鎴戠殑妯″瀷鏈夊蹇紵鈥斺€旀繁搴﹀涔犵綉缁滄ā鍨嬬殑杩愮畻澶嶆潅搴︺€佺┖闂村崰鐢ㄥ拰鍐呭瓨璁块棶鎯呭喌璁＄畻

Posted 2021-04-07 鏈哄櫒瀛︿範AI绠楁硶宸ョ▼

鍚慉I杞瀷鐨勭▼搴忓憳閮藉叧娉ㄤ簡杩欎釜鍙?/span>馃憞馃憞馃憞

---

绔崍鎯呮祿锛岀步椋橀

鍏嶈垂棰嗗彇  瀹氬埗鐝嶈吹鍝佺墝棣欑步

瀹炵墿瑙佷笅鍥撅紝棰嗗彇鏂瑰紡瑙佹湰鏂囨湯銆?/span>

---

娣卞害缃戠粶鐨勮绠楁秷鑰楁槸瀛︽湳 paper 鐩稿灏戣鐨勮瘽棰樸€傚綋鐒讹紝鏃╂湡缃戠粶绮惧害涓嶅鐨勬儏鍐典笅璁ㄨ鍘嬬缉涔熸病鏈夋剰涔夈€傚伐绋嬪笀闇€瑕佸疄鐜版ā鍨嬪苟璁╃綉缁滃敖鍙兘鍦板湪鍚勭被鐜涓嬪伐浣滐紝妯″瀷鐨勮祫婧愭秷鑰楁儏鍐靛拰杩愯閫熷害闈炲父鍏抽敭銆?/p>

鍘熸枃浠ョЩ鍔ㄧ鐨勬ā鍨嬪簲鐢ㄤ负渚嬶紝鍒楀嚭浜嗗洓涓富瑕侀棶棰橈細

绌洪棿鍗犵敤鈥斺€斿崟涓ā鍨嬬殑鍙傛暟鏂囦欢瑕佸崰鐢ㄥ澶х┖闂?br class="mq-15">鍐呭瓨鍗犵敤鈥斺€旇繍琛屽湪鎵嬫満鎴栧钩鏉夸笂鏃堕渶瑕佸崰鐢ㄥ澶х殑 RAM
杩愯閫熷害鈥斺€斿挨鍏惰€冭檻瀹炴椂鐨勮棰戝拰澶у浘鍍忓鐞嗘儏褰?br class="mq-17">鑰楃數鎯呭喌鈥斺€旀垜鍙笉鎯宠鏆栨墜瀹?/p>

妗堜緥锛氫綔鑰呯殑涓€浣嶅鎴锋渶杩戠敤 MobileNetV2 鏇挎崲鎺変簡 V1 妯″瀷锛屾寜鐞嗚V2 鐨勮绠楅噺杩滃皬浜?V1 锛?/span>

锛堟敞锛氬彲鍙傝€?/p>

https://www.zhihu.com/question/265709710/answer/299136290锛宧ttps://www.reddit.com/r/MachineLearning/comments/8a7sf6/d_mobilenet_v2_paper_said_depthwise_separable/銆?/p>

瀹樻柟宸茬粡鏀惧嚭妯″瀷 https://github.com/tensorflow/models/tree/master/research/slim/nets/mobilenet 椤甸潰涓婁篃鏈夊疄楠屾祴璇曠粨鏋溿€傜湅瀹屽叏鏂囦篃浼氬彂鐜?V2 涓嶆瘮 V1 鎱€傦級

1.璁＄畻娑堣€?/p>

鍙互鐢?FLOPS锛坒loating point operations per second锛屾瘡绉掓诞鐐硅繍绠楁暟锛夋潵琛￠噺妯″瀷鐨勯€熷害銆傚彟涓€绉嶆柟娉曟槸 MACCs锛坢ultiply-accumulate operations锛屼箻-鍔犳搷浣滐級锛屼篃鍙?MAdds銆備絾璇寸┛浜嗭紝閮芥槸鐐圭Н杩愮畻鑰屽凡銆?/p>

