20161120-MLHands-on

2016年11月20日にGCPUG Tokushimaで開催した機械学習ハンズオンの資料です。

犬猫の画像を分類する畳み込みニューラルネットワーク（CNN）を構築します。
データセットにはCNNThe Oxford-IIIT Pet Datasetを使います。

Transcript

1. C-LIS CO., LTD.
   ػցֶशϋϯζΦϯ ($16(5PLVTIJNB
   

   View Slide

2. C-LIS CO., LTD.
   
   ༗ࢁܓೋʢ,FJKJ"3*:"."ʣ
   $-*4$0 -5%
   "OESPJEΞϓϦ։ൃνϣοτσΩϧ
   ػցֶशॳ৺ऀ
   Photo : Koji MORIGUCHI (AUN CREATIVE FIRM)
   

   View Slide

3. C-LIS CO., LTD.
   TM
   
   

   View Slide

4. 
   10% 1SJOU0O%FNBOE
   ిࢠॻ੶
   Impress R&D͔Βൃചத
   http://amzn.to/2axRog
   

   View Slide

5. 
   *OUSPEVDUJPO
   ʰػցֶशʱͱ͸ͳʹ͔
   

   View Slide

6. ػցֶशʢ͖͔͍͕͘͠Ύ͏ɺӳNBDIJOFMFBSOJOHʣͱ͸ɺਓ޻஌ೳʹ͓͚
   Δݚڀ՝୊ͷҰͭͰɺਓ͕ؒࣗવʹߦ͍ͬͯΔֶशೳྗͱಉ༷ͷػೳΛίϯ
   ϐϡʔλͰ࣮ݱ͠Α͏ͱ͢Δٕज़ɾख๏ͷ͜ͱͰ͋Δɻ
   
   ػցֶश
   ʢWikipedia: https://ja.wikipedia.org/wiki/%E6%A9%9F%E6%A2%B0%E5%AD%A6%E7%BF%92ʣ
   

   View Slide

7. σΟʔϓϥʔχϯάɺਂ૚ֶशʢӳEFFQMFBSOJOHʣͱ͸ɺଟ૚ߏ଄ͷχϡʔ
   ϥϧωοτϫʔΫʢσΟʔϓχϡʔϥϧωοτϫʔΫɺӳEFFQOFVSBM
   OFUXPSLʣΛ༻͍ͨػցֶशͰ͋Δɻ
   
   σΟʔϓϥʔχϯά
   ʢWikipedia: https://ja.wikipedia.org/wiki/%E3%83%87%E3%82%A3%E3%83%BC%E3%83%97%E3%83%A9%E3%83%BC%E3%83%8B%E3%83%B3%E3%82%B0ʣ
   

   View Slide

8. χϡʔϥϧωοτϫʔΫ ਆܦճ࿏໢ɺӳOFVSBMOFUXPSL //
   ͸ɺ೴ػೳʹ
   ݟΒΕΔ͍͔ͭ͘ͷಛੑΛܭࢉػ্ͷγϛϡϨʔγϣϯʹΑͬͯදݱ͢Δ͜ͱ
   Λ໨ࢦͨ͠਺ֶϞσϧͰ͋Δɻ
   
   χϡʔϥϧωοτϫʔΫ
   ʢWikipedia: https://ja.wikipedia.org/wiki/%E3%83%8B%E3%83%A5%E3%83%BC%E3%83%A9%E3%83%AB%E3%83%8D%E3%83%83%E3%83%88%E3%83%AF%E3%83%BC%E3%82%AFʣ
   

   View Slide

9. 
   ཁ͢Δʹ਺ֶ
   

   View Slide

10. C-LIS CO., LTD.
    TM
    5FOTPS'MPXJTBO0QFO4PVSDF4PGUXBSF-JCSBSZ
    GPS.BDIJOF*OUFMMJHFODF
    
    

