pixiv App Night 20190611

6月11日に開催された「pixiv App Night」の発表資料です。

「わたしは機械学習プロジェクトで
技術的負債を抱えました」

Transcript

1. C-LIS CO., LTD.
   

   View Slide

2. QJYJW"QQ/JHIU
   
   Θͨ͠͸ػցֶशϓϩδΣΫτͰ

   ٕज़తෛ࠴Λ๊͑·ͨ͠
   5FOTPS'MPXͰझຯͷը૾ऩूαʔόʔΛ࡞Δɹ೥݄߸
   

   View Slide

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

   View Slide

4. 5FOTPS'MPXൃදʢ೥݄ʣ
   
   
   

   View Slide

5. ػցֶशΛར༻ͨ͋͠Ε͜Ε
   
   
   IUUQLJWBOUJVNIBUFCMPKQFOUSZ
   TensorFlowͰΞχϝΏΔΏΓͷ੍࡞ձࣾΛࣝผ͢Δ
   IUUQCPIFNJBIBUFOBCMPHDPNFOUSZ
   σΟʔϓϥʔχϯάͰ͓ͦদ͞Μͷ࿡ͭࢠ͸ݟ෼͚ΒΕΔͷ͔ʁ
   IUUQCPIFNJBIBUFOBCMPHDPNFOUSZ
   IUUQDISJTUJOBIBUFOBCMPHDPNFOUSZ
   Deep LearningͰϥϒϥΠϒʂΩϟϥΛࣝผ͢Δ
   

   View Slide

6. 
   
   ؟ڸ່ͬͷΠϥετΛ

   Πϯλʔωοτ͔ΒࣗಈͰऩू͍ͨ͠ʂ
   

   View Slide

7. ؟ڸ່ͬ൑ఆ
   
   
   

   View Slide

8. ؟
   ڸ
   ͬ
   ່
   ˜ࠜઇΕ͍
   

   View Slide

9. 
   
   

   View Slide

10. 
    
    σϞ
    

    View Slide

11. 
    
    ධՁ༻αʔόʔ
    ܇࿅ɾֶश༻αʔόʔ܈
    σʔληοτసૹ
    ʢTFRecordʣ ֶशࡁϞσϧऔಘ
    ը૾औಘ
    ը૾औಘ
    ϥϕϧ
    ෇͚
    σʔληοτ؅ཧ
    αʔόʔ
    ը૾ऩूݩ
    ը૾ऩू
    ϥϕϧ
    ෇͚
    σʔληοτ

    ؅ཧΞϓϦ
    playground.megane.ai
    ֶशࡁΈϞσϧ഑ஔ
    ը૾ૹ৴
    ൑ఆ݁Ռ
    

    View Slide

12. ϥϕϧͷछྨ
    PSJHJOBM@BSU
    OTGX
    GBWPSJUF
    QIPUP
    JMMVTU
    DPNJD
    
    
    GBDF
    GFNBMF
    NFHBOF
    TDISPPM@VOJGPSN
    CMB[FS@VOJGPSN
    TBJMPS@VOJGPSN
    HM
    LFNPOP
    NBMF
    CM
    DBU
    EPH
    GPPE
    EJTMJLF
    

    View Slide

13. https://twitter.com/_meganeco
    .FHBOF$P1MBZHSPVOE
    
    
    https://playground.megane.ai/
    

    View Slide

14. 
    
    ̏೥ؒͷࣦഊ
    

    View Slide

15. ֶशϓϩηεʹOPIVQΛ͚ͭ๨ΕΔ
    
    
    $ nohup python train.py ...
    

    View Slide

16. ϗʔϜσΟϨΫτϦΛਧ͖ඈ͹͢
    
    
    $ rsync -av --delete ./scripts/ meganeco2.c-lis.net:~/
    

    View Slide

17. αʔόʔηΩϡϦςΟΛమนʹ
    
    
    $ iptables -F
    IUUQXXXJOWJUSPKQCMPHJOEFYDHJ-JOVY@IUN
    

    View Slide

18. def download(self, url, output_dir):
    response = requests.get(url)
    response.raise_for_status()
    fd, file_path = tempfile.mkstemp(dir=output_dir)
    os.close(fd)
    with open(file_path, mode='wb') as fp:
    fp.write(response.content)
    response.close()
    return file_path
    
    
    ϑΝΠϧσΟεΫϦϓλΛރׇͤ͞Δ
    IUUQTOJTIJNVSBIBUFOBEJBSZPSHFOUSZ
    

    View Slide

19. 
    
    ੵΈ্͛ΒΕٕͨज़తෛ࠴ͷ਺ʑ
    

    View Slide

20. ٕज़తෛ࠴
    ϞμϯͰͳ͍"OESPJEΞϓϦͷίʔυ
    5FOTPS'MPXରԠ
    ֶशͱධՁϓϩηεͷ෼཭
    ;͘Ε͕͋Δσʔληοτ
    
    
    

    View Slide

21. 
    