浠€涔堝彨涔?鍔狅紵绁炵粡缃戠粶閲岀殑杩愮畻澶ч兘鏄繖鏍风殑锛?/p>

w 鍜?x 閮芥槸鍚戦噺锛寉 鏄爣閲忋€備笂寮忔槸鍏ㄨ繛鎺ュ眰鎴栧嵎绉眰鐨勫吀鍨嬭繍绠椼€備竴娆′箻-鍔犺繍绠楀嵆涓€娆′箻娉?涓€娆″姞娉曡繍绠楋紝鎵€浠ヤ笂寮忕殑 MACCs 鏄痭 銆?/p>

涓嶈繃鍙互鐪嬪埌锛屽姞娉曡繍绠楃殑娆℃暟骞堕潪 n 鑰屾槸 n-1 銆備絾鑰冭檻 MACCs 鏃跺彲浠ョ被姣旂畻娉曞鏉傚害浼扮畻鐨?big-O 锛屽嵆缁撴灉鍙互鏄繎浼肩殑銆?/p>

鑰屾崲鍒?FLOPS 鐨勬儏鍐碉紝鐐圭Н鍋氫簡 2n-1 FLOPS锛屽嵆 n-1 娆″姞娉曞拰 n 娆′箻娉曘€傚彲浠ョ湅鍒帮紝MACCs 澶х害鏄?FLOPS 鐨勪竴鍗娿€?/p>

1.1 鍏ㄨ繛鎺ュ眰

鍏ㄨ繛鎺ュ眰鐨勮绠?/p>

鏉冮噸 W

鏄竴涓?I脳J 鐭╅樀锛岃緭鍏?x 鏄?I 缁村疄鍊煎悜閲忥紝b 鏄?J 缁村亸缃€傝緭鍑?y 涔熸槸 J缁村疄鍊煎悜閲忋€侳C 灞傜殑 MACCs 涔熶笉闅捐绠椼€?/p>

涓婃枃渚嬪瓙鏄悜閲忎笌鍚戦噺鐨勭偣绉紝FC 鏄悜閲忎笌鐭╅樀鐨勭偣绉紝姣忎竴缁勭偣绉彂鐢熷湪杈撳叆 x

鍚屾潈閲?W 鏌愪竴鍒椾箣闂达紝璁℃湁 I MACCs锛屼竴鍏辫璁＄畻 J 缁勭偣绉紝鎵€浠?FC 灞傜殑 MACCs 鎬昏 I脳J锛岃窡鏉冮噸鐨勫昂瀵镐竴鑷淬€?/p>

鍋忕疆椤?b

瀵?MACCs 鐨勫奖鍝嶅彲浠ュ拷鐣ヤ笉璁°€傝€屼笂闈篃鎻愬埌 MACCs 涓姞娉曟瘮涔樻硶灏戜竴娆★紝 b

鍒氬ソ琛ヤ笂浜嗚繖涓己銆?/p>

鎵€浠ワ紝瀵笽鐨勮緭鍏ャ€佹潈閲嶄负 I脳J 鐨勬潈閲嶇煩闃靛拰 J 鐨勮緭鍑猴紝MACCs 涓?I脳J 锛孎LOPS 涓?(2I鈭?)脳J銆?/p>

涓句緥锛?/p>

涓€涓叏杩炴帴灞傦紝杈撳叆 100 缁达紝杈撳嚭 300 缁达紝MACCs 鏈?300脳100=30,000

銆備笉杩囷紝濡傛灉涓€涓叏杩炴帴灞傜揣鎺ョ潃鍗风Н灞傦紝杈撳叆鍙兘娌℃湁鎸囧畾闀垮害 I 浣嗘湁 feature map 鐨勫昂瀵告瘮濡傦紙512, 7, 7锛夈€傚湪 Keras 閲屽氨闇€瑕佸啓涓€琛?Flatten 鎶婂畠灞曞钩锛岃繖鏍锋鏃剁殑 I 灏辨槸 512脳7脳7浜嗐€?/p>

