20170419-TFUG

4月19日のTensorFlow User Groupの発表資料です。

Transcript

1. C-LIS CO., LTD.
   

   View Slide

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

   View Slide

3. View Slide

4. 
   ݄̐̕೔ʢ೔ʣ ΞΩόɾεΫΤΞ
   ʰٕज़ॻయ̎ ͓ʔ̍̎ʱʹͯ
   TensorFlow͸͡Ί·ͨ͠

   Super Resolution ʔ ௒ղ૾
   https://goo.gl/Z3Exjo
   ׬ച͠·ͨ͠
   

   View Slide

5. ৞ΈࠐΈΦʔτΤϯίʔμʔ
   
   conv1
   9x9x3
   64
   padding:
   SAME
   conv2
   9x9x64
   128
   conv4
   5x5x128
   64
   conv3-0
   7x7x128
   128
   conv3-1
   7x7x128
   128 dconv1
   9x9x64
   3
   padding:
   SAME
   dconv2
   9x9x128
   64
   dconv4
   5x5x64
   128
   dconv3-1
   7x7x128
   128
   dconv3-2
   7x7x128
   128
   conv3-2
   7x7x128
   128
   dconv3-2
   7x7x128
   128
   256x256 256x256
   

   View Slide

6. View Slide

7. 
   5FOTPS'MPX6TFS(SPVQ
   
   5FOTPS'MPX

   (16
   

   View Slide

8. 
   5FOTPS'MPXͱ(16
   

   View Slide

9. લఏ৚݅
   (16$6%"$PNQVUF$BQBCJMJUZ
   $6%"5PPMLJU
   DV%//W
   QJQJOTUBMMUFOTPSqPXHQV
   
   https://www.tensorflow.org/install/install_linux
   

   View Slide

10. (16΁ͷॲཧͷׂΓ౰ͯ
    XJUIUGEFWJDF HQVEJ
    

    QBTT

    
    

    View Slide

11. Ϛϧν(16
    UPXFS@HSBET<>
    GPSJJOSBOHF '-"(4OVN@HQVT
    

    XJUIUGEFWJDF HQVEJ
    

    XJUIUGOBNF@TDPQF T@E 508&3@/".& J
    
    BTTDPQF

    MPTT@@UPXFS@MPTT TDPQF GJMF@MJTU 

    .0%&-*/165@4*;& .0%&-065165@4*;& 

    '-"(4CBUDI@TJ[F
    

    

    UGTVNNBSZTDBMBS UPUBM@MPTT MPTT
    

    

    UGHFU@WBSJBCMF@TDPQF
    SFVTF@WBSJBCMFT
    

    

    HSBETPQUDPNQVUF@HSBEJFOUT MPTT
    

    UPXFS@HSBETBQQFOE HSBET
    
    
    

    View Slide

12. Ϛϧν(16
    BWH@HSBET@@BWFSBHF@HSBEJFOUT UPXFS@HSBET
    

    

    BQQMZ@HSBEJFOU@PQ@@BQQMZ@HSBEJFOUT BWH@HSBET PQU HMPCBM@TUFQ
    

    

    WBSJBCMF@BWFSBHFTUGUSBJO&YQPOFOUJBM.PWJOH"WFSBHF 

    HMPCBM@TUFQ
    

    WBSJBCMFT@BWFSBHFT@PQ
    WBSJBCMF@BWFSBHFTBQQMZ UGUSBJOBCMF@WBSJBCMFT
    
    

    

    USBJO@PQUGHSPVQ BQQMZ@HSBEJFOU@PQ WBSJBCMFT@BWFSBHFT@PQ
    
    
    

