NxでMNISTの手書き数字画像分類を試す / Training MNIST Datasets with Nx

NervesJP #15 Nxを触ってみる回
https://nerves-jp.connpass.com/event/205125/

Transcript

1. ܀ྛ݈ଠ࿠ʢ(.0ϖύϘגࣜձࣾɺ๺཮ઌ୺Պֶٕज़େֶӃେֶʣ
   /FSWFT+1/YΛ৮ͬͯΈΔճʢ೥݄೔ʣ
   NxͰMNISTͷखॻ͖਺ࣈը૾෼ྨΛࢼ͢
   

   View Slide

2. ܀ྛ݈ଠ࿠BLB͋ΜͪΆ
   IUUQTLFOUBSPLVSJCBZBTIJDPN
   ɾ(.0ϖύϘגࣜձࣾऔక໾$50
   ɾҰൠࣾஂ๏ਓ೔ຊ$50ڠձཧࣄ
   ๺཮ઌ୺Պֶٕज़େֶӃେֶʢ+"*45ʣത
   ࢜લظ՝ఔࡏֶதͷࣾձਓֶੜͰ΋͋
   Δɻ
   *P5ؔ࿈ͷݚڀΛ४උ͍ͯ͠Δͱ͜Ζʢ
   ݄ʹ/FSWFT͕ग़ͯ͘Δݚڀใࠂ࿦จʹͭ
   ͍ͯൃද͠·͢ʣɻ
   ࣗݾ঺հ
   2
   

   View Slide

3. /Y /VNFSJDBM&MJYJS
   JTOPXQVCMJDMZBWBJMBCMF%BTICJU#MPH
   IUUQTEBTICJUDPCMPHOYOVNFSJDBMFMJYJSJTOPXQVCMJDMZBWBJMBCMF
   

   View Slide

4. *OUSPEVDJOH/Y+PTÉ7BMJNc-BNCEB%BZT
   IUUQTZPVUVCFG1,.N+Q"(8D
   

   View Slide

5. ಈըΛ؍ͯΔ͚ͩͰ͸Θ͔Βͳ͍ͷͰ
   ϥΠϒίʔσΟϯάΛࣸܦͰ࠶ݱͨ͠
   

   View Slide

6. +OOOY+PTÉ`T/FVSBM/FUXPSLXJUI/Y
   IUUQTHJUIVCDPNLFOUBSPKOOOY
   

   View Slide

7. ͜Μͳײ͡ͰKeras෩ʹࢼͤ·͢
   ./*45σʔλͷಡΈࠐΈ
   [x_train, y_train, x_test, y_test] = Jnnnx.MNIST.Dataset.load_data()
   σʔλͷ੔ܗͱਖ਼نԽ
   x_train =
   x_train
   |> Nx.reshape({60000, 28*28}, names: [:batch, :input])
   |> Nx.divide(255)
   x_test =
   x_test
   |> Nx.reshape({10000, 28*28}, names: [:batch, :input])
   |> Nx.divide(255)
   POFIPUFODPEJOH
   y_train = y_train |> Jnnnx.Utils.to_categorical(10, names: [:batch, :output])
   y_test = y_test |> Jnnnx.Utils.to_categorical(10, names: [:batch, :output])
   τϨʔχϯάσʔλΛ༻ֶ͍ͯश
   params = Jnnnx.fit(x_train, y_train, epoch: 5, batch_size: 50, learning_rate: 0.01)
   ςετσʔλΛ༻͍ͯධՁ
   score = Jnnnx.evaluate(params, x_test, y_test)
   IO.puts("Accuracy: #{Nx.to_scalar(score)}")
   

   View Slide

8. ૉͷ&MJYJS$16ʢ&9-"Λ༻͍ͳ͍ʣͰ࣮ߦͨ݁͠Ռ˞
   ֶश݁ՌʢΤϙοΫ5ɺֶश཰0.01ɺֶशʹཁͨ࣌ؒ͠: ൒೔͙Β͍ʣ
   ˞&9-"$16΋ಈ͔ͯ͠Έ͕ͨɺܻ୆Ͱ଎͘ͳΔʢ࣍ϖʔδʣͱ͸͍͑ݩ͕஗ա͗ΔͷͰಉֶ͡शΛ΋͏Ұ౓΍Δ͜ͱ͸͠ͳ͔ͬͨɻ
   