    View Slide

11. 
    ϋϯζΦϯ
    

    View Slide

12. w 1ZUIPOʢ(PPHMF$PNQVUFS1MBUGPSN5PPMT͕ରԠͷͨΊʣ
    w ґଘύοέʔδ
    w TJY
    w 1*-1JMMPX
    w MYNM
    w QZUIPOHqBHT
    w 5FOTPS'MPXɹɹ˞ཁ֬ೝ
    
    ։ൃ؀ڥͷ֬ೝ
    

    View Slide

13. 
    ͋ΔͱΑ͍΋ͷ
    https://www.jetbrains.com/pycharm/
    

    View Slide

14. 
    ը૾෼ྨ
    ʢग़యIUUQTHJUIVCDPNDB[BMBNOJTUʣ ʢग़యIUUQTXXXDTUPSPOUPFEV_LSJ[DJGBSIUNMʣ
    MNIST CIFAR-10
    

    View Slide

15. 
    ݘೣ෼ྨ
    

    View Slide

16. σʔλͱਖ਼ղϥϕϧΛ༻ҙ͢Δ
    σʔλΛ5'3FDPSEܗࣜʹม׵͢Δ
    ϞσϧΛ࡞੒͢Δ
    σʔλΛධՁʢJOGFSFODFʣ͢Δ
    ϞσϧΛ܇࿅͢Δ
    ύϥϝʔλʔΛอଘ͢Δ
    ܇࿅ࡁΈϞσϧΛݕূ͢Δ
    
    େ·͔ͳखॱ
    

    View Slide

17. 
    εςοϓ̍
    σʔλͱਖ਼ղϥϕϧΛ༻ҙ͢Δ
    

    View Slide

18. 
    σʔληοτͷμ΢ϯϩʔυ
    http://www.robots.ox.ac.uk/~vgg/data/pets/
    

    View Slide

19. 
    JNBHFT
    

    View Slide

20. 
    BOOPUBUJPOT
    ೣ: Abyssinian
    ೣ: Ragdoll
    ೣ: Russian_Blue
    ݘ: Shiba
    ݘ: saint_bernard
    ݘ: samoyed
    

    View Slide

21. 
    30*ʢ3FHJPO0G*OUFSFTUʣ
    

    View Slide

22. 
    ίʔυͷμ΢ϯϩʔυ
    IUUQTHPPHM,TW,I2
    

    View Slide

23. 
    εςοϓ̎
    σʔλΛ5'3FDPSEܗࣜʹม׵͢Δ
    

    View Slide

24. 
    લॲཧ
    Abyssinian_1.jpg
    

    View Slide

25. 
    OXIIIT
    Abyssinian_1.jpg
    
    OXFORD-IIIT Pet Dataset
    OXIIIT
    flickr
    
    
    600
    400
    3
    
    0
    
    BOOPUBUJPOTYNMT
    
    cat
    Frontal
    0
    0
    
    333
    72
    425
    158
    
    0
    
    
    Abyssinian_1.xml
    

    View Slide

26. 
    30*ͷ੾Γൈ͖
    

    View Slide

27. 
    ίʔυΛมߋ
    bndboxͷ֤஋͔Βը૾Λ੾Γൈ͍ͯอଘ͢Δ
    dataset/crop_roi_image.py
    

    View Slide

28. FLAGS = gflags.FLAGS

    gflags.DEFINE_string('image_dir', None, 'ॲཧ͢Δը૾ͷσΟϨΫτϦ')

    gflags.DEFINE_string('xml_dir', None, 'ॲཧ͢ΔXMLͷσΟϨΫτϦ')

    gflags.DEFINE_string('output_dir', None, 'ग़ྗ͢ΔσΟϨΫτϦ')
    
    HqBHT
    $ python crop_roi_image.py \
    --image_dir /Users/keiji_ariyama/Downloads/catsanddogs/images \
    --xml_dir /Users/keiji_ariyama/Downloads/catsanddogs/annotations/xmls \
    --output_dir /Users/keiji_ariyama/Desktop/test
    