    View Slide

13. 5FOTPS#PBSE
    

    View Slide

14. (16̍ͭʹׂΓ౰ͯΔॲཧ
    Input Pipeline
    tf.train.shuffle_batch
    Ϟσϧ

    model.inference
    ޡࠩؔ਺

    __loss
    ޯ഑ͷܭࢉ

    compute_gradient
    __tower_loss
    ޯ഑
    ʢgradsʣ
    ֶश

    σʔλ
    ڭࢣ

    σʔλ
    ਪ࿦݁Ռ ޡࠩ
    

    View Slide

15. શମͷॲཧ
    TOWER
    ޯ഑
    ʢgradsʣ
    TOWER
    ޯ഑
    ʢgradsʣ
    TOWER
    ޯ഑
    ʢgradsʣ
    TOWER
    ޯ഑
    ʢgradsʣ
    ޯ഑ͷฏۉΛܭࢉ

    __average_gradients
    શମͷޯ഑
    ʢavg_gradsʣ
    ࠷దԽ

    __apply_gradients
    params
    params
    params
    params
    

    View Slide

16. 
    %FNP
    

    View Slide

17. ৞ΈࠐΈΦʔτΤϯίʔμʔ
    
    conv2
    9x9x64
    128
    conv4
    5x5x128
    64
    conv3-0
    7x7x128
    128
    conv3-1
    7x7x128
    128
    dconv2
    9x9x128
    64
    dconv4
    5x5x64
    128
    dconv3-1
    7x7x128
    128
    dconv3-2
    7x7x128
    128
    conv3-2
    7x7x128
    128
    dconv3-2
    7x7x128
    128
    256x256 256x256
    conv1
    9x9x3
    64
    padding:
    SAME
    dconv1
    9x9x64
    3
    padding:
    SAME
    

    View Slide

18. ख࣋ͪͷ(16αʔόʔ
    $16*OUFM$PSFJ,
    .FNPSZ(#
    44%(#
    (F'PSDF(59(#

    $6%"DPSF
    
    

    View Slide

19. ͘͞ΒͷߴՐྗίϯϐϡʔςΟϯάʢ࣮ݧػʣ
    $169FPO$PSFʷ
    .FNPSZ(#
    44%(#
    (F'PSDF(595*5"/9ʢ1BTDBMΞʔΩςΫνϟʣ(#ʷ

    $6%"DPSF
    (F'PSDF(595Jʢ1BTDBMΞʔΩςΫνϟʣ(#ʷ

    $6%"DPSF
    
    

    View Slide

20. 
    https://blog.keiji.io/2016/05/cuda_error_no_device.html
    

    View Slide

21. ύλʔϯ"ɿυϥΠόʔͷΞοϓσʔτ
    
    

    View Slide

22. E tensorflow/stream_executor/cuda/cuda_dnn.cc:347] Loaded runtime CuDNN
    library: 4007 (compatibility version 4000) but source was compiled with
    5103 (compatibility version 5100).
    
    ύλʔϯ#ɿ5FOTPS'MPXͷΞοϓσʔτ
    https://github.com/tensorflow/tensorflow/issues/4251
    

    View Slide

23. 
    ύλʔϯ$
    

    View Slide

24. View Slide

25. View Slide

26. 
    खݩͰ࣮ߦՄೳͳϞσϧͱɺ

    ࢴ໘ʹܝࡌՄೳͳϞσϧ͸ҧ͏
    

    View Slide

27. ։ൃػ
    ͘͞ΒͷߴՐྗαʔόʔʢ2VBE(16Ϟσϧʣ
    $169FPO$PSFʷ
    .FNPSZ(#
    44%(#
    (F'PSDF(595*5"/9ʢ.BYXFMMΞʔΩςΫνϟʣ(#ʷ

    $6%"DPSF
    
    128 sample / batch
    learning rate 0.001
    

    View Slide

28. https://www.sakura.ad.jp/press/2017/0417_koukaryoku_hourly/
    

    View Slide

29. ݕূػ

    %FMM"-*&/8"3&93
    $16*OUFM$PSFJ,
    .FNPSZ(#
    44%(#
    (F'PSDF(59(#

    $6%"DPSF
    
    128 sample / batch
    learning rate 0.001
    

    View Slide

30. ஗͍ʜʜ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    
    128 sample / batch
    learning rate 0.001
    

    View Slide

31. (PPHMF$MPVE1MBUGPSN
    W$16
    .FNPSZ
    44%
    (F'PSDF5FTMB,ʢ(#ʷʣ

    $6%"DPSF ʷ
    
    

    View Slide

32. 
    

