TFUG KANSAI 20190928

2019年9月28日、Sansan株式会社大阪支店で開催されたTensorFlow User Group KANSAI Meet upの発表資料です。

TensorFlow 2.0について。
TensorFlowで趣味の画像収集サーバーを作る 2019年9月号

Transcript

1. C-LIS CO., LTD.
   

   View Slide

2. C-LIS CO., LTD.
   
   ༗ࢁܓೋʢ,FJKJ"3*:"."ʣ
   $-*4$0 -5%
   Photo : Koji MORIGUCHI (MORIGCHOWDER)
   "OESPJEΞϓϦ։ൃνϣοτσΩϧ
   ʮझຯͰػցֶशΛ΍͍ͬͯΔऀͩʯ
   ΍ͬͯ·ͤΜ
   

   View Slide

3. C-LIS CO., LTD.
   
   IUUQTLOPXMFEHFTBLVSBBEKQ
   

   View Slide

4. 
   5FOTPS'MPX6TFS(SPVQ,"/4"*
   BU4BOTBOגࣜձࣾ
   
   5FOTPS'MPXʹҠߦͨ͠࿩
   5FOTPS'MPXͰझຯͷը૾ऩूαʔόʔΛ࡞Δ೥݄߸
   

   View Slide

5. 
   5FOTPS'MPX͸

   ͱͯ΋ྑ͍ϑϨʔϜϫʔΫͰ͢ɻ
   

   View Slide

6. conv2d (from tensorflow.python.layers.convolutional) is deprecated and
   will be removed in a future version.
   Instructions for updating: Use `tf.keras.layers.Conv2D` instead.
   flatten (from tensorflow.python.layers.core) is deprecated and
   will be removed in a future version.
   Instructions for updating: Use keras.layers.flatten instead.
   dense (from tensorflow.python.layers.core) is deprecated and
   will be removed in a future version.
   Instructions for updating: Use keras.layers.dense instead.
   dropout (from tensorflow.python.layers.core) is deprecated
   and will be removed in a future version.
   Instructions for updating: Use keras.layers.dropout instead.
   
   JTEFQSFDBUFEBOEXJMMCFSFNPWFEJOBGVUVSFWFSTJPO
   

   View Slide

7. def layers(tag_name, image, training=False, reuse=None):
   scope = '%s/%s' % (tag_name, MODEL_NAME)
   with tf.variable_scope(scope, reuse=reuse):
   conv = tf.layers.conv2D(image,
   BASE_CHANNEL, [3, 3], [1, 1],
   padding='SAME',
   activation=tf.nn.relu,
   use_bias=True,
   name='conv')
   
   UGMBZFST
   

   View Slide

8. def layers(tag_name, image, training=False, reuse=None):
   scope = '%s/%s' % (tag_name, MODEL_NAME)
   with tf.variable_scope(scope, reuse=reuse):
   conv = tf.keras.layers.Conv2D(BASE_CHANNEL, [3, 3], [1, 1],
   padding='SAME',
   activation=tf.nn.relu,
   use_bias=True,
   name='conv')(image)
   
   UGLFSBTMBZFST
   

   View Slide

9. 
   IUUQTUXJUUFSDPNLFJKJ@BSJZBNBTUBUVT
   

   View Slide

10. 
    IUUQTXXXGBDFCPPLDPNHSPVQTQFSNBMJOL
    

    View Slide

11. 
    

    View Slide

12. ౖΓͷ(JU)VC*TTVF
    
    IUUQTHJUIVCDPNUFOTPSqPXUFOTPSqPXJTTVFT
    tf.keras.layers (in TF 1.13.1): variable_scope does not work
    

    View Slide

13. def layers(tag_name, image, training=False, reuse=None):
    scope = '%s/%s' % (tag_name, MODEL_NAME)
    with tf.variable_scope(scope, reuse=reuse):
    conv = tf.keras.layers.Conv2D(BASE_CHANNEL, [3, 3], [1, 1],
    padding='SAME',
    activation=tf.nn.relu,
    use_bias=True,
    name='conv')(image)
    
    UGWBSJBCMF@TDPQF
    

    View Slide

14. def _build_train_graph(train_path_list):
    image_batch, label_batch = _load_dataset(train_path_list, is_train=True)
    logits = model.layers('megane', image_batch, training=True)
    train_op = _optimize(logits, label_batch)
    return train_op
    def _build_eval_graph(eval_path_list):
    image_batch, label_batch = _load_dataset(eval_path_list, is_train=False)
    logits = model.layers('megane', image_batch, reuse=True)
    accuracy = _calc_accuracy(logits, label_batch)
    return accuracy
    