    View Slide

29. 
    ը૾ʹϥϕϧΛ෇Ճͯ͠5'3FDPSE΁ม׵
    + Cat (1)
    Abyssinian_1.jpg
    TFRecord
    

    View Slide

30. #Image CLASS-ID SPECIES BREED ID
    #ID: 1:37 Class ids
    #SPECIES: 1:Cat 2:Dog
    #BREED ID: 1-25:Cat 1:12:Dog
    #All images with 1st letter as captial are cat images
    #images with small first letter are dog images
    Abyssinian_196 1 1 1
    Abyssinian_197 1 1 1
    Abyssinian_19 1 1 1
    Abyssinian_1 1 1 1
    
    BOOPUBUJPOTUSBJOWBMUYU
    Abyssinian_1.xml
    

    View Slide

31. 
    ίʔυΛมߋ
    5'3FDPSEʹSPECIES ͷ஋Λ௥Ճ
    dataset/convert_to_tfrecord.py
    

    View Slide

32. 
    ςετσʔλΛ෼཭͢Δ
    σʔληοτ
    ςετσʔλ
    

    View Slide

33. 
    εςοϓ̏
    ϞσϧΛ࡞੒͢Δ
    

    View Slide

34. 
    ৞ࠐΈχϡʔϥϧωοτϫʔΫʢ̘̣̣ʣ
    DPOW
    QPPM GD PVUQVU
    JOQVU
    

    View Slide

35. 
    ৞ࠐΈʢ$POWPMVUJPOʣ૚
    

    View Slide

36. # conv1

    with tf.variable_scope('conv1') as scope:

    weights = _get_weights([5, 5, 3, 32], stddev=0.01)

    conv1 = tf.nn.conv2d(image_node, weights, [1, 1, 1, 1], padding='SAME')

    biases = tf.get_variable('biases', [32],

    initializer=tf.constant_initializer(0.0))

    bias = tf.nn.bias_add(conv1, biases)

    conv1 = tf.nn.relu(bias, name=scope.name)

    
    ৞ࠐΈʢ$POWPMVUJPOʣ૚
    

    View Slide

37. 
    ৞ࠐΈʢ$POWPMVUJPOʣ૚
    ৞ΈࠐΉϑΟϧλͷܗʢshapeʣ
    ຕ਺
    εϥΠυ෯
    ύσΟϯάͷ༗ແ
    ׆ੑԽؔ਺
    

    View Slide

38. 
    ίʔυΛมߋ
    ৞ࠐΈ૚ͷઃఆΛܾΊΔ
    model.py
    

    View Slide

39. 
    ϓʔϦϯάʢ1PPMJOHʣ૚
    εϥΠυ෯2
    

    View Slide

40. pool1 = tf.nn.max_pool(conv1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1],

    padding='SAME', name='pool1')

    
    ϓʔϦϯάʢ1PPMJOHʣ૚
    

    View Slide

41. 
    ϓʔϦϯάʢ1PPMJOHʣ૚
    ϓʔϦϯάͷछྨ
    Χʔωϧʢ΢Οϯυ΢ʣͷେ͖͞
    εϥΠυ෯
    ύσΟϯάͷ༗ແ
    

    View Slide

42. 
    ίʔυΛมߋ
    ϓʔϦϯά૚ͷઃఆΛܾΊΔ
    model.py
    

    View Slide

43. 
    શ݁߹ʢ'VMMZDPOOFDUFEʣ૚
    ※ ʮதؒ૚ʯʮӅΕ૚ʯͱ͍͏৔߹΋͋Δ
    :
    :
    :
    :
    :
    

    View Slide

44. reshape = tf.reshape(pool1, [batch_size, -1])

    dim = reshape.get_shape()[1].value

    

