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。

Transcript

1. C-LIS CO., LTD.
2. C-LIS CO., LTD.  ༗ࢁ
   ܓೋ
   ʢ,FJKJ
   "3*:"."ʣ $-*4
   $0
   -5% "OESPJEΞϓϦ։ൃνϣοτσΩϧ Photo by Koji MORIGUCHI (MORIGCHOWDER) ػցֶश͸ͪΐͬͱ΍ͬͨ͜ͱ͋Γ·͢ Twitter΍ͬͯ·ͤΜ
3. ͘͞ΒͷػցֶशφΠτ
   
     5FOTPS'MPXͰ
 /4'8ը૾ݕग़
4. 5FOTPS'MPXʢ೥݄ൃදʣ  ػց஌ೳ޲͚ܭࢉϑϨʔϜϫʔΫ
   ࠷৽όʔδϣϯ
   ʢ೥݄ʣ
5.  ษڧձ΍Ζ͏ͥ
6.  (PPHMF
   %FWFMPQFS
   (SPVQ
7.  IUUQTHEHLPCFEPPSLFFQFSKQFWFOUT
8.  Πϯλʔωοτ͔Β
   ޷Έͷը૾ΛࣗಈͰऩू͍ͨ͠
9. © ࠜઇΕ͍ ؟ ڸ ͬ ່
10. ؟ڸ່ͬ൑ఆ  1 0
11. σʔληοτʢ೥݄࣌఺ʣ ؟ڸ່ͬ
    ɹຕ
    ඇ؟ڸ່ͬ
    ຕ  ؟ڸ່ͬ ඇ؟ڸ່ͬ ޡݕग़ ؟ڸ່ͬ ඇ؟ڸ່ͬ
12. { "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 
13. ߏ੒  Downloader  σʔληοτ Region + Label ઃఆ rsync
14. ཧ૝ͷߏ੒  Downloader Face Detection Megane Detection ֬ೝɾमਖ਼ ೝࣝ݁Ռ ֶशʢ܇࿅ʣ λΠϜϥΠϯ ϝσΟΞ σʔληοτ ֶशʢ܇࿅ʣ TensorFlow  rsync
15.  ௅ઓͷաఔΛಉਓࢽʹ
16. ͞·͟·ͳ՝୊ σʔληοτ͕(#Λ௒͑ͨ͋ͨΓ͔ΒϩʔΧϧ΁ͷಉظ͕ࠔ೉ʹɻ
    ྖҬʢ3FHJPOʣͷઃఆͱϥϕϧͷ෇༩͸૝૾Ҏ্ʹෛՙ͕ߴ͍ɻ
    
17.  ը૾͕ສຕΛಥഁ
    σʔλ੔ཧ͕ࢸٸͷ՝୊ʹ
18.  ໨ඪΛ࠶֬ೝ
19.  Πϯλʔωοτ͔Β
    ޷Έͷ؟ڸ່ͬը૾ΛࣗಈͰऩू͍ͨ͠
20. Ҏલͷߏ੒  Downloader  σʔληοτ Region + Label ઃఆ rsync
21.  ྖҬʴϥϕϧ
22. ৽͍͠ߏ੒  Downloader  σʔληοτ Tagઃఆ
23.  λά megane girl
24. ؟ڸ່ͬ൑ผϞσϧ  Ϟσϧ 1.00 0.00
25.  %BUBTFU.BOBHFS
    GPS
    "OESPJE
26.  σϞ
27. https://twitter.com/35s_00/status/930366666973757441
28.  https://twitter.com/_meganeco
29.  /4'8ʢ/PU
    4BGF
    'PS
    8PSLʣ
30. /4'8ը૾ 
31. ͞·͟·ͳϦεΫ ࡞ۀͷϊΠζ
    ਫ਼ਆతͳෛՙ
    ๏తϦεΫ 
32.  /4'8ը૾ͷݕग़
33. ֶश༻σʔληοτʢ/4'8ʣ ਖ਼ྫɿ 
    ෛྫɿ   ← NSFWը૾
34.  ܇࿅ɾֶश
35. ڭࢣ༗Γֶश  ◦ × Ϟσϧ 1.00 0.00
36. 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
37.  Sigmoid
38. # 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') 
39. 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]) 
40. 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 
41. 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 
42. 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)  ޡࠩؔ਺ͱ࠷దԽΞϧΰϦζϜ
43.  ֶशΛ্ख͘ਐΊΔ޻෉
44. ਖ਼ྫɾෛྫͷൺ཰ ਖ਼ྫɿ 
    ෛྫɿ   ← NSFWը૾ NSFW
45. 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
46. ֶश؀ڥʢ͘͞ΒͷߴՐྗίϯϐϡʔςΟϯάʣ $16
    9FPO
    