    Ϟσϧͷ࠶ར༻
    

    View Slide

15. 
    5FOTPS'MPX
    

    View Slide

16. LABEL_BYTE = 1
    IMAGE_BYTES = 32 * 32 * 3
    RECORD_BYTES = LABEL_BYTE + IMAGE_BYTES
    def _load_dataset(data_path_list):
    def _process_record(record):
    value = tf.io.decode_raw(record, tf.uint8)
    label = value[0]
    image = value[1:]
    image = tf.reshape(image, (3, 32, 32))
    image = tf.transpose(image, (1, 2, 0))
    image = tf.cast(image, tf.float32)
    image = image / 255
    return image, label
    dataset = tf.data.FixedLengthRecordDataset(
    data_path_list,
    RECORD_BYTES)
    return dataset.map(_process_record)
    
    %BUBTFU"1*
    

    View Slide

17. class MyModel(Model):
    def __init__(self):
    super(MyModel, self).__init__()
    self.conv1 = Conv2D(32, 3, activation='relu')
    self.flatten = Flatten()
    self.d1 = Dense(128, activation='relu')
    self.d2 = Dense(10, activation='softmax')
    def call(self, x):
    x = self.conv1(x)
    x = self.flatten(x)
    x = self.d1(x)
    return self.d2(x)
    # Create an instance of the model
    model = MyModel()
    
    Ϟσϧͷఆٛ
    

    View Slide

18. loss_object = tf.keras.losses.SparseCategoricalCrossentropy()
    optimizer = tf.keras.optimizers.Adam()
    
    ଛࣦؔ਺ɾ࠷దԽΞϧΰϦζϜ
    

    View Slide

19. train_loss = tf.keras.metrics.Mean(name='train_loss')
    train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy')
    @tf.function
    def train_step(images, labels):
    with tf.GradientTape() as tape:
    predictions = model(images)
    loss = loss_object(labels, predictions)
    gradients = tape.gradient(loss, model.trainable_variables)
    optimizer.apply_gradients(zip(gradients, model.trainable_variables))
    train_loss(loss)
    train_accuracy(labels, predictions)
    
    ֶशͱධՁ
    

    View Slide

20. test_loss = tf.keras.metrics.Mean(name='test_loss')
    test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='test_accuracy')
    @tf.function
    def test_step(images, labels):
    predictions = model(images)
    t_loss = loss_object(labels, predictions)
    test_loss(t_loss)
    test_accuracy(labels, predictions)
    
    ֶशͱධՁ
    

    View Slide

21. EPOCHS = 5
    SUMMARY_DIR = './summary'
    TRAIN_BATCH_SIZE = 32
    TEST_BATCH_SIZE = 32
    import time
    train_dataset = _load_dataset(train_path_list).batch(TRAIN_BATCH_SIZE)
    test_dataset = _load_dataset(test_path_list).batch(TEST_BATCH_SIZE)
    
    ֶशͱධՁ
    

    View Slide

22. for epoch in range(EPOCHS):
    for images, labels in train_dataset:
    train_step(images, labels)
    for test_images, test_labels in test_dataset:
    test_step(test_images, test_labels)
    template = 'Epoch {}, Train Loss: {}, Train Accuracy: {}, Test Loss: {}, Test Accuracy: {}'
    print(template.format(epoch+1,
    train_loss.result(),
    train_accuracy.result()*100,
    test_loss.result(),
    test_accuracy.result()*100))
    # Reset the metrics for the next epoch
    train_loss.reset_states()
    train_accuracy.reset_states()
    test_loss.reset_states()
    test_accuracy.reset_states()
    
    

    View Slide

23. 
    2VBOUJ[BUJPOBXBSF5SBJOJOH
    

    View Slide

24. ϞσϧʢύϥϝʔλʔʣͷྔࢠԽ
    
    CJUුಈখ਺఺਺ΛCJU੔਺ʹม׵ʢྔࢠԽʣ
    

    View Slide

25. g = tf.get_default_graph()
    tf.contrib.quantize.create_training_graph(input_graph=g,
    quant_delay=100000)
    
    2VBOUJ[BUJPOBXBSF5SBJOJOH
    •ϑΝΠϯνϡʔχϯάͰ࢖͏
    •quant_delayΛ஫ҙਂ͘ઃఆ͢Δ
    •άϥϑͷॻ͖׵͑Λߦ͏ͨΊɺ

    ύϥϝʔλʔϑΝΠϧͷޓ׵ੑ͸ࣦΘΕΔ
    

    View Slide

26. 
    2VBOUJ[BUJPOBXBSF5SBJOJOH
    •ϑΝΠϯνϡʔχϯάͰ࢖͏
    •quant_delayΛ஫ҙਂ͘ઃఆ͢Δ
    •άϥϑͷॻ͖׵͑Λߦ͏ͨΊɺ