    View Slide

33. 
    

    View Slide

34. 
    ( Д ) ʄ ʄŴƅƃŕ
    

    View Slide

35. ௨ৗ͸ΫΥʔλͰ੍ݶ͞Ε͍ͯΔ
    
    

    View Slide

36. 
    :
    :
    

    View Slide

37. 
    Hello,
    Thank you for requesting additional quota.
    To help us ensure that this is a legitimate request and that the resources you requested are available to you,
    please do one of the following:
    Make a payment [1] of $35 or the same amount in your currency from the Transaction History page [2] and
    reply to this message when the charge clears. Your payment will be applied to any charges you incur in the
    future and will be visible as a credit in your account.
    Alternatively, reply to this message with the project ID of another project that you own that has cleared a
    charge of at least the amount mentioned.
    To learn more about project quota requests and cores quota requests, check out the Project Quota Request
    FAQ page at [3] and the Request cores quota increase FAQs page at [4].
    We appreciate your patience.
    Sincerely,
    Cloud Platform Support
    [1] - https://support.google.com/cloud/answer/6294016 [2] - https://console.cloud.google.com/billing [3] -
    https://support.google.com/cloud/answer/6330231 [4] - https://support.google.com/cloud/answer/6376374
    

    View Slide

38. 
    Hello,
    Your quota request for Project '632906615582' has been approved and your project quota has been adjusted
    accordingly.
    Changed Quota:
    +--------------------+-----------------+
    | Region: asia-east1 | NVIDIA_K80_GPUS |
    +--------------------+-----------------+
    | Changes | 0 -> 1 |
    +--------------------+-----------------+
    To verify the quota change, please navigate to
    https://console.developers.google.com/project/632906615582/compute/quotas and verify what are your
    current project quotas.
    Happy Computing!
    Google for Work Support
    http://support.google.com/enterprisehelp/
    

    View Slide

39. (PPHMF$MPVE1MBUGPSN
    W$16
    .FNPSZ(#
    44%(#
    (F'PSDF5FTMB,ʢ(#ʷʣ

    $6%"DPSF ʷ
    
    ͋·ΓมΘΒͣ……
    

    View Slide

40. 
    ݪҼ͸ʁ
    

    View Slide

41. OWJEJBTNJ
    
    ]/7*%*"4.*%SJWFS7FSTJPO]
    ]
    ](16/BNF1FSTJTUFODF.]#VT*E%JTQ"]7PMBUJMF6ODPSS&$$]
    ]'BO5FNQ1FSG1XS6TBHF$BQ].FNPSZ6TBHF](166UJM$PNQVUF.]
    ]]
    ](F'PSDF(590GG]0O]/"]
    ]$188].J#.J#]%FGBVMU]
    
    
    
    ]1SPDFTTFT(16.FNPSZ]
    ](161*%5ZQF1SPDFTTOBNF6TBHF]
    ]]
    ](VTSMJCYPSH9PSH.J#]
    ]$QZUIPO.J#]
    
    
    

    View Slide

42. ˈUPQ
    
    

    View Slide

43. 
    (16͕ಈ͍ͯͳ͍
    

    View Slide

44. DPOGJHUG$POGJH1SPUP 

    BMMPX@TPGU@QMBDFNFOU5SVF 

    MPH@EFWJDF@QMBDFNFOU'-"(4MPH@EFWJDF@QMBDFNFOU
    

    XJUIUG4FTTJPO DPOGJHDPOGJH
    BTTFTT

    DIFDLQPJOUUGUSBJOHFU@DIFDLQPJOU@TUBUF USBJO@EJS
    

    
    MPH@EFWJDF@QMBDFNFOU
    

    View Slide

45. I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2_26: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    save/RestoreV2_25: (RestoreV2): /job:localhost/replica:0/task:0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2_25: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    save/RestoreV2_24: (RestoreV2): /job:localhost/replica:0/task:0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2_24: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    save/RestoreV2_23: (RestoreV2): /job:localhost/replica:0/task:0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2_23: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    save/RestoreV2_22: (RestoreV2): /job:localhost/replica:0/task:0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2_22: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    save/RestoreV2_21: (RestoreV2): /job:localhost/replica:0/task:0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2_21: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    save/RestoreV2_20: (RestoreV2): /job:localhost/replica:0/task:0/cpu:0
    
