C-LIS CO., LTD.

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

C-LIS CO., LTD. IUUQTLOPXMFEHFTBLVSBBEKQ

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

5FOTPS'MPX͸
 ͱͯ΋ྑ͍ϑϨʔϜϫʔΫͰ͢ɻ

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

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

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

IUUQTUXJUUFSDPNLFJKJ@BSJZBNBTUBUVT

IUUQTXXXGBDFCPPLDPNHSPVQTQFSNBMJOL

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

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

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༻

5FOTPS'MPX

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*

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

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

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

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

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

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()

2VBOUJ[BUJPOBXBSF5SBJOJOH

ϞσϧʢύϥϝʔλʔʣͷྔࢠԽ CJUුಈখ਺఺਺ΛCJU੔਺ʹม׵ʢྔࢠԽʣ

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

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

IUUQTUXJUUFSDPNLFJKJ@BSJZBNBTUBUVT

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

ʹ͢ΔͳΒͤΊͯɺ
 Yʹ͋Δػೳ͸࢖͑ΔΑ͏ʹ͍ͯͩ͘͠͞ɻ (PPHMF͞Μ

5FOTPS'MPXͰझຯͷը૾ऩूαʔόʔΛ࡞Δ ೥݄߸

σϞ

֓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

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

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

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

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

IUUQTUXJUUFSDPN@NFHBOFDP

IUUQTUXJUUFSDPNLFJKJ@BSJZBNBTUBUVT

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

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, ...

σʔληοτੜ੒ Resampling

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, ...

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

IUUQTUXJUUFSDPNLFJKJ@BSJZBNBTUBUVT NFHBOFͷ൑ఆਫ਼౓͕

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

IUUQNFHBOFLLPLZPEBOPSHDPVODJM ׬ച͠·ͨ͠

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