    ύϥϝʔλʔϑΝΠϧͷޓ׵ੑ͸ࣦΘΕΔ
    Quantization-aware ༗Γ

    quant-delay MAX_VALUE
    Quantization-aware ༗Γ

    quant-delay 0
    Quantization-aware ແ͠
    

    View Slide

27. 
    IUUQTUXJUUFSDPNLFJKJ@BSJZBNBTUBUVT
    

    View Slide

28. quantization-aware training Λαϙʔτ͍ͯͨ͠ܭࢉάϥϑͷॻ͖׵͑ؔ਺͸ɺTensorFlow 2.0ʹΑΔϞσϧΛ
    αϙʔτ͠·ͤΜɻ ·ͨɺTensorFlow Lite ͷ quantization API ͸ɺKeras API Λ௨ͯ͡ quantization-aware
    training Λαϙʔτ͢Δํ޲Ͱ࡞Γ௚͠ͱ߹ཧԽΛקΊ͍ͯΔ࠷தͰ͢ɻ ৽͍͠ quantization API ͕ϩʔϯν
    ͞ΕΔ·Ͱ͸ɺ͜ΕΒͷଐੑ͸ 2.0 API ͔Β࡟আ͞Ε·͢ɻ ॻ͖׵͑ؔ਺ʹΑͬͯϞσϧΛม׵͍ͨ͠৔߹
    ͸ tf.compat.v1.TFLiteConverter Λ࢖͍ͬͯͩ͘͞ɻ
    
    IUUQTXXXUFOTPSqPXPSHMJUFDPOWFSUQZUIPO@BQJRVBOUJ[BUJPOBXBSF@USBJOJOH
    

    View Slide

29. 
    ʹ͢ΔͳΒͤΊͯɺ

    Yʹ͋Δػೳ͸࢖͑ΔΑ͏ʹ͍ͯͩ͘͠͞ɻ
    (PPHMF͞Μ
    

    View Slide

30. 
    5FOTPS'MPXͰझຯͷը૾ऩूαʔόʔΛ࡞Δ
    ೥݄߸
    

    View Slide

31. 
    σϞ
    

    View Slide

32. ֓ཁ
    
    ػցֶशج൫
    ετϨʔδ

    αʔόʔ
    σʔλऩूݩ
    αʔϏε
    Android

    ΫϥΠΞϯτ
    Web
    ΠϯλʔϑΣʔε
    σʔληοτ؅ཧαʔόʔ
    σʔλϕʔε
    σʔλͷऩू
    MeganeCo
    Playground
    ֶशσʔλͷੜ੒
    ֶशࡁϞσϧ

    ʹΑΔਪ࿦
    Quad GPUs Radeon RX Vega 64 x 2
    6TB HDD
    256GB SSD
    ֶश༻σʔλʢTFRecordʣ
    ֶशࡁϞσϧ
    

    View Slide

33. ػցֶशج൫
    
    Quad GPUs Radeon RX Vega 64 x 2
    ͘͞ΒͷVPSʢ2Gʣ
    Master
    Kubernetes
    Argo workflow
    

    View Slide

34. ߴՐྗίϯϐϡʔςΟϯάαʔόʔʢݕূ࣮ݧػʣ
    $16ɹɹɹɹ9FPO&WίΞʷ
    .FNPSZɹɹ(#
    ετϨʔδɹ44%(#ʷʢ3"*%ʣ
    (16ɹɹɹ/7*%*"5*5"/9(#
    ɹɹɹɹɹ/7*%*"5*5"/9(#
    ɹɹɹɹɹ/7*%*"(F'PSDF(595J(#
    ɹɹɹɹɹ/7*%*"(F'PSDF(595J(#
    
    ʢDriver Version: 410.48ʣ
    

    View Slide

35. ߴՐྗίϯϐϡʔςΟϯάαʔόʔʢݕূ࣮ݧػʣ
    $16ɹɹɹɹ9FPO&WίΞʷ
    .FNPSZɹɹ(#
    ετϨʔδɹ44%(#ʷʢ3"*%ʣ
    (16ɹɹɹ".%3BEFPO7FHB(#
    ".%3BEFPO7FHB(#
    
    

    View Slide

36. 
    