   View Slide

9. 4PGUNBYؔ਺ͷ࣮ߦํࣜ͝ͱͷϕϯνϚʔΫ݁Ռ
   IUUQTHJUIVCDPNFMJYJSOYOYUSFFNBJOOYOVNFSJDBMEFGJOJUJPOT
   

   View Slide

10. ˔ +PTÉͷϥΠϒίʔσΟϯάಈըΛ؍ͳ͕Βࣸܦͨ͠
    ˔ ػցֶशϥΠϒϥϦͷΑ͏ʹ࢖͑ΔΑ͏ʹ੔ཧͨ͠
    ˓ ࣸܦͨ͠ίʔυΛϥΠϒϥϦͬΆ͍ϑΝΠϧߏ଄Ͱ഑ஔ
    ˓ ϋΠύʔύϥϝλΛؔ਺ͷҾ਺ͱͯ͠౉ͤΔΑ͏ʹͨ͠
    ˔ ./*45ͷσʔληοτΛऔಘ͢ΔϞδϡʔϧΛ௥Ճͨ͠
    ˔ ֶशͨ͠ϞσϧΛɺςετσʔλʹΑͬͯධՁ͢Δؔ਺Λ௥Ճͨ͠
    ˠಈըͰσϞͯͨ͠ίʔυͷݩʹͳ͍ͬͯΔ΋ͷͱࢥΘΕΔ΋ͷ͕
    FYMBͷ΄͏ͷFYBNQMFTʹ͋ͬͨʂ˞
    ΍ͬͨ͜ͱ
    ˞IUUQTHJUIVCDPNFMJYJSOYOYCMPCNBJOFYMBFYBNQMFTNOJTUFYT
    

    View Slide

11. ˔ ݱঢ়Ͱ͸ϨΠϠʔͷߏ੒ɺ׆ੑԽؔ਺ɺଛࣦؔ਺౳ΛܾΊଧͪʹ͠
    ͍ͯΔ͕ɺࣗ༝ʹ૊Έ߹ΘͤΒΕΔΑ͏ʹ͢Δ͜ͱ
    ˓ ͦͷ͋ͨΓ·Ͱ΍Δͱ΋͏গ͠ϥΠϒϥϦͬΆ͘ͳΔ
    ˓ ͍·͸୯ʹॲཧΛͦΕͬΆ͘ݟ͑ΔΑ͏ʹ·ͱΊ͚ͨͩ
    ˔ &9-"Λ࢖ͬͯ(16Ͱܭࢉ͢Δ͜ͱ˞
    ·ͩ΍ͬͯͳ͍͜ͱ
    ˞+FUTPO/BOP(#Ͱࢼ͔͕ͨͬͨ͠ɺϕλϕλ৮ͬͨΓ͍͔ͨͤ͠ىಈ͠ͳ͘ͳͬ
    

    View Slide

12. ͜Ε͔Βඞཁͳ΋ͷ͕੝Γͩ͘͞Μʂ
    IUUQTZPVUVCFG1,.N+Q"(8D U
    

    View Slide

13. ˔ ݱঢ়͸ɺςϯιϧͷܭࢉ΍ࣗಈඍ෼౳ͷɺσΟʔϓϥʔχϯάΛ͢Δ
    ্ͰجຊͱͳΔϏϧσΟϯάϒϩοΫ͕Ͱ͖ͨͱ͜Ζ
    ˔ 5FOTPSGMPX,FSBTɺ1Z5PSDIɺTDJLJUMFBSOͷΑ͏ͳػցֶशϑϨʔ
    ϜϫʔΫ͕͋Δͱ͍͍ͳ͋
    ˠͦΜͳؾ࣋ͪ΋͋ͬͯࠓճɺࡶʹࢼͯ͠ΈͨΓͨ͠ͷͰͨ͠
    ˠʰθϩ͔Β࡞Δ%FFQ-FBSOJOH⁠ʕϑϨʔϜϫʔΫฤʱΛಡΈͳ
    ͕Βࢼ͠ʹ࡞Γ࢝ΊͯΈ͚ͨͲ͏·͍͜ͱઃܭͰ͖ͳ͍ͯͬͨ͘Μ͋
    ͖ΒΊ·ͨ͠ʢؔ਺ܕݴޠ೴ʹͳΓ͖Εͯͳ͍ʜʜʣ
    ࠓޙ΁ͷظ଴
    

    View Slide