    # fc2

    with tf.variable_scope('fc2') as scope:

    weights = _get_weights([dim, 384], stddev=0.04)

    biases = _get_biases([384], value=0.1)

    fc2 = tf.nn.relu(tf.nn.bias_add(tf.matmul(reshape, weights), biases),

    name=scope.name)

    
    શ݁߹ʢ'VMMZDPOOFDUFEʣ૚
    

    View Slide

45. 
    શ݁߹ʢ'VMMZDPOOFDUFEʣ૚
    :
    :
    :
    :
    :
    ϊʔυͷ਺
    ׆ੑԽؔ਺
    

    View Slide

46. 
    ίʔυΛมߋ
    શ݁߹૚ͷϊʔυ਺ΛܾΊΔ
    model.py
    

    View Slide

47. 
    ग़ྗ૚
    :
    :
    :
    :
    :
    

    View Slide

48. with tf.variable_scope('output') as scope:

    weights = _get_weights(shape=[384, 2], stddev=1 / 384.0)

    biases = _get_biases([2], value=0.0)

    logits = tf.add(tf.matmul(fc2, weights), biases, name='logits')

    
    ग़ྗ૚
    

    View Slide

49. 
    ग़ྗ૚
    :
    :
    :
    :
    :
    ෼ྨ͢ΔΫϥεͷ਺
    

    View Slide

50. 
    ίʔυΛมߋ
    ग़ྗ૚ͷϊʔυ਺ΛܾΊΔ
    model.py
    

    View Slide

51. 
    εςοϓ̐
    σʔλΛධՁʢJOGFSFODFʣ͢Δ
    

    View Slide

52. ظ଴͢Δ஋ͱJOGFSFODFͷ݁ՌͷࠩʢޡࠩʣΛٻΊΔɻ
    
    ޡࠩؔ਺
    

    View Slide

53. def __loss(logits, label):

    labels = tf.cast(label, tf.int64)

    cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(

    logits, labels, name='cross_entropy_per_example')

    cross_entropy_mean = tf.reduce_mean(cross_entropy, name='cross_entropy')

    return cross_entropy_mean

    
    ޡࠩؔ਺
    

    View Slide

54. 
    ίʔυΛมߋ
    εςοϓ͝ͱʹϩε཰ʢޡࠩʣΛදࣔ͢Δ
    train.py
    

    View Slide

55. 
    εςοϓ̑
    ϞσϧΛ܇࿅͢Δ
    

    View Slide

56. ޡࠩʹج͍ͮͯύϥϝʔλʔΛௐ੔͢Δɻ
    w ޯ഑߱Լ๏
    w "EBN0QUJNJ[FS
    w "EB(SBEͳͲ
    
    ֶशʢ܇࿅ʣΞϧΰϦζϜ
    

    View Slide

57. global_step = tf.Variable(0, trainable=False)

    loss = __loss(logits, specie_batch)

    train_op = tf.train.GradientDescentOptimizer(FLAGS.learning_rate).minimize(

    loss, global_step)

    
    ֶशʢ܇࿅ʣΞϧΰϦζϜ
    

    View Slide

58. 
    ίʔυΛมߋ
    train_opΛ࣮ߦ͢Δ
    train.py
    

    View Slide

59. # fc2

    with tf.variable_scope('fc2') as scope:

    weights = _get_weights([dim, 384], stddev=0.04)

    biases = _get_biases([384], value=0.1)

    fc2 = tf.nn.relu(tf.nn.bias_add(tf.matmul(reshape, weights), biases),

    name=scope.name)

    
    ύϥϝʔλʔ
    

    View Slide

60. 
    ॏΈʢXFJHIUʣͱόΠΞεʢCJBTʣ
    :
    :
    :
    :
    :
    w b
    

    View Slide

61. 
    ׆ੑԽؔ਺
    :
    :
    :
    :
    :
    

