少ないデータで画像分類の性能を上げるには

CNNを使った画像分類において少ないデータで過学習を抑制し性能を上げる方法を紹介

Transcript

1. 
   "    "      PYTHON'&IN# RESTART#5 2019/2/23 !$%
2.  
   GOOGLE COLAB   
3. 60? • 2C0?%"' A@/"!3>.6 - "! ,8   • %"' A@/"! =B5:< D.   :< • %"'#&$!49A@/"! 7 )(*+  • cf. A@0?1;
4.  /% -* " ) 
     " ' , #+ -* &$# 
     &$.# (+ !
5. )#/ & • 
    (  • 
   .-  • 4" 
    !2 • %0" +  , %*", L1/L2%0" •
   13$ ,' https://towardsdatascience.com/deep-learning-3-more-on-cnns-handling-overfitting-2bd5d99abe5d
6.   • 6"(!  +
   1-37. $) • Keras Dropout:% *4 • 2/, 
   9508 &'#  https://keras.io/layers/core/#dropout
7. "BATCH NORMALIZATION • &  #"% ' • Keras keras.layers.BatchNormalization()  • Google! $ 
8. L1/L2 /7* • 64(0"!$ +,  #1  "!"8-&) • 2/7*kernel_regularizer • 3.*/7*activity_regularizer • Keras'5 /7*(0 • L1/7*keras.regularizers.l1(p) • L2/7*keras.regularizers.l2(p • 9%Elastic Netkeras.regularizers.l1_l2(l1=p1, l2=p2) https://keras.io/regularizers/
9. 
    GOOGLE COLAB   
10. @> • 15 04 #%' • 9B3* $' • KerasImageDataGenerator2,)!CA-&<8 • 0/F6%.  width_shift_range, height_shift_range • D(+=brightness_range
    E%:?shear_range • zoom_range
     "channel_shift_range • 0/F6%.7;horizontal_flip, vertical_flip https://keras.io/preprocessing/image/#imagedatagenerator-class
11.  
     GOOGLE COLAB 
12. 2?5%!$- • #!$E@0 )#," *& +(6!$ • /=+',9 :B;A ImageNethttp://www.image- net.org/ ILSVRC2012."$:B1203D1000+ E@0  7 • KerasXceptionVGG16VGG19ResNet50InceptionV3 InceptionResNetV2MobileNetDenseNetNASNetMobileNetV2 C >41<8 https://keras.io/applications/
13. 8H; %:6-4 • CLO' • 8H; %(1P>KA
    ,3CLGO'  "$9?  • '3 CLGB %,350* 8H) • DF8H • =I/ 8H; %2  %<> MJ ) • !&#& • DF8H.
    8H; %'3@1+50* N7E 
14.  
    
15. LOQ+)8; • !A%.( ON29 +)6B   • IJ@>2
    29R51 G  +)
    : ?C =D  • HIJ@29R5 1% 
    7E@=D : ?C99%  • 4<3 +)KF $,'0-/0&#0*'0 -/0&OQ+) 
    4<M • LOQ+)>P"