Light Weight CNN妯″瀷鐨勫垎鏋愪笌鎬荤粨
Posted
tags:
篇首语：本文由小常识网(cha138.com)小编为大家整理，主要介绍了Light Weight CNN妯″瀷鐨勫垎鏋愪笌鎬荤粨相关的知识，希望对你有一定的参考价值。
鏍囩锛?a href='http://www.mamicode.com/so/1/die' title='die'>die
htm max tiny point enc resid res gate 鏈枃閫夋嫨浜?涓猯ight weight CNN妯″瀷锛屽苟瀵瑰畠浠殑璁捐鎬濊矾鍜屾€ц兘杩涜浜嗗垎鏋愪笌鎬荤粨锛岀洰鐨勫湪浜庝负鍦ㄥ畬鎴愬浘鍍忚瘑鍒换鍔℃椂妯″瀷鐨勯€夋嫨涓庤璁℃柟闈㈡彁渚涚浉鍏崇殑鍙傝€冭祫鏂欍€?/p>1 绠€浠?/strong>
鑷狝lexNet[1]鍦↙SVRC-2010 ImageNet[22]鍥惧儚鍒嗙被浠诲姟涓婂彇寰楃獊鐮存€ц繘灞曚箣鍚庯紝鏋勫缓鏇存繁鏇村ぇ鐨刢onvolutional neural networks锛圕NN锛夊嚑涔庢垚浜嗕竴绉嶄富瑕佺殑瓒嬪娍[2-9]銆傞€氬父锛岃幏寰梥tate-of-the-art鍑嗙‘鐜囩殑妯″瀷閮芥湁鎴愮櫨涓婂崈鐨勭綉璺眰浠ュ強鎴愬崈涓婁竾鐨勪腑闂寸壒寰侀€氶亾锛岃繖渚垮鑷翠簡涓€浜沜omplex CNN妯″瀷[7-9]