¼òµ¥µÄÒôƵʶ±ð- b8 [ A3 E, R
" X6 a& y, d+ [: {2 @# I3 h& C7 f% q! d
½ñÌìµÄÎÄÕ½«ÏòÄúչʾÈçºÎ¹¹½¨¿ÉÒÔʶ±ð 10 ¸ö²»Í¬×ִʵĻù±¾ÓïÒôʶ±ðÍøÂç¡£ÐèҪעÒâµÄÊÇ£¬ÕæÕýµÄÓïÒôºÍÒôƵʶ±ðϵͳҪ¸´Ôӵö࣬µ«¾ÍÏñÓÃÓÚʶ±ðͼÏñµÄ MNIST£¬Õâ¸ö»ù±¾ÓïÒôʶ±ðÍøÂçÄܹ»°ïÖúÄú»ù±¾Á˽âËùÉæ¼°µÄ¼¼Êõ¡£Ñ§Íê±¾½Ì³Ìºó£¬Äú½«»ñµÃÒ»¸öÄ£ÐÍ£¬¸ÃÄ£Ðͻ᳢ÊÔ½«Ê±³¤Îª 1 ÃëµÄÒôƵƬ¶Î¹éÀàΪÎÞÉù¡¢Î´Öª×Ö´Ê¡¢¡°yes¡±¡¢¡°no¡±¡¢¡°up¡±¡¢¡°down¡±¡¢¡°left¡±¡¢¡°right¡±¡¢¡°on¡±¡¢¡°off¡±¡¢¡°stop¡± »ò ¡°go¡±¡£Äú»¹¿ÉÒÔÔÚ Android Ó¦ÓÃÖÐÔËÐиÃÄ£ÐÍ¡£! F% I8 O" a3 Q0 k5 C/ f I! b
: ~7 n$ j8 O# H" m9 `" j9 Q; {& q' m) q3 g2 t$ P. W
×¼±¸
5 n) G' ~# S9 ~ÄúӦȷ±£°²×°ÁË TensorFlow£»´ËÍ⣬ÓÉÓڽű¾»áÏÂÔØ³¬¹ý 1GB µÄѵÁ·Êý¾Ý£¬Òò´ËÄúÐèҪȷ±£¼ÆËã»úÓµÓÐÎȶ¨µÄ»¥ÁªÍøÁ¬½ÓºÍ×ã¹»µÄ¿ÉÓÿռ䡣ѵÁ·¹ý³Ì¿ÉÄÜÐèÒª¼¸¸öСʱ£¬Òò´ËÇëÈ·±£ÄúµÄ¼ÆËã»ú¿ÉÒÔÍê³ÉÕâô³¤Ê±¼äµÄѵÁ·²Ù×÷¡£
( w# a) c8 e5 J# h- U! g3 T& n/ o
8 O, l1 g0 @$ d) |# O5 d6 |% B- F( v' k" z7 C
ѵÁ·% r! C1 G4 i* a
Òª¿ªÊ¼ÑµÁ·¹ý³Ì£¬Çëתµ½ TensorFlow Ô´´úÂëÊ÷£¬È»ºóÔËÐÐÒÔϽű¾£º
# G" e$ k6 g& B" d9 Z! J' H! p/ ]python tensorflow/examples/speech_commands/train.py+ a9 ^+ q, m9 K% S7 s0 p7 I
+ B7 Z* ]: i3 j5 x6 [
¸Ã½Å±¾»áÏÈÏÂÔØÓïÒôÖ¸ÁîÊý¾Ý¼¯£¬ÆäÖаüº¬³¬¹ý 105000 ¸ö WAVE ÒôƵÎļþ£¬ÒôƵÄÚÈÝÊÇÓÐÈË˵³ö 30 ¸ö²»Í¬µÄ×Ö´Ê¡£ÕâЩÊý¾ÝÓÉ Google ÊÕ¼¯£¬²¢ÒÀ¾Ý CC BY Ðí¿É·¢²¼£¬Äú¿ÉÒÔÌá½» 5 ·ÖÖÓ×Ô¼ºµÄ¼ÒôÀ´°ïÖú¸Ä½ø¸ÃÊý¾Ý¡£¹éµµÊý¾Ý³¬¹ý 2GB£¬Òò´ËÕⲿ·Ö¹ý³Ì¿ÉÄÜÐèÒªÒ»¶Îʱ¼ä£¬µ«ÄúÓ¦¸Ã¿ÉÒÔ¿´µ½½ø¶ÈÈÕÖ¾£»ÏÂÔØÍê³Éºó£¬ÄúÎÞÐèÔÙ´ÎÖ´Ðд˲½Öè¡£ÈçÐèÏêϸÁ˽â¸ÃÊý¾Ý¼¯£¬Çë²ÎÔÄ https://arxiv.org/abs/1804.03209- B- K# ^/ b. V$ n7 U k