1.2 婵€娲诲嚱鏁?/p>

FC 瀹屼簡鎺ヤ笅鏉ラ€氬父鏈変釜婵€娲诲嚱鏁帮紝ReLU 鎴栬€?Sigmoid銆傛縺娲诲嚱鏁扮殑璁＄畻娌℃湁鐐圭Н锛屾墍浠ュ彧鐢?FLOPS 琛￠噺銆?/p>

瀵硅緭鍑轰负 J FC 灞傦紝ReLU 鏈?J

FLOPS锛?/p>

鎴戜滑鎶婂姞鍑忎箻闄ゃ€佹寚鏁般€佸钩鏂规牴绛夌瓑杩愮畻閮界畻浣滀竴娆?FLOPS锛岃繖閲屾湁闄ゆ硶銆佸姞娉曘€佹寚鏁板拰鍑忔硶鍥涚杩愮畻锛屾墍浠?FLOPS 灏辨槸 J脳4銆?/p>

鐩稿浜庡叏杩炴帴鐨勭煩闃佃繍绠楋紝婵€娲诲嚱鏁扮殑璁＄畻閲忛€氬父蹇界暐涓嶈锛堝崥涓绘敞锛氫笉涓€瀹氾紝鐪嬫儏鍐碉級銆?/p>

1.3 鍗风Н灞?/p>

鍗风Н灞傝鍗曠嫭绠楄€屼笉鏄敤鍏ㄨ繛鎺ュ眰鐨勭粨璁猴紝鏄洜涓鸿緭鍏ヨ嚦灏戞槸涓夌淮鐨勶細H脳W脳C銆傚浜庤繖鏍风殑鍗风Н灞傦紝MACCs 鏈夛細

瑙ｉ噴涓€涓嬶細

杈撳嚭鐨?feature map 閲屾瘡涓€氶亾涓婃湁 Hout脳Wout涓厓绱狅紝鏉冮噸浠?K脳K澶у皬鐨勭獥鍙ｏ紝鍦ㄦ墍鏈夌殑 Cin涓€氶亾涓婂仛鐐圭Н锛屽叡鏈?Cout涓嵎绉牳锛屼笂杩版搷浣滈噸澶嶄簡 Cout 娆?/p>

鍚屾牱锛岃繖閲屼篃蹇界暐浜嗗亸缃拰婵€娲诲嚱鏁般€備笉搴旇蹇界暐鐨勬槸 stride锛堟闀匡級銆乨ilation factors锛堟紡瀛?鑶ㄨ儉鍗风Н锛夈€乸adding锛堝～鍏咃級锛岃繖灏辨槸涓轰粈涔堢洿鎺ヤ粠杈撳嚭灏哄 Hout脳Wout

寮€濮嬬畻鐨勫師鍥犫€斺€旈兘宸茬粡鑰冭檻鍦ㄥ唴浜嗐€?/p>

涓句緥锛?/p>

3脳3鍗风Н锛?28 涓?filer锛岃緭鍏ョ殑 feature map 鏄?112脳112脳64锛宻tride=1锛宲adding=same锛孧ACCs 鏈夛細

3脳3脳64脳112脳112脳128=924,844,032

鎺ヨ繎鍗佷嚎鐨勪箻-鍔犳搷浣溿€?/p>

1.4 Batch Normalization

璁＄畻鍏紡锛?/p>

棣栧厛浠ヨ緭鍏ヤ负鍗风Н灞傜殑鎯呭喌涓轰緥銆?/p>

姣忎釜閫氶亾涓婇兘瀛樺湪涓€缁?mean 銆乥eta 銆乬amma 銆乿ariance 锛孋涓€氶亾灏辨湁 C脳4涓彲瀛︿範鐨勫弬鏁般€傝€屼笖 BN 鏄綔鐢ㄥ湪姣忎竴涓厓绱犱笂鐨勶紝杩欐牱鐪嬫潵锛岄€犳垚鐨?FLOPS 搴旇涓嶅皯銆?/p>