    View Slide

62. # fc2

    with tf.variable_scope('fc2') as scope:

    weights = _get_weights([dim, 384], stddev=0.04)

    biases = _get_biases([384], value=0.1)

    fc2 = tf.nn.relu(tf.nn.bias_add(tf.matmul(reshape, weights), biases),

    name=scope.name)

    
    ׆ੑԽؔ਺
    

    View Slide

63. 
    3F-6ʢ3FDUJpFE-JOFBSʣ
    ʢग़య: http://cs231n.github.io/neural-networks-1/ʣ
    

    View Slide

64. 
    4JHNPJE
    ʢग़య: https://commons.wikimedia.org/wiki/File:Sigmoid-function-2.svgʣ
    

    View Slide

65. 
    )ZQFSCPMJD5BOHFOU
    ʢग़య: https://commons.wikimedia.org/wiki/File:Hyperbolic_Tangent.svgʣ
    

    View Slide

66. 
    εςοϓ̒
    ύϥϝʔλʔΛอଘ͢Δ
    

    View Slide

67. saver = tf.train.Saver(tf.all_variables())

    

    with tf.Session() as sess:

    sess.run(tf.initialize_all_variables())

    step = 0

    while step < FLAGS.max_steps:

    _, step, loss_value = sess.run([train_op, global_step, loss])

    

    if step % 100 == 0:

    print('Step: %d, loss: %.4f' % (step, loss_value))

    saver.save(sess, os.path.join(FLAGS.checkpoint_dir, 'cad_checkpoint'),
    global_step=step)
    
    4BWFSʹΑΔύϥϝʔλʔͷอଘ
    

    View Slide

68. 
    ίʔυΛมߋ
    ύϥϝʔλʔΛอଘ͢Δ
    train.py
    

    View Slide

69. 
    εςοϓ̓
    ܇࿅ࡁΈϞσϧΛݕূ͢Δ
    

    View Slide

70. def __restore(sess, saver):

    ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)

    if ckpt and ckpt.model_checkpoint_path:

    saver.restore(sess, ckpt.model_checkpoint_path)
    
    4BWFSʹΑΔύϥϝʔλʔͷಡΈࠐΈ
    

    View Slide

71. saver = tf.train.Saver(tf.all_variables())

    

    with tf.Session() as sess:

    __restore(sess, saver)
    
    4BWFSʹΑΔύϥϝʔλʔͷಡΈࠐΈ
    

    View Slide

72. logits = model.inference(image_batch, batch_size=FLAGS.batch_size)

    top_k_op = tf.nn.in_top_k(logits, specie_batch, 1)

    for step in range(num_iter):
    predictions = sess.run(top_k_op)

    true_count += np.sum(predictions)

    
    ݕূ
    

    View Slide

73. 
    ίʔυΛมߋ
    ൑ఆʹࣦഊͨ͠ϑΝΠϧ໊Λ

    දࣔ͢Δ
    eval.py
    

    View Slide

74. 
    ܇࿅ʢֶशʣͱ͸
    σʔλ inference logits
    ਖ਼ղϥϕϧ
    ޡࠩؔ਺
    ޡࠩʢϩεʣ
    ֶशΞϧΰϦζϜ
    ύϥϝʔλʔ
    (weights / biases)
    