7 K* R, ?. z8 S( e, c1 Q( U* G
ÏÂÔØÍê³Éºó£¬Äú½«¿´µ½ÈçÏÂÈÕÖ¾ÐÅÏ¢£º4 H* `: n& m6 Q, H- m6 o
I0730 16:53:44.766740 55030 train.py:176] Training from step: 1. @ I N x/ H. o% U
I0730 16:53:47.289078 55030 train.py:217] Step #1: rate 0.001000, accuracy 7.0%, cross entropy 2.611571
?6 }2 s y1 H- |2 k# A! H
. ~# ?: |9 @$ u4 kÕâ±íÃ÷³õʼ»¯¹ý³ÌÒѾÍê³É£¬ÑµÁ·Ñ»·ÒѾ¿ªÊ¼¡£Äú½«¿´µ½¸ÃÈÕÖ¾Êä³öÿ¸öѵÁ·²½µÄÐÅÏ¢¡£ÏÂÃæÏêϸ˵Ã÷Á˸ÃÈÕÖ¾ÐÅÏ¢µÄº¬Ò壺
1 L9 T) x. a+ `5 w
0 f# ], r; j4 C1 I3 c* l, p5 k) CStep #1 ±íÃ÷ÕýÔÚ½øÐÐѵÁ·Ñ»·µÄµÚÒ»²½¡£ÔÚ´ËʾÀýÖÐ×ܹ²ÓÐ 18000 ¸öѵÁ·²½£¬Äú¿ÉÒԲ鿴²½±àºÅÀ´Á˽⻹ÓжàÉÙ²½¼´¿ÉÍê³É¡£
4 c2 Q; _' F& f3 Y" ]
& I- C2 E+ ^* M+ e+ Y6 @rate 0.001000 ÊÇ¿ØÖÆÍøÂçÈ¨ÖØ¸üÐÂËٶȵÄѧϰËÙÂÊ¡£ÔÚѵÁ·µÄÔçÆÚ½×¶Î£¬ËüÊÇÒ»¸öÏà¶Ô½Ï´óµÄÊý×Ö (0.001)£¬µ«ÔÚѵÁ·ÖÜÆÚµÄºóÆÚ»á¼õÉÙµ½ÔÀ´µÄÊ®·ÖÖ®Ò»£¬¼´ 0.0001¡£1 s; g% \, ]4 [! S
8 M2 \! e' \# s- t5 I
accuracy 7.0% ±íʾģÐÍÔÚ±¾ÑµÁ·²½ÖÐÔ¤²âÕýÈ·µÄÀà±ðÊýÁ¿¡£¸Ãֵͨ³£»áÓнϴóµÄ²¨¶¯£¬µ«Ó¦¸Ã»áËæ×ÅѵÁ·µÄ½øÐÐ×ÜÌåÓÐËùÌá¸ß¡£¸ÃÄ£ÐÍ»áÊä³öÒ»¸öÊý×ÖÊý×飬ÿ¸ö±êÇ©¶ÔÓ¦Ò»¸öÊý×Ö£¬Ã¿¸öÊý×Ö¶¼±íʾÊäÈë¿ÉÄܹéÈë¸ÃÀà±ðµÄÔ¤²â¸ÅÂÊ¡£¿Éͨ¹ýÑ¡ÔñµÃ·Ö×î¸ßµÄÌõÄ¿À´ÌôѡԤ²â±êÇ©¡£µÃ·ÖʼÖÕ½éÓÚ 0 µ½ 1 Ö®¼ä£¬ÖµÔ½¸ß±íʾ½á¹ûµÄÖÃÐŶÈÔ½¸ß¡£# U- I/ ~* D+ { T' H) u
|- W; x& u7 t
cross entropy 2.611571 ÊÇÓÃÓÚÖ¸µ¼ÑµÁ·¹ý³ÌµÄËðʧº¯ÊýµÄ½á¹û¡£ËüÊÇÒ»¸öµÃ·Ö£¬Í¨¹ý½«µ±Ç°ÑµÁ·ÔËÐеĵ÷ÖÏòÁ¿ÓëÕýÈ·±êÇ©½øÐбȽϼÆËã¶ø³ö£¬¸ÃµÃ·ÖÓ¦ÔÚѵÁ·ÆÚ¼ä³ÊÏ»¬Ç÷ÊÆ¡£
3 M* [0 z! k8 t1 n' K9 e9 b# x* M2 B/ L8 q/ G# r
¾¹ý 100 ²½Ö®ºó£¬ÄúÓ¦¿´µ½ÈçÏÂËùʾµÄÐУº