浣嗘湁瓒ｇ殑鏄紝鍦?BN 鐩存帴杩炴帴鍗风Н灞傜殑鎯呭喌涓嬶紝鍗?Conv-BN-ReLU 鏃讹紝閫氳繃涓€缁勬帹瀵硷紝鍙互灏?BN 鐨勮绠楁暣鍚堝埌鍗风Н灞傚綋涓紙娉ㄦ剰杩欐槸 inference 鐨勬儏鍐碉紝璺熻缁冮樁娈靛樊鍒緢澶э級锛屼粠鑰屾秷鍘荤殑 BN 灞傞€犳垚鐨?FLOPS銆傚鏋滄槸 Conv-ReLU-BN 鐨勭粨鏋勮繖涓€濂楀氨琛屼笉閫氫簡銆?/p>

锛?BN 灞傜殑璁＄畻缁撳悎鍒?Conv 灞備腑鍘伙紝BN 灞傜殑 FLOPS 娑堝け浜嗭紝Conv 灞傞渶瑕佷箻涓€涓父绯绘暟锛?/p>

鍗充粠缁撴灉涓婃潵璇达紝鍦?inference 鏃舵ā鍨嬩腑鐨?BN 灞傚疄闄呰娑堝幓浜嗐€?/p>

1.5 鍏朵粬灞?/p>

鍍?Pooling 灞傝櫧鐒剁‘瀹炲緢鍏抽敭锛屼絾娌℃湁鐢ㄥ埌鐐圭Н杩愮畻锛屾墍浠?MACCs 涓嶈兘寰堝ソ鍦拌　閲忚繖閮ㄥ垎璁＄畻娑堣€椼€傚鏋滅敤 FLOPS锛屽彲浠ュ彇 feature map 鐨勫昂瀵哥劧鍚庝箻涓€涓父绯绘暟銆?/p>

濡?maxpooling 灞傦紝stride=2銆乫ilter_sz=2锛堝嵆杈撳嚭淇濇寔鐩稿悓灏哄锛夛紝112 x 112 x 128 鐨刦eature map锛孎LOPS 灏辨槸 112 x 112 x 128 = 1,605,632 銆傜浉瀵瑰嵎绉眰鍜屽叏杩炴帴灞傜殑杩愮畻锛岃繖涓绠楅噺姣旇緝灏忥紝鎵€浠ヤ篃鍙互蹇界暐涓嶈銆?/p>

RNN 杩欓噷涓嶅仛璁ㄨ銆傜畝鍗曟潵璇达紝浠?LSTM 涓轰緥锛岃绠椾富瑕佹槸涓や釜澶х殑鐭╅樀涔樻硶锛宻igmoid锛宼anh 鍜屼竴浜涘厓绱犵骇鐨勬搷浣溿€傚彲浠ョ湅鎴愪袱涓叏杩炴帴灞傜殑杩愮畻锛屾墍浠?MACCs 涓昏鍙栧喅浜庤緭鍏ャ€佽緭鍑哄拰闅愮姸鎬佸悜閲忕殑灏哄銆傜偣绉繍绠楄繕鏄崰浜嗗ぇ澶淬€?/p>

