[DL Hacks]Pruning Convolutional Neural Networks for Resource Efficient Inference

1 1 7 E 27 1 1 CI LMP N R

z y • L : LDI N – – sp – SV KR G = R T D 4GGJDJ NS 6NG ND - ? 8 R ?JM 1JL 7 N 8 TSX DJS E C , 7TN DJS E C ISSPR. ou – S PI N ? AJ tk 629= – • T L 6 61 v – • LTSJ N L n • • TNJN 2 N ISSP. D N 2 NG P N J VN S G ND RS - TM0JE/ 752JV) L 6 61 5?2 N E M NE PTS DID NG D M LTSJ N L N T L N SV gifhla ub SD P R NS SJ N R (( P L M LDI N P TNJN KR PEG a em a c bougifh dwr 2

• https://openreview.net/forum?id=SJGCiw5gl
• http://on-demand.gputechconf.com/gtc/2017/presentation/s7442-pavlo-molchanov-pruning-convolutional-neural-networks.pdf

mjp a bc • d cr – – i o e • o : g y : i r G e u :i n C s o – – • .- hl n – – - U P ./ t / . .- . . . . . . 3

• https://github.com/jacobgil/pytorch-pruning
• https://jacobgil.github.io/deeplearning/pruning-deep-learning

p • U ir – P t n – C – e U • e – n G r g ( /) n VGG16 on Birds 200 dataset • Better u – – ca u 4

N • C • – 5

N • C • – 6

N • C • – 7

• [Han+, ‘15] 3x3 1 0 3 0 5 0 7 0 0 [Han+, ‘16] h/ps://www.slideshare.net/ren4yu/deep-neural-network-79382352 P.34-P.43 8

• https://www.slideshare.net/ren4yu/deep-neural-network-79382352

6 24 4 1 1- • • pruning – P • pruning N 4 1 1-G V = • – :!: 21 2 6 21 16 9

• P – – • - > > > I 10

T e - • • - - • - - T – • - ℎ" - ] - - L n C - 3 n i i - p p t +)0(93 t [ i o cu , gradient ac,va,on 0 -3 =0 – l ad > 0 - 3 rs 11

• Taylor Oracle 12

Better Be#er Taylor 13

- Update 14

15

• – CPU, GPU 16

cfD • A T 7 9 3 4 ie 2 7 42 A 4 5 daIH 44 /4A 9 /4 8 – .4 4 7 – M • – :7 43A A N 7 9 h 9734 4 4 :9 44 /4A 9 /4 4 A 344 4A 9 4 8 8 17

• http://eyeriss.mit.edu/tutorial.html
• https://www.slideshare.net/ren4yu/deep-neural-network-79382352