1 H+ o, h+ n" _8 G, q* w1 ]I0730 16:54:41.813438 55030 train.py:252] Saving to "/tmp/speech_commands_train/conv.ckpt-100"
0 C% n: C Q9 ^
! s3 S4 K( C Z9 L8 k9 n) Z´ËÐлὫµ±Ç°µÄѵÁ·È¨Öر£´æµ½¼ì²éµãÎļþÖС£Èç¹ûѵÁ·½Å±¾ÖжÏÁË£¬Äú¿ÉÒÔ²éÕÒÉϴα£´æµÄ¼ì²éµã£¬È»ºó½« --start_checkpoint=/tmp/speech_commands_train/conv.ckpt-100 ÓÃ×÷ÃüÁîÐвÎÊýÖØÆô¸Ã½Å±¾£¬ÒÔ±ã´Ó¸Ãµã¿ªÊ¼¡£6 v. r n, e+ I! I/ G; Z
# i; s/ F9 Z7 O( d5 m2 A
5 }: x/ @2 K) f0 I `»ìÏý¾ØÕó2 M+ t+ a& N+ Y, H7 g, V4 `
¾¹ý 400 ²½Ö®ºó£¬Äú»á¿´µ½ÒÔÏÂÈÕÖ¾ÐÅÏ¢£º; v! w6 U& V7 |" S- ?( w
I0730 16:57:38.073667 55030 train.py:243] Confusion Matrix:
7 \0 s8 o. z' j! V' z1 m/ u [[258 0 0 0 0 0 0 0 0 0 0 0]
5 E) n& \" c6 n6 {% J4 M8 b5 x$ F5 M [ 7 6 26 94 7 49 1 15 40 2 0 11]! \8 `: U5 u( R% Q# j1 |
[ 10 1 107 80 13 22 0 13 10 1 0 4]! [ g* F( l2 n( Y3 H8 Y9 n& b! ~4 Q/ C
[ 1 3 16 163 6 48 0 5 10 1 0 17]/ W6 Y8 h: C/ \0 f5 N! G
[ 15 1 17 114 55 13 0 9 22 5 0 9]
( k% @8 c- J* Y ?7 q [ 1 1 6 97 3 87 1 12 46 0 0 10]
; t1 X) h2 S5 m# [ [ 8 6 86 84 13 24 1 9 9 1 0 6]" E8 P c& h) H0 f
[ 9 3 32 112 9 26 1 36 19 0 0 9]
& [% K9 M0 W( K& o [ 8 2 12 94 9 52 0 6 72 0 0 2]
1 b8 O& t/ f# L [ 16 1 39 74 29 42 0 6 37 9 0 3]# m) ]8 O, q6 b% X6 z6 R H
[ 15 6 17 71 50 37 0 6 32 2 1 9]
4 G% {3 J0 C0 k S- p0 r [ 11 1 6 151 5 42 0 8 16 0 0 20]]
+ E) H5 D) Q1 z/ y j. y/ E f4 A( k3 _
