ޡࠩؔަࠩΤϯτϩϐʔ࠷దԽΞϧΰϦζϜ"EBNֶशόοναΠζ
طଘͷσʔληοτʹਪʢJOGFSFODFʣΛ࣮ߦDownloaderσʔληοτTagઃఆinferencetrainerֶशࡁΈϞσϧ ֶश༻σʔληοτ
ਪ݁Ռ/4'8 Ұൠը૾ NSFW 8.6%
ֶश༻σʔληοτʢ/4'8ʣਖ਼ྫɿ ɹˠɹ ෛྫɿ ɹˠɹ
܇࿅ɾֶशʹ͔͔Δܭࢉ࣌ؒ
σϞ(16ɾ$16ͷൺֱ
$16ɾ(16ͷൺֱʢCBUDI4J[Fʣ5*5"/9TFDTUFQ9FPO$PSFTFDTUFQࠓճͷϞσϧͷֶशʹ͍ͭͯ5*5"/9ͷํ͕ഒ͍ʂ
$16ɾ(16ͷൺֱʢCBUDI4J[Fʣ5*5"/9 (595JTFDTUFQ9FPO$PSFTFDTUFQࠓճͷϞσϧͷֶशʹ͍ͭͯ(16ʷͷํ͕ഒ͍ʂ
ࠓޙͷ՝
σʔληοταʔόʔͷ৴པੑ্
JOGFSFODFʢਪʣͷͨΊͷܭࢉࢿݯͷ֬อDownloaderσʔληοτTagઃఆinferencetrainerֶशࡁΈϞσϧ ֶ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.PWJEJVTҠߦDownloaderσʔληοτTagઃఆtrainerֶशࡁΈϞσϧֶश༻σʔληοτinference
ΫϥεఆϞσϧconv3x3x64stride 1conv3x3x64 stride 1ReLU ReLUconv3x3x128 stride 1conv3x3x128 stride 1ReLUconv3x3x256 stride 1conv3x3x256 stride 1ReLUoutput20256x256x1max_pool2x2stride 2max_pool2x2stride 2ReLU ReLU Sigmoidmax_pool2x2stride 2conv3x3x64 stride 1ReLUfc768ReLUbnbnbn
C-LIS CO., LTD.ຊࢿྉɺ༗ݶձࣾγʔϦεͷஶ࡞Ͱ͢ɻຊࢿྉͷશ෦ɺ·ͨҰ෦ʹ͍ͭͯɺஶ࡞ऀ͔ΒจॻʹΑΔڐΛಘͣʹෳ͢Δ͜ͱې͡ΒΕ͍ͯ·͢ɻ5IF"OESPJE4UVEJPJDPOJTSFQSPEVDFEPSNPEJpFEGSPNXPSLDSFBUFEBOETIBSFECZ(PPHMFBOEVTFEBDDPSEJOHUPUFSNTEFTDSJCFEJOUIF$SFBUJWF$PNNPOT"UUSJCVUJPO-JDFOTF໊֤ɾϒϥϯυ໊ɺձ໊ࣾͳͲɺҰൠʹ֤ࣾͷඪ·ͨొඪͰ͢ɻຊࢿྉதͰɺɺɺäΛׂѪ͍ͯ͠·͢ɻ5IF"OESPJESPCPUJTSFQSPEVDFEPSNPEJpFEGSPNXPSLDSFBUFEBOETIBSFECZ(PPHMFBOEVTFEBDDPSEJOHUPUFSNTEFTDSJCFEJOUIF$SFBUJWF$PNNPOT"UUSJCVUJPO-JDFOTFhttps://speakerdeck.com/keiji/20171209-sakura-ml-night
keiji
smasato
orimanabu
moritalous
bengo4com
koheiawa
asei
tomoaki25
jark
mtx2s
yoshidashingo
kddisnowvillage
kc_nakanishi
samanthasiow
wjessup
iamctodd
vanstee
chriscoyier
mclloyd
jennybc
smashingmag
brettharned
thoeni
cherdarchuk
aarron