2. 鍐呭瓨鍗犵敤

鍐呭瓨甯﹀鍏跺疄姣?MACCs 鏇撮噸瑕併€傜洰鍓嶇殑璁＄畻鏈虹粨鏋勪笅锛屽崟娆″唴瀛樿闂瘮鍗曟杩愮畻鎱㈠緱澶氱殑澶氥€?/p>

瀵规瘡涓€灞傜綉缁滐紝璁惧闇€瑕侊細

娑夊強澶ч噺鐨勫唴瀛樿闂€傚唴瀛樻槸寰堟參鐨勶紝鎵€浠ョ綉缁滃眰鐨勫唴瀛樿鍐欏閫熷害鏈夊緢澶х殑褰卞搷锛屽彲鑳芥瘮璁＄畻鑰楁椂杩樿澶氥€?/p>

2.1 鏉冮噸鐨勫唴瀛樺崰鐢?/p>

鍏ㄨ繛鎺ュ眰鏈?I x J 澶у皬鐨勬潈閲嶇煩闃碉紝鍔犱笂鍋忕疆鍚戦噺鍏辫 (I + 1) x J 銆?/p>

鍗风Н灞傜殑 kernel 閫氬父鏄鏂瑰舰鐨勶紝瀵?kernel_sz = K 鍜岃緭鍏ラ€氶亾涓?Cin 銆佽緭鍑?Cout 鍜岄澶栫殑 Cout 涓亸缃殑鎯呭喌锛屽叡鏈?(K x K x Cin + 1) x Cout 涓弬鏁般€傚姣斾箣涓嬪嵎绉眰鐨勫弬鏁伴噺杩滃皬浜庡叏杩炴帴銆?/p>

涓句緥锛?/p>

鍏ㄨ繛鎺ュ眰鏈?096涓緭鍏ュ拰4096涓緭鍑猴紝鎵€浠ユ潈閲嶆暟 (4096+1) x 4096 = 16.8M 銆?/p>

3 x 3 銆?8涓嵎绉牳锛屽湪64x64 銆?2涓€氶亾鐨勮緭鍏ヤ笂璁＄畻锛屽叡鏈?3 x 3 x 32 x 48 + 48 = 13, 872 涓潈閲嶃€?/p>

娉ㄦ剰鍒版澶勫嵎绉眰鐨勮緭鍏ュ疄闄呮槸鍏ㄨ繛鎺ュ眰鐨?2鍊嶏紙閫氶亾锛夛紝杈撳嚭鏄?8鍊嶏紝鐒堕箙鏉冮噸鏁板彧鏈夊悗鑰呯殑鍗冨垎涔嬩竴涓嶅埌銆傚叏杩炴帴灞傜殑鍐呭瓨鍗犵敤鐪熺殑寰堝彲鎬曘€?/p>

浣滆€呮敞锛氬嵎绉眰鍙互鐪嬩綔涓€涓彈闄愯繛鎺ョ殑鍏ㄨ繛鎺ュ眰锛屽嵆鏉冮噸瀵?k x k 浠ュ鐨勮緭鍏ョ疆闆讹紝涓嶄娇鐢ㄣ€?/p>

2.2 feature maps 鍜屼腑闂寸粨鏋?/p>

CS231n 鐨?Lesson 9 涓撻棬鑺变簡寰堝绡囧箙璁?feature map 鐨勮绠楋紝鍙互鍙傝€冦€?/p>

杩樻槸涓句緥璇存槑銆傚嵎绉眰鐨勮緭鍏ユ槸 224x224x3 锛屾妸鎵€鏈夎繖浜涘€艰鍑烘潵闇€瑕佽闂?150,528 娆″唴瀛樸€傚鏋滃嵎绉牳鏄?KxKxCout 锛岃繕瑕佷箻涓婅繖涓郴鏁帮紙鍥犱负姣忔鍗风Н閮借璁块棶涓€閬嶏級銆傛嬁 stride=2, kernel 鏁颁负32鐨勬儏鍐垫潵璇达紝杈撳嚭鐨?feature map 灏哄涓?112x112x32锛屽叡璁?401,408 娆″唴瀛樿闂€?/p>