    View Slide

75. 
    ܇࿅ʢֶशʣͱ͸
    inference
    0
    1
    

    View Slide

76. 
    ܇࿅ʢֶशʣͱ͸
    inference
    0
    1
    

    View Slide

77. 
    $MPVE.-Ͱ܇࿅͢Δ
    

    View Slide

78. 
    (PPHMF$MPVE1MBUGPSNͷ
    ར༻Λ։࢝
    

    View Slide

79. 
    ϓϩδΣΫτͷ࡞੒
    

    View Slide

80. 
    Ϋʔϙϯͷొ࿥
    https://cloud.google.com/redeem
    

    View Slide

81. C-LIS CO., LTD.
    ˞ϓϩδΣΫτ࡞੒࣌ʹબ୒ͨ͠

    ੥ٻઌΞΧ΢ϯτͱ

    Ұக͍ͯ͠Δ͔֬ೝ
    

    View Slide

82. 
    $MPVE.-ͷ༗ޮԽ
    

    View Slide

83. 
    $MPVE.-ͷ༗ޮԽ
    

    View Slide

84. 
    $MPVE4%,ͷΠϯετʔϧ
    https://cloud.google.com/sdk/
    

    View Slide

85. 
    $MPVE4%,ͷΠϯετʔϧ
    

    View Slide

86. $ gcloud auth
    $ gcloud beta ml init-project
    Cloud ML needs to add its service accounts to your project
    (catsanddogs-149703) as Editors. This will enable Cloud Machine
    Learning to access resources in your project when running your
    training and prediction jobs.
    Do you want to continue (Y/n)?
    
    $MPVE.-ͷϓϩδΣΫτΛॳظԽ
    

    View Slide

87. $ PROJECT_ID=$(gcloud config list project --format "value(core.project)")
    $ BUCKET_NAME="cats_and_dogs"
    $ gsutil mb -l us-central1 gs://$BUCKET_NAME
    
    $MPVE.-༻ͷόέοτΛ࡞੒
    ※ usͷΑ͏ͳMulti-regionalͰ͸ͳ͘RegionalͰࢦఆ͢Δඞཁ͕͋Δ

    https://cloud.google.com/ml/docs/how-tos/getting-set-up
    

    View Slide

88. 
    ࡞੒ͨ͠όέοτʹ܇࿅σʔλΛసૹ
    

    View Slide

89. TSDσΟϨΫτϦʹ __init__.py Λ࡞੒͢Δɻ
    
    εΫϦϓτΛύοέʔδԽ
    

    View Slide

90. ϑΝΠϧγεςϜ΁ΞΫηε͢ΔॲཧΛɺ͢΂ͯ5FOTPS'MPXͷ

    tensorflow.python.lib.io.file_ioΛ࢖͏Α͏ʹ͢Δɻ
    
    5FOTPS'MPXͷίʔυΛมߋ
    

    View Slide

91. $ gcloud beta ml jobs submit training cat_and_dogs \
    --package-path=/Users/keiji_ariyama/Developments/catsanddogs/src \
    --module-name=src.train \
    --staging-bucket=gs://cats_and_dogs \
    --region=us-central1 \
    -- \
    --train_dir gs://cats_and_dogs/cad_train \
    --checkpoint_dir gs://cats_and_dogs/cad_checkpoint \
    --max_steps 100
    
    $MPVE.-༻ʹδϣϒΛొ࿥͢Δ
    

    View Slide

92. 
    ΄Βɺ؆୯ʂ
    

    View Slide

93. ݱࡏSpeciesͰ෼ྨ͍ͯ͠Δ΋ͷΛ

    BreedͰ෼ྨ͢ΔΑ͏ʹมߋ͢Δɻ
    
    νϟϨϯδ
    

    View Slide

94. C-LIS CO., LTD.
    ֤੡඼໊ɾϒϥϯυ໊ɺձ໊ࣾͳͲ͸ɺҰൠʹ֤ࣾͷ঎ඪ·ͨ͸ొ࿥঎ඪͰ͢ɻຊࢿྉதͰ͸ɺ˜ɺšɺäΛׂѪ͍ͯ͠·͢ɻ
    ຊࢿྉ͸ɺ༗ݶձࣾγʔϦεͷஶ࡞෺Ͱ͋ΓɺΫϦΤΠςΟϒίϞϯζͷදࣔඇӦརܧঝ6OQPSUFEϥΠηϯεͷݩͰެ։͍ͯ͠·͢ɻ
    

    View Slide

95. 
    ݄೔ʙ
    https://devfestkansai.connpass.com/event/42864/
    

    View Slide