µÚÒ»²¿·ÖÊÇ»ìÏý¾ØÕó¡£ÒªÀí½âËüµÄ¾ßÌ庬Ò壬ÄúÊ×ÏÈÐèÒªÁ˽âËùÓõıêÇ©¡£ÔÚ±¾Ê¾ÀýÖУ¬ËùÓõıêÇ©ÊÇ ¡°silence¡±¡¢¡°unknown¡±¡¢¡°yes¡±¡¢¡°no¡±¡¢¡°up¡±¡¢¡°down¡±¡¢¡°left¡±¡¢¡°right¡±¡¢¡°on¡±¡¢¡°off¡±¡¢¡°stop¡± ºÍ ¡°go¡±¡£Ã¿Áдú±íÒ»×鱻ģÐÍÔ¤²âΪÿ¸ö±êÇ©µÄÑù±¾£¬Òò´ËµÚÒ»Áдú±íÔ¤²âΪÎÞÉùµÄËùÓÐÒôƵƬ¶Î£¬µÚ¶þÁдú±íÔ¤²âΪδ֪×ִʵÄËùÓÐÒôƵƬ¶Î£¬µÚÈýÁдú±íÔ¤²âΪ ¡°yes¡± µÄËùÓÐÒôƵƬ¶Î£¬ÒÀ´ËÀàÍÆ¡£, R$ r( s7 m: H& |( T, f
& M5 B7 n" F. g
ÿÐбíʾÒôƵƬ¶Îʵ¼Ê¹éÈëµÄ±êÇ©¡£µÚÒ»ÐÐÊǹéÈëÎÞÉùµÄËùÓÐÒôƵƬ¶Î£¬µÚ¶þÐÐÊǹéÈëδ֪×ִʵÄËùÓÐÒôƵƬ¶Î£¬µÚÈýÐÐÊǹéÈë ¡°yes¡± µÄËùÓÐÒôƵƬ¶Î£¬ÒÀ´ËÀàÍÆ¡£
! O+ q- X# L) I7 j9 A
5 Z/ h' e) l+ h. V! h# e2 X) e# [) S´Ë¾ØÕó±Èµ¥¸ö׼ȷÂʵ÷ָü¼ÓÓÐÓã¬ÒòΪËü¿ÉÒԺܺõØ×ܽáÍøÂç³öÏֵĴíÎó¡£ÔÚ´ËʾÀýÖУ¬Äú¿ÉÒÔ·¢ÏÖ£¬³ýÁ˵ÚÒ»¸öÊýÖµÒÔÍ⣬µÚÒ»ÐÐÖеÄËùÓÐÌõÄ¿¾ùΪ 0¡£ÒòΪµÚÒ»ÐбíʾËùÓÐʵ¼ÊÎÞÉùµÄÒôƵƬ¶Î£¬ÕâÒâζ×ÅËùÓÐÒôƵƬ¶Î¶¼Î´±»´íÎ󵨱ê¼ÇΪ×Ö´Ê£¬Òò´ËÎÒÃÇδµÃ³öÈκÎÓйØÎÞÉùµÄ¼Ù¸ºÀý¡£Õâ±íÊ¾ÍøÂçÒѾ¿ÉÒԺܺõØÇø·ÖÎÞÉùºÍ×Ö´Ê¡£
+ F5 e. O& b$ M N3 S- ]) P4 j1 ]1 c9 ~8 K* j! a) u" W7 ~( V) F
Èç¹ûÎÒÃÇÍùÏ¿´£¬¾Í»á·¢ÏÖµÚÒ»ÁÐÓдóÁ¿·ÇÁãÖµ¡£¸ÃÁбíʾԤ²âΪÎÞÉùµÄËùÓÐÒôƵƬ¶Î£¬Òò´ËµÚÒ»¸öµ¥Ôª¸ñÍâµÄÕýÊýÊÇ´íÎóµÄÔ¤²â¡£Õâ±íʾһЩʵ¼ÊÊÇÓïÒô×ִʵÄÒôƵƬ¶Î±»Ô¤²âΪÎÞÉù£¬Òò´ËÎÒÃǵóöÁ˺ܶà¼ÙÕýÀý¡£6 `& K9 d# a( }, r
3 B( `) E8 T; h, ^" v) ?8 JÍêÃÀµÄÄ£ÐÍ»áÉú³É»ìÏý¾ØÕ󣬳ýÁË´©¹ýÖÐÐĵĶԽÇÏßÉϵÄÌõÄ¿ÒÔÍ⣬ËùÓÐÆäËûÌõÄ¿¶¼Îª 0¡£·¢ÏÖÆ«ÀëÕâ¸öģʽµÄµØ·½ÓÐÖúÓÚÄúÁ˽âÄ£ÐÍ×îÈÝÒ×ÔÚÄÄЩ·½Ãæ»ìÏý£»È·¶¨ÎÊÌâËùÔÚºó£¬Äú¾Í¿ÉÒÔͨ¹ýÌí¼Ó¸ü¶àÊý¾Ý»òÇåÀíÀà±ðÀ´½â¾öÕâЩÎÊÌâ¡£( W: i7 U" f4 P; y4 [! A2 Q& ~
. h: v+ N8 @3 q! I5 A5 g# L
6 v5 J& _* _1 G9 ~( B$ yÑéÖ¤* o& w( z6 U" d* H( @3 Q