鎵€浠ワ紝姣忓眰鐨勫唴瀛樿闂€绘暟濡備笅锛?/p>

杩欑鎯呭喌涓?weights 閮ㄥ垎涔熶細鍙樺緱寰堝ぇ锛屾墍浠ユ槸涓嶈兘蹇界暐鐨勩€?/p>

raw_convnet

杩欏箙鍥炬槸閫氳繃寮€婧愮殑宸ュ叿draw_convnet(https://github.com/gwding/draw_convnet)鐢熸垚鐨勩€傚湪娓呮鏁翠釜鍓嶅悜璁＄畻缃戠粶涓殑姣忎竴涓眰鐨勮緭鍏ヨ緭鍑轰互鍙婂弬鏁拌缃悗鍙互鑷繁鎵嬪姩鐢诲嚭璁＄畻鍥惧嚭鏉ワ紝瀵逛簬鍙傛暟閲忚绠楀氨寰堢洿瑙備簡銆?/span>

feature map澶у皬璁＄畻

LeNet-5

涓嬮潰鏄竴涓閫氶亾鍥惧儚鐨勮緭鍏eNet-5缃戠粶鍓嶅悜璁＄畻妯℃嫙鍥撅細

缃戠姸绔嬩綋鏍煎瓙琛ㄧずkernel锛屽叾浠栭鑹叉柟鍥捐〃绀篺eature map(Input琛ㄧず杈撳叆灞傦紝鍙互鐪嬪仛鐗规畩鐨刦eature map)涓€涓猭ernel瀵瑰簲涓€涓猣eature map鍙傛暟閲忎富瑕佷负kernel澶у皬 姣忎釜kernel甯︿竴涓猙ias

鏁翠釜缃戠粶鍗犳嵁鏉冮噸鐨勪负Convolution/Innerproduct 涓ゅ眰锛屽垎鍒绠楀弬鏁伴噺涓猴細

鐢?bytes鐨刦loat绫诲瀷鏉ュ瓨鍌ㄥ弬鏁帮紝鍒欐€荤殑鍙傛暟閲忓ぇ灏忎负锛?/p>

500 + 25000 + 400000 + 5000 + 锛?0 + 50 + 500 + 10锛?= 431080

瀛楄妭鏁颁负锛?/p>

431080 x 4 = 1724320 鈮?1683.90625kb 鈮?1.64M

瀵规瘮瀹為檯LeNet-5缃戠粶鍩轰簬caffe璁粌鍑烘潵鐨勬ā鍨嬪ぇ灏忎负锛?.64 MB (1,725,025 瀛楄妭)锛屽熀鏈帴杩戯紝鍥犱负妯″瀷涓彲鑳借繕甯︽湁闄勫姞鐗规€у弬鏁般€?/p>

2.3 Fusion

杩欎竴鑺傜殑鎰忔€濇槸锛屽儚 ReLU 杩欐牱姣旇緝绠€鍗曠殑杩愮畻锛屽鏋滀笉鍋氫紭鍖栵紝鍦ㄨ绠楁椂杩戜箮鏄粠杈撳叆鍒拌緭鍑哄仛浜嗕竴娆℃嫹璐濄€傝绠楀彲浠ヨ涓轰笉鑰楁椂闂达紝浣嗗唴瀛樿闂繕鏄湁娑堣€楃殑锛屾墍浠ュ彲浠ユ妸杩欎竴姝ュ悓鍗风Н灞傜殑璁＄畻鍚堟垚锛屼粠鑰岃妭鐪佷簡涓€杞唴瀛樿鍐欍€?/p>

3. MobileNet V2 vs. V1

杩欓儴鍒嗕綔鑰呰浜嗕粬璁や负 V2 涓嶄細姣?V1 蹇殑鍒嗘瀽杩囩▼銆傜粨璁鸿窡寮€澶村崥涓诲紩鐨勫浘鐩歌繎锛屽嵆涔樺瓙閮戒负1.0鏃讹紝V2鏄樉钁楀揩浜嶸1鐨勶紝浣哣2鍦ㄤ箻瀛愪负1.4鏃堕€熷害姣擵1绋嶆參銆?/p>