    

    View Slide

46. I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2_8: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    save/RestoreV2_7: (RestoreV2): /job:localhost/replica:0/task:0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2_7: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    save/RestoreV2_6: (RestoreV2): /job:localhost/replica:0/task:0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2_6: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    save/RestoreV2_5: (RestoreV2): /job:localhost/replica:0/task:0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2_5: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    save/RestoreV2_4: (RestoreV2): /job:localhost/replica:0/task:0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2_4: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    save/RestoreV2_3: (RestoreV2): /job:localhost/replica:0/task:0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2_3: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    save/RestoreV2_2: (RestoreV2): /job:localhost/replica:0/task:0/cpu:0
    
    

    View Slide

47. I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2_2: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    save/RestoreV2_1: (RestoreV2): /job:localhost/replica:0/task:0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2_1: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    save/RestoreV2: (RestoreV2): /job:localhost/replica:0/task:0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] save/RestoreV2: (RestoreV2)/
    job:localhost/replica:0/task:0/cpu:0
    conv3/biases/ExponentialMovingAverage: (VariableV2): /job:localhost/replica:0/task:0/
    cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] conv3/biases/
    ExponentialMovingAverage: (VariableV2)/job:localhost/replica:0/task:0/cpu:0
    save/Assign_22: (Assign): /job:localhost/replica:0/task:0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] save/Assign_22: (Assign)/
    job:localhost/replica:0/task:0/cpu:0
    conv3/biases/ExponentialMovingAverage/read: (Identity): /job:localhost/replica:0/task:
    0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] conv3/biases/
    
    ҎԼུ
    

    View Slide

48. ΦϖϨʔγϣϯ͕࣮ߦ͞ΕΔσόΠε
    (16
    $16
    
    

    View Slide

49. I tensorflow/core/common_runtime/simple_placer.cc:841] tower_0/image_loader_vvv/
    ReaderReadV2: (ReaderReadV2)/job:localhost/replica:0/task:0/cpu:0
    tower_0/image_loader_vvv/DecodeJpeg: (DecodeJpeg): /job:localhost/replica:0/task:
    0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] tower_0/image_loader_vvv/
    DecodeJpeg: (DecodeJpeg)/job:localhost/replica:0/task:0/cpu:0
    tower_0/image_loader_vvv/random_crop/Shape: (Shape): /job:localhost/replica:0/
    task:0/gpu:0
    
    EFDPEF@KQFH
    

    View Slide

50. 
    

    View Slide

51. 
    

    View Slide

52. 
    

    View Slide

53. 
    

    View Slide

54. I tensorflow/core/common_runtime/simple_placer.cc:841] tower_0/image_loader_vvv/
    ReaderReadV2: (ReaderReadV2)/job:localhost/replica:0/task:0/cpu:0
    tower_0/image_loader_vvv/DecodeJpeg: (DecodeJpeg): /job:localhost/replica:0/task:
    0/cpu:0
    I tensorflow/core/common_runtime/simple_placer.cc:841] tower_0/image_loader_vvv/
    DecodeJpeg: (DecodeJpeg)/job:localhost/replica:0/task:0/cpu:0
    tower_0/image_loader_vvv/random_crop/Shape: (Shape): /job:localhost/replica:0/
    task:0/gpu:0
    