ÔÚ»ìÏý¾ØÕóÖ®ºó£¬ÄúÓ¦¿´µ½ÈçÏÂËùʾµÄÐУº4 ]* v+ I6 F$ @ B0 ~9 n+ e
I0730 16:57:38.073777 55030 train.py:245] Step 400: Validation accuracy = 26.3% (N=3093)" m0 y6 X' j5 a
9 _4 l6 t5 u6 ~* F3 M×îºÃ½«Êý¾Ý¼¯·Ö³ÉÈý¸öÀà±ð¡£×î´óµÄ×Ó¼¯£¨±¾Ê¾ÀýÖÐΪԼ 80% µÄÊý¾Ý£©ÓÃÓÚѵÁ·ÍøÂ磬½ÏСµÄ×Ó¼¯£¨±¾Ê¾ÀýÖÐΪԼ 10% µÄÊý¾Ý£¬³ÆÎª ¡°ÑéÖ¤¡± ¼¯£©Ô¤ÁôÏÂÀ´ÒÔÆÀ¹ÀѵÁ·ÆÚ¼äµÄ׼ȷÂÊ£¬ÁíÒ»¸ö×Ó¼¯£¨Ê£Ï嵀 10%£¬³ÆÎª ¡°²âÊÔ¡± ¼¯£©ÓÃÀ´ÆÀ¹ÀѵÁ·Íê³ÉºóµÄ׼ȷÂÊ¡£
# e1 {% H) k9 A# H
6 R$ Y2 |$ e$ {) U' r2 P! |' t$ `5 ^Ö®ËùÒÔ²ÉÓÃÕâÖÖ²ð·Ö·½·¨£¬ÊÇÒòΪʼÖÕ´æÔÚÕâÑùÒ»ÖÖ·çÏÕ£ºÍøÂçÔÚѵÁ·ÆÚ¼ä¿ªÊ¼¼ÇÒäÊäÈ롣ͨ¹ý½«ÑéÖ¤¼¯·ÖÀ뿪À´£¬¿ÉÒÔÈ·±£Ä£ÐÍÄܹ»´¦ÀíËü֮ǰ´Óδ¼û¹ýµÄÊý¾Ý¡£²âÊÔ¼¯ÊÇÒ»ÖÖ¶îÍâµÄ±£»¤´ëÊ©£¬¿ÉÒÔÈ·±£Äú²»½öÒÔÊʺÏѵÁ·¼¯ºÍÑéÖ¤¼¯µÄ·½Ê½µ÷ÕûÄ£ÐÍ£¬¶øÇÒʹģÐÍÄܹ»·º»¯µ½·¶Î§¸ü¹ãµÄÊäÈë¡£
+ y- l- O9 S3 c: ~
" X \. f& X# B/ XѵÁ·½Å±¾»á×Ô¶¯½«Êý¾Ý¼¯·ÖΪÕâÈý¸öÀà±ð£¬ÉÏÃæµÄÈÕÖ¾ÐлáÏÔʾÔÚÑéÖ¤¼¯ÉÏÔËÐÐʱģÐ͵Ä׼ȷÂÊ¡£ÀíÏëÇé¿öÏ£¬¸Ã׼ȷÂÊÓ¦¸ÃÏ൱½Ó½üѵÁ·×¼È·ÂÊ¡£Èç¹ûѵÁ·×¼È·ÂÊÓÐËùÌá¸ßµ«Ñé֤׼ȷÂÊûÓУ¬Ôò±íÃ÷´æÔÚ¹ýÄâºÏ£¬Ä£ÐÍֻѧϰÁËÓйØÑµÁ·ÒôƵƬ¶ÎµÄÐÅÏ¢£¬¶øÃ»ÓÐѧϰÄÜ·º»¯µÄ¸ü¹ã·ºÄ£Ê½¡£! W% @. ?0 w' j; B5 _3 Y
% i2 p+ e& a5 ]7 ^3 x. X
* {4 v1 w& u' q3 j- \; W9 h+ H
Tensorboard
3 @0 a4 b: N5 MʹÓà Tensorboard ¿ÉÒԺܺõع۲ìѵÁ·½ø¶È¡£Ä¬ÈÏÇé¿öÏ£¬½Å±¾»á½«Ê¼þ±£´æµ½ /tmp/retrain_logs£¬Äú¿ÉÒÔͨ¹ýÔËÐÐÒÔÏÂÃüÁî¼ÓÔØÕâЩʼþ£º) n" C S0 J% @4 ]/ a3 `
tensorboard --logdir /tmp/retrain_logs
' x1 X, }7 d" t) c' O
t, n, q/ f8 ~+ oÈ»ºó£¬ÔÚä¯ÀÀÆ÷ÖÐתµ½ http://localhost:6006£¬Äú½«¿´µ½ÏÔʾģÐͽø¶ÈµÄͼ±í¡£
3 f1 y- W- [0 R7 r: M
& d; L. W5 U' p/ r
|