鑷充簬鍘熷洜鍢涳紝绠€鍗曟潵璇村氨鏄?V2 鐨勫眰鏁版洿娣憋紝姣忓眰鐨勮緭鍏ヨ緭鍑哄弬鏁拌鍐欏鑷村唴瀛樿闂噺澶у銆傚洜姝や綔鑰呰涓哄奖鍝?inference 閫熷害鐨勭摱棰堝叾瀹炰笉鍦?MACCs锛岃€屾槸鍐呭瓨璁块棶鏁帮紙memory accesses锛夈€?/p>

V2 with multiplier=1.4 鐨勯€熷害鐣ユ參浜?V1锛屼絾绮惧害楂樺嚭涓嶅皯锛沄2 with multiplier=1.0 閫熷害姣?V1 蹇緢澶氥€傚彲浠ユ牴鎹渶瑕佽繘琛屽彇鑸嶃€傚畼鏂归〉闈笂涔熺粰浜嗗緢澶氬疄楠屽弬鑰冦€?/p>

鐒跺悗浣滆€呭 VGG16 鍋氫簡涓€鐐硅€冨療锛岀粨璁哄緢鏈夋剰鎬濄€?/p>

VGG16 缁忓父琚綋浣滃浘鍍忔柟闈㈢殑鐗瑰緛鎻愬彇鍣紝缁撴瀯寰堢畝鍗曪紝灞傛暟涔熶笉澶氾紝鐪嬭捣鏉ュソ鍍忚绠楁瘮杈冨銆佸唴瀛樿闂細灏戜竴浜涳紝鐪熺殑鏄繖鏍峰悧锛熷姣?MobileNet锛堣緭鍏ユ寜绉诲姩璁惧16:9鐨勮鏍硷紝鏄?26x224锛屽彲浠ョ畻鍑轰互涓嬬粨鏋滐細

鎵€浠ユ洿澶х殑 feature map 瀵艰嚧浜嗘洿澶氱殑鍐呭瓨璁块棶銆?/p>

4 缁撹

璁烘枃涓?MobileNet V2 涓昏姣旇緝浜?MACCs 鍜屽弬鏁伴噺锛屾寚鍑哄洜涓鸿繖涓ら」瑙勬ā鏇村皬鎵€浠ラ€熷害鏇村揩銆備絾瀹為檯涓婅繕瑕佽€冭檻鍐呭瓨璁块棶鐨勬儏鍐点€?/p>

鍙﹀鏈枃缁欏嚭鐨?MACCs銆佸唴瀛樿闂€佸弬鏁伴噺閮芥槸浼拌鍊硷紝鍙敤浜庡悓绫绘ā鍨嬬殑澶嶆潅搴︽瘮杈冿紝鍑轰簡杩欎釜璇鏄鏃犳剰涔夌殑銆?/p>

杩涗竴姝ラ槄璇?/p>

璁烘枃锛?/p>

http://machinethink.net/blog/how-fast-is-my-model/ 

http://blog.csdn.net/cheese_pop/article/details/51955915

http://timdettmers.com/2015/03/26/convolution-deep-learning/

---

绀煎搧 1:  鐝嶈吹鍝佺墝 浜旀柟鏂?棣欑步绀肩洅  1鐩?nbsp;

绀煎搧 2:   鐭ュ悕鏈哄櫒瀛︿範绠楁硶缃戣  

閰嶅20璁茶棰?浠ｇ爜+璇句欢

棰嗗彇鏂瑰紡

閭€璇?浣嶅ソ鍙嬫壂鎻忎笅鏂逛簩缁寸爜鐧婚檰鑵捐浜?绀惧尯锛屽苟鍏虫敞 鏈哄櫒瀛︿範AI绠楁硶宸ョ▼

璇︾粏淇℃伅鍙仈绯诲皬缂?nbsp;寰俊  hai299014