    EFDPEF@KQFH
    

    View Slide

55. DPOGJHUG$POGJH1SPUP 

    BMMPX@TPGU@QMBDFNFOU5SVF 

    MPH@EFWJDF@QMBDFNFOU'-"(4MPH@EFWJDF@QMBDFNFOU
    

    XJUIUG4FTTJPO DPOGJHDPOGJH
    BTTFTT

    DIFDLQPJOUUGUSBJOHFU@DIFDLQPJOU@TUBUF USBJO@EJS
    

    
    BMMPX@TPGU@QMBDFNFOU
    

    View Slide

56. BMMPX@TPGU@QMBDFNFOU
    
    If you would like TensorFlow to automatically choose an existing and
    supported device to run the operations in case the specified one doesn't exist,
    you can set allow_soft_placement to True in the configuration option
    when creating the session.
    https://www.tensorflow.org/tutorials/using_gpu
    

    View Slide

57. EFG@MPBE@JNBHF GJMF@MJTU JOQVU@TJ[F PVUQVU@TJ[F DIBOOFMT TDBMF
    

    XJUIUGOBNF@TDPQF JNBHF@MPBEFS@WWW
    

    GJMFOBNF@RVFVFUGUSBJOTUSJOH@JOQVU@QSPEVDFS GJMF@MJTU 

    TIVGGMF5SVF
    

    SFBEFSUG8IPMF'JMF3FBEFS
    

    @ WBMVFSFBEFSSFBE GJMFOBNF@RVFVF
    

    

    JNBHFUGJNBHFEFDPEF@KQFH WBMVF DIBOOFMTDIBOOFMT
    

    DSPQQFE@JNBHFUGSBOEPN@DSPQ JNBHF JOQVU@TJ[F JOQVU@TJ[F DIBOOFMT
    
    

    

    MS@JNBHFUGJNBHFSFTJ[F@JNBHFT DSPQQFE@JNBHF JOQVU@TJ[FTDBMF 

    JOQVU@TJ[FTDBMF
    

    NFUIPEUGJNBHF3FTJ[F.FUIPE#*$6#*$
    

    

    MS@JNBHFUGJNBHFSFTJ[F@JNBHFT MS@JNBHF JOQVU@TJ[F JOQVU@TJ[F
    

    NFUIPEUGJNBHF3FTJ[F.FUIPE#*$6#*$
    

    

    PGGTFU@MFGU JOQVU@TJ[FPVUQVU@TJ[F
    

    PGGTFU@UPQ JOQVU@TJ[FPVUQVU@TJ[F
    

    HSPVOE@USVUIUGJNBHFDSPQ@UP@CPVOEJOH@CPY DSPQQFE@JNBHF 

    PGGTFU@UPQ PGGTFU@MFGU 

    PVUQVU@TJ[F PVUQVU@TJ[F
    

    

    SFUVSOMS@JNBHF HSPVOE@USVUI
    
    

    View Slide

58. 
    510x510
    decode_jpeg
    random_crop
    

    View Slide

59. 
    όονʹ͖ͭɺ

    ϑΝΠϧͷಡΈࠐΈɾσίʔυ͕

    ճൃੜ͢Δ
    

    View Slide

60. EFG@MPBE@JNBHF GJMF@MJTU JOQVU@TJ[F PVUQVU@TJ[F DIBOOFMT TDBMF CBUDI@TJ[F
    

    XJUIUGOBNF@TDPQF JNBHF@MPBEFS@WWW
    

    GJMFOBNF@RVFVFUGUSBJOTUSJOH@JOQVU@QSPEVDFS GJMF@MJTU 

    TIVGGMF5SVF
    

    SFBEFSUG8IPMF'JMF3FBEFS
    

    @ WBMVFSFBEFSSFBE GJMFOBNF@RVFVF
    

    JNBHFUGJNBHFEFDPEF@KQFH WBMVF DIBOOFMTDIBOOFMT
    

    
    

    View Slide

61. MS@JNBHFT<>

    HSPVOE@USVUIT<>

    

    GPSJJOSBOHF CBUDI@TJ[F
    

    DSPQQFE@JNBHFUGSBOEPN@DSPQ JNBHF JOQVU@TJ[F JOQVU@TJ[F DIBOOFMT
    
    