    $PSF
    ʷ
    
    .FNPSZ
    (#
    44%
    (#
    (F'PSDF
    (59
    5*5"/
    9ʢ1BTDBM
    ΞʔΩςΫνϟʣ
    (#
    ʷ
    
    (F'PSDF
    (59
    
    5Jʢ1BTDBM
    ΞʔΩςΫνϟʣ
    (#
    ʷ
     
47. ֶश৚݅ ޡࠩؔ਺
    ަࠩΤϯτϩϐʔ
    ࠷దԽΞϧΰϦζϜ
    "EBN
    ֶश཰
    
    όοναΠζ
     
48. 
49. 
50. طଘͷσʔληοτʹਪ࿦ʢJOGFSFODFʣΛ࣮ߦ  Downloader σʔληοτ Tagઃఆ inference trainer ֶशࡁΈϞσϧ ֶश༻σʔληοτ
51. ਪ࿦݁Ռ /4'8
    
    
    
    
    
     
    Ұൠը૾
        NSFW
 8.6%
52. ֶश༻σʔληοτʢ/4'8ʣ ਖ਼ྫɿ ɹˠɹ 
    ෛྫɿ ɹˠɹ  
53. ܇࿅ɾֶशʹ͔͔Δܭࢉ࣌ؒ 
54. 
55.  σϞ
    (16ɾ$16ͷൺֱ
56. $16ɾ(16ͷ଎౓ൺֱʢCBUDI4J[F
    ʣ 5*5"/
    9
    
    TFD
    
    TUFQ
    9FPO
    
    $PSF
    
    
    
    
    
    TFD
    
    TUFQ
     ࠓճͷϞσϧͷֶशʹ͍ͭͯ͸
    5*5"/
    9
    ͷํ͕ഒ଎͍ʂ
57. $16ɾ(16ͷ଎౓ൺֱʢCBUDI4J[F
    
    
    ʣ 5*5"/
    9
    
    
    (59
    5J
    
    
    
    
    
    TFD
    
    TUFQ
    9FPO
    
    $PSF
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    TFD
    
    TUFQ
     ࠓճͷϞσϧͷֶशʹ͍ͭͯ͸
    (16ʷ
    ͷํ͕ഒ଎͍ʂ
58.  ࠓޙͷ՝୊
59. σʔληοταʔόʔͷ৴པੑ޲্ 
60. JOGFSFODFʢਪ࿦ʣͷͨΊͷܭࢉࢿݯͷ֬อ  Downloader σʔληοτ Tagઃఆ inference trainer ֶशࡁΈϞσϧ ֶश༻σʔληοτ
61. 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', ]
62. .PWJEJVT 
63. ਪ࿦Λ.PWJEJVT΁Ҡߦ  Downloader σʔληοτ Tagઃఆ trainer ֶशࡁΈϞσϧ ֶश༻σʔληοτ inference
64. Ϋϥε൑ఆϞσϧ  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
65. C-LIS CO., LTD. ຊࢿྉ͸ɺ༗ݶձࣾγʔϦεͷஶ࡞෺Ͱ͢ɻຊࢿྉͷશ෦ɺ·ͨ͸Ұ෦ʹ͍ͭͯɺஶ࡞ऀ͔ΒจॻʹΑΔڐ୚Λಘͣʹෳ੡͢Δ͜ͱ͸ې͡ΒΕ͍ͯ·͢ɻ 5IF
    "OESPJE
    4UVEJP
    JDPO
    JT
    SFQSPEVDFE
    PS
    NPEJpFE
    GSPN
    XPSL
    DSFBUFE
    BOE
    TIBSFE
    CZ
    (PPHMF
    BOE
    VTFE
    BDDPSEJOH
    UP
    UFSNT
    EFTDSJCFE
    JO
    UIF
    $SFBUJWF
    $PNNPOT
    
    "UUSJCVUJPO
    -JDFOTF ֤੡඼໊ɾϒϥϯυ໊ɺձ໊ࣾͳͲ͸ɺҰൠʹ֤ࣾͷ঎ඪ·ͨ͸ొ࿥঎ඪͰ͢ɻຊࢿྉதͰ͸ɺ˜ɺšɺäΛׂѪ͍ͯ͠·͢ɻ 5IF
    "OESPJE
    SPCPU
    JT
    SFQSPEVDFE
    PS
    NPEJpFE
    GSPN
    XPSL
    DSFBUFE
    BOE
    TIBSFE
    CZ
    (PPHMF
    BOE
    VTFE
    BDDPSEJOH
    UP
    UFSNT
    EFTDSJCFE
    JO
    UIF
    $SFBUJWF
    $PNNPOT
    
    "UUSJCVUJPO
    -JDFOTF https://speakerdeck.com/keiji/20171209-sakura-ml-night