5. 7A 3 P 9 2 % 1 I I 21 % 5 . 9 9 0 2 % 21 26 8
8. u - H T s P 3 > S bQ - H> S 3 b
10. I 6 I 3 S I D K I I I I I A 0
12. P P • • 6 • 6 • 60 A CP • 6 • 6 • 6 • 60 N9 D , D + 3 • P P 4 1L f , + R e
13. API IS NIS • • R P • L C • • U IS C IS - F -
15. Explore New Features Feature Refinement (Reuse Features)
16. ResNeXt-101 (64x4d) 320 MB 15.5 4.45 4.4 DPN-98 Model Size 236 MB - 26% GFLOPs 11.7 - 25% - 8% 12.1 GB GPU Memory 11.1 GB 20.4 / 5.3 Top 1 / Top 5 Error on x224 20.2 / 5.2 Single model, Single center-crop, Top-5 val error rate on x299 error rate 4.35 4.25% 4.3 4.25 4.16% 4.2 4.15 DPN-98 is Fast!! Training = 86 img/sec 4.1 (per node, 4 x K80 cards) 4.05 4 ResNeXt-101 (64x4d) Very Deep PolyNet - DPN-98
17. a it • u • N k l a N C • a • N /s a T M u -
18. - - - / • • •D •D •N C 5
21. OU OU + • U • Q B T XC 4 • • O • • • D NA R e HA D U U
22. 1 . 1 3 1 / - . 2 1- , , , , , , , , , , , , 1 2 1 / 1 1 2 2 ( )2 2 -2 , 2. 2 1 3 2 , 2 1 2 ACDB
23. /) (- -- /3 /1 . /1 . 2 3-3 12 + 2 /1 . /1 . 2 3-3 12 /1 . + 2
25. . *- - - . 3 . *- ) - ) - . -) . *- ) -3 ) 3. ( - - -) ) - ) 3 ) - . -. ) . *- - ) - ) *
28. G U 3 3 0S H K L , ( ) 16 B 5A 0+ A 5 C , B :<8 P A8 ./D
33. • • • • • • • • • • • • • • • - • •