    MS@JNBHFUGJNBHFSFTJ[F@JNBHFT DSPQQFE@JNBHF JOQVU@TJ[FTDBMF 

    JOQVU@TJ[FTDBMF
    

    NFUIPEUGJNBHF3FTJ[F.FUIPE#*$6#*$
    

    MS@JNBHFUGJNBHFSFTJ[F@JNBHFT MS@JNBHF JOQVU@TJ[F JOQVU@TJ[F
    

    NFUIPEUGJNBHF3FTJ[F.FUIPE#*$6#*$
    

    MS@JNBHFUGFYQBOE@EJNT MS@JNBHF EJN
    

    MS@JNBHFTBQQFOE MS@JNBHF
    

    

    PGGTFU@MFGU JOQVU@TJ[FPVUQVU@TJ[F
    

    PGGTFU@UPQ JOQVU@TJ[FPVUQVU@TJ[F
    

    HSPVOE@USVUIUGJNBHFDSPQ@UP@CPVOEJOH@CPY DSPQQFE@JNBHF 

    PGGTFU@UPQ PGGTFU@MFGU 

    PVUQVU@TJ[F PVUQVU@TJ[F
    

    HSPVOE@USVUIUGFYQBOE@EJNT HSPVOE@USVUI EJN
    

    HSPVOE@USVUITBQQFOE HSPVOE@USVUI
    

    

    MS@JNBHFTUGDPODBU MS@JNBHFT BYJT
    

    HSPVOE@USVUITUGDPODBU HSPVOE@USVUIT BYJT
    

    

    SFUVSOMS@JNBHFT HSPVOE@USVUIT

    
    

    View Slide

62. 
    510x510
    decode_jpeg
    random_crop
    

    View Slide

63. MPBE@JNBHF
    
    [batch_size, height, width, channels]
    [height, width, channels]
    

    View Slide

64. MS@JNBHF@CBUDI HSPVOE@USVUI@CBUDIUGUSBJOTIVGGMF@CBUDI 

    

    CBUDI@TJ[F'-"(4CBUDI@TJ[F 

    DBQBDJUZDBQBDJUZ 

    FORVFVF@NBOZ5SVF 

    NJO@BGUFS@EFRVFVF'-"(4NJO@BGUFS@EFRVFVF 

    OVN@UISFBET'-"(4OVN@UISFBET
    

    
    UGUSBJOTIV⒐F@CBUDI
    

    View Slide

65. 
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    4UFQ -PTT 5JNFNTTUFQ
    128 sample / batch
    learning rate 0.001
    

    View Slide

66. 
    ύλʔϯ$ɿ$16ͷΦʔόʔϔου
    

    View Slide

67. ·ͱΊ
    (16͸଎͍ɻίʔυ͕গ͠൥ࡶʹͳΔ͚ͲɺಘΒΕΔ଎͞͸ັྗ
    ͘͞ΒͷߴՐྗ͸շదɻଟগɺඇޮ཰తͳίʔυΛॻ͍ͯ΋ؾ͔ͮͳ͍͘Β͍
    ($&ͷ(16Πϯελϯε΋଎͍ɻW$16ͷ਺ͱ͔ϝϞϦ༰ྔͱ͔ɺ՝୊ʹର͢
    Δ࠷దεϖοΫΛಋ͘ϊ΢ϋ΢͕ཉ͍͠
    .-&OHJOFʜʜʁ
    
    

    View Slide

68. Impress R&DࣾΑΓ

    Print On Demand ˍ ిࢠॻ੶Ͱ
    ݄̑Լ० ൃചܾఆ
    ݄̐̕೔ʢ೔ʣ ΞΩόɾεΫΤΞ
    ʰٕज़ॻయ̎ ͓ʔ̍̎ʱʹͯ
    TensorFlow͸͡Ί·ͨ͠

    Super Resolution ʔ ௒ղ૾
    ׬ച͠·ͨ͠
    

    View Slide

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

    View Slide