    View Slide

37. 
    Label input_size
    Positive
    sample
    Negative
    sample
    Train step
    (100 step)
    Sensitivity Specificity IoU
    megane 256x256 19,601 46,950 3,920 0.9242 0.9578 0.8726
    favorite 256x256 16,921 29,960 1,050 0.8545 0.7926 0.7377
    illust 256x256 24,421 38,587 1,825 0.9601 0.9707 0.9434
    photo 256x256 16,296 30,606 1,725 0.9651 0.9339 0.9141
    comic 256x256 18,646 4,067 900 0.9950 0.9727 0.9068
    face 256x256 19,573 25,033 2,000 0.9629 0.9168 0.9413
    nsfw 256x256 18,002 41,258 3,730 0.8399 0.8748 0.7179
    nsfw(mobile) 128x128 18,002 41,258 2,230 0.8313 0.8139 0.6637
    kemono 256x256 5,863 21,174 1,225 0.8460 0.8383 0.5908
    sailor_uniform 256x256 3,060 10,242 1,980 0.8501 0.8954 0.6904
    blazer_uniform 256x256 958 8,681 550 0.7126 0.8555 0.3229
    cat 256x256 5,167 21,417 725 0.8992 0.9325 0.7472
    food 256x256 5,873 20,165 995 0.9198 0.9487 0.8173
    food(mobile) 128x128 5,873 20,165 1,570 0.9355 0.9366 0.8099
    

    View Slide

38. 
    IUUQTUXJUUFSDPN@NFHBOFDP
    

    View Slide

39. 
    

    View Slide

40. 
    

    View Slide

41. 
    IUUQTUXJUUFSDPNLFJKJ@BSJZBNBTUBUVT
    

    View Slide

42. def distort(image, model_input_size, channels,
    use_random_crop,
    use_random_colorized):
    with tf.name_scope('distort'):
    if use_random_crop:
    with tf.name_scope('random_crop'):
    image_size_ratio = tf.random.uniform([], 100, 120, tf.int32)
    image_size = tf.cast(model_input_size * image_size_ratio / 100, tf.int32)
    image = tf.image.resize(image, (image_size, image_size))
    image = tf.image.random_crop(
    image,
    (model_input_size, model_input_size, channels)
    )
    else:
    image = tf.image.resize(image,
    (model_input_size, model_input_size))
    
    %BUB"VHNFOUBUJPO
    

    View Slide

43. python ./create_dataset.py \
    --base_dir /dataset/source/ \
    --output_tfrecords_dir $TFRECORDS_DIR \
    --output_catalogs_dir $CATALOGS_DIR \
    --image_size 256
    
    σʔληοτੜ੒
    256x256
    TFRecord
    labels 1.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, ...
    masks 1, 0, 1, 0, 0, 1, 0, ...
    

    View Slide

44. σʔληοτੜ੒
    
    Resampling
    

    View Slide

45. python ./create_dataset.py \
    --base_dir /dataset/source/ \
    --output_tfrecords_dir $TFRECORDS_DIR \
    --output_catalogs_dir $CATALOGS_DIR \
    --image_size 512
    
    σʔληοτੜ੒
    512x512
    TFRecord
    labels 1.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, ...
    masks 1, 0, 1, 0, 0, 1, 0, ...
    

    View Slide

46. def distort(image, request_image_size, channels,
    use_random_crop,
    use_random_colorized):
    with tf.name_scope('distort'):
    if use_random_crop:
    with tf.name_scope('random_crop'):
    image_size_ratio = tf.random.uniform([], 100, 120, tf.int32)
    image_size = tf.cast(request_image_size * image_size_ratio / 100, tf.int32)
    image = tf.image.resize(image, (image_size, image_size))
    image = tf.image.random_crop(
    image,
    (request_image_size, request_image_size, channels)
    )
    else:
    image = tf.image.resize(image,
    (request_image_size, request_image_size))
    
    %BUB"VHNFOUBUJPO
    

    View Slide

47. 
    IUUQTUXJUUFSDPNLFJKJ@BSJZBNBTUBUVT
    NFHBOFͷ൑ఆਫ਼౓͕
    

    View Slide

48. ࠓޙͷ՝୊
    5FOTPS'MPX΁ͷຊ֨Ҡߦ
    LFSBTUVOFSͷಋೖʢϋΠύʔύϥϝʔλʔɾαʔνʣ
    (SBE$".ʢ(SBEJFOUXFJHIUFE$MBTT"DUJWBUJPO.BQQJOHʣͷಋೖ
    
    

    View Slide

49. 
    IUUQNFHBOFLLPLZPEBOPSHDPVODJM
    ׬ച͠·ͨ͠
    

    View Slide

50. 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͸ɺ˜ɺšɺäΛׂѪ͍ͯ͠·͢ɻ
    5IF"OESPJESPCPUJTSFQSPEVDFEPSNPEJpFEGSPNXPSLDSFBUFEBOETIBSFECZ(PPHMFBOEVTFEBDDPSEJOHUPUFSNTEFTDSJCFEJOUIF$SFBUJWF$PNNPOT"UUSJCVUJPO-JDFOTF
    

    View Slide