    ධՁ༻αʔόʔ
    ܇࿅ɾֶश༻αʔόʔ܈
    σʔληοτసૹ
    ʢTFRecordʣ ֶशࡁϞσϧऔಘ
    ը૾औಘ
    ը૾औಘ
    ϥϕϧ
    ෇͚
    σʔληοτ؅ཧ
    αʔόʔ
    ը૾ऩूݩ
    ը૾ऩू
    ϥϕϧ
    ෇͚
    σʔληοτ

    ؅ཧΞϓϦ
    playground.megane.ai
    ֶशࡁΈϞσϧ഑ஔ
    ը૾ૹ৴
    ൑ఆ݁Ռ
    

    View Slide

22. ϞμϯͰͳ͍"OESPJEΞϓϦͷίʔυʢٕज़ʣ
    ,PUMJO
    3Y,PUMJO
    3PPN
    #VUUFS,OJGF
    5FOTPS'MPX-JUF
    
    
    $PSPVUJOFT
    "OESPJE"SDIJUFDUVSF$PNQPOFOUT
    /FVSBM/FUXPSLT"1*
    

    View Slide

23. 
    
    5FOTPS'MPXରԠ
    

    View Slide

24. WARNING:tensorflow:From /usr/local/lib/python3.5/dist-packages/tensorflow/python/
    framework/op_def_library.py:263: colocate_with (from tensorflow.python.framework.ops)
    is deprecated and will be removed in a future version.
    Instructions for updating:
    Colocations handled automatically by placer.
    WARNING:tensorflow:From /usr/local/lib/python3.5/dist-packages/tensorflow/python/
    framework/op_def_library.py:263: colocate_with (from tensorflow.python.framework.ops)
    is deprecated and will be removed in a future version.
    Instructions for updating:
    Colocations handled automatically by placer.
    WARNING:tensorflow:From model_res5.py:28: conv2d (from
    tensorflow.python.layers.convolutional) is deprecated and will be removed in a future
    version.
    Instructions for updating:
    Use keras.layers.conv2d instead.
    
    
    

    View Slide

25. 
    
    def layers(tag_name, image, keep_prob=0.5, training=False):
    scope = '%s/%s' % (tag_name, NAME)
    with tf.variable_scope(scope):
    conv = tf.layers.conv2d(image, BASE_CHANNEL, [1, 1], [1, 1],
    padding='SAME',
    activation=tf.nn.relu,
    use_bias=True,
    trainable=training,
    name='conv_top')
    def layers(tag_name, image, rate=0.5, training=False):
    scope = '%s/%s' % (tag_name, NAME)
    with tf.variable_scope(scope):
    conv = tf.keras.layers.Conv2D(BASE_CHANNEL, [1, 1], [1, 1],
    padding='SAME',
    activation=tf.nn.relu,
    use_bias=True,
    name='conv_top')(image)
    

    View Slide

26. 
    
    IUUQTUXJUUFSDPNLFJKJ@BSJZBNBTUBUVT
    

    View Slide

27. 
    
    ֶशͱධՁϓϩηεͷ෼཭
    

    View Slide

28. 
    
    ධՁ༻αʔόʔ
    ܇࿅ɾֶश༻αʔόʔ܈
    σʔληοτసૹ
    ʢTFRecordʣ ֶशࡁϞσϧऔಘ
    ը૾औಘ
    ը૾औಘ
    ϥϕϧ
    ෇͚
    σʔληοτ؅ཧ
    αʔόʔ
    ը૾ऩूݩ
    ը૾ऩू
    ϥϕϧ
    ෇͚
    σʔληοτ

    ؅ཧΞϓϦ
    playground.megane.ai
    ֶशࡁΈϞσϧ഑ஔ
    ը૾ૹ৴
    ൑ఆ݁Ռ
    

    View Slide

29. 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

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

    View Slide

31. 
    
    

    View Slide

32. ̎ͭͷϓϩηεʢֶशͱධՁʣ
    
    
    train process
    eval process
    ֶशࡁΈύϥϝʔλʔ
    ֶशσʔληοτ
    ධՁσʔληοτ
    ࠷ྑύϥϝʔλʔ
    ߋ৽
    ධՁ
    มߋΛ؂ࢹ
    ߋ৽
    

    View Slide

33. - name: train-single-rocm
    container:
    image: rocm/tensorflow:rocm2.2-tf1.13-python3
    command: [/workdir/scripts/train_single.sh]
    env:
    - name: DATASET_DIR
    value: "/dataset/tfrecords_classifier_latest"
    - name: CATALOGS_DIR
    value: "{{inputs.parameters.catalogs-dir}}"
    - name: TRAIN_DIR
    value: "data/{{inputs.parameters.train-dir}}"
    - name: SUMMARY_DIR
    value: "data/{{inputs.parameters.summary-dir}}"
    - name: CONFIG_FILE
    value: "scripts/{{inputs.parameters.config-file}}"
    - name: BATCH_SIZE
    value: "{{inputs.parameters.batch-size}}"
    - name: LERANING_RATE
    value: "{{inputs.parameters.learning-rate}}"
    - name: STEPS
    value: "{{inputs.parameters.steps}}"
    resources:
    limits:
    amd.com/gpu: 2
    requests:
    amd.com/gpu: 2
    
    
    - name: train-single-cuda
    container:
    image: tensorflow/tensorflow:1.13.1-gpu-py3
    command: [/workdir/scripts/train_single.sh]
    env:
    - name: DATASET_DIR
    value: "/dataset/tfrecords_classifier_latest"
    - name: CATALOGS_DIR
    value: "{{inputs.parameters.catalogs-dir}}"
    - name: TRAIN_DIR
    value: "data/{{inputs.parameters.train-dir}}"
    - name: SUMMARY_DIR
    value: "data/{{inputs.parameters.summary-dir}}"
    - name: CONFIG_FILE
    value: "scripts/{{inputs.parameters.config-file}}"
    - name: BATCH_SIZE
    value: "{{inputs.parameters.batch-size}}"
    - name: LERANING_RATE
    value: "{{inputs.parameters.learning-rate}}"
    - name: STEPS
    value: "{{inputs.parameters.steps}}"
    resources:
    limits:
    nvidia.com/gpu: 2
    requests:
    nvidia.com/gpu: 2
    

    View Slide

34. #!/bin/sh
    HIP_VISIBLE_DEVICES=0 CUDA_VISIBLE_DEVICES=0 \
    python3 sources/server/picture_single_discriminator/eval.py \
    --learning_config $CONFIG_FILE \
    --tfrecords_dir $DATASET_DIR \
    --catalogs_dir $CATALOGS_DIR \
    --train_dir $TRAIN_DIR \
    --summary_dir $SUMMARY_DIR &
    HIP_VISIBLE_DEVICES=1 CUDA_VISIBLE_DEVICES=1 \
    python3 sources/server/picture_single_discriminator/train.py \
    --learning_config $CONFIG_FILE \
    --tfrecords_dir $DATASET_DIR \
    --catalogs_dir $CATALOGS_DIR \
    --train_dir $TRAIN_DIR \
    --summary_dir $SUMMARY_DIR \
    --batch_size $BATCH_SIZE \
    --learning_rate $LERANING_RATE \
    --num_gpus 1 \
    --step $STEPS
    
    
    

    View Slide

35. ධՁϓϩηεΛഇࢭ
    
    
    train process
    eval process
    ֶशࡁΈύϥϝʔλʔ
    ֶशσʔληοτ
    ධՁσʔληοτ
    ࠷ྑύϥϝʔλʔ
    ߋ৽
    ධՁ
    มߋΛ؂ࢹ
    ߋ৽
    ×
    

    View Slide

36. ֶशɾධՁϓϩηε
    
    
    train & eval process
    ֶशσʔληοτ
    ධՁσʔληοτ
    ࠷ྑύϥϝʔλʔ
    ߋ৽
    

    View Slide

37. 
    
    ;͘Ε͕͋Δσʔληοτ
    

    View Slide

38. 
    
    ධՁ༻αʔόʔ
    ܇࿅ɾֶश༻αʔόʔ܈
    σʔληοτసૹ
    ʢTFRecordʣ ֶशࡁϞσϧऔಘ
    ը૾औಘ
    ը૾औಘ
    ϥϕϧ
    ෇͚
    σʔληοτ؅ཧ
    αʔόʔ
    ը૾ऩूݩ
    ը૾ऩू
    ϥϕϧ
    ෇͚
    σʔληοτ

    ؅ཧΞϓϦ
    playground.megane.ai
    ֶशࡁΈϞσϧ഑ஔ
    ը૾ૹ৴
    ൑ఆ݁Ռ
    

    View Slide

39. σʔληοτ؅ཧαʔόʔ
    ໿ສຕͷը૾ϑΝΠϧʢ໿5#ʣ
    ໿ສ݅ͷϥϕϧ
    όοΫΞοϓମ੍ͷ੔උ
    
    
    

    View Slide

40. 
    
    

    View Slide

41. ٕज़తෛ࠴Λͳͥղܾ͠ͳ͍ͷ͔
    
    
    ͍ͭ·Ͱ΍ΔʁɹͲ͜·Ͱ΍Δʁɹকདྷͷల๬͸ʁ
    ҆ఆͯ͠Քಇ͍ͯ͠ΔγεςϜΛյͨ͘͠ͳ͍

    ʢJGJUXPSLT EPOUpYJUʣ
    όοΫΞοϓϓϥϯΛ༻ҙͰ͖ͳ͍ʢܦࡁ߹ཧੑʣ
    

    View Slide

42. 
    
    ͍ͭ·Ͱ΍ΔʁɹͲ͜·Ͱ΍Δʁɹকདྷͷల๬͸ʁ
    

    View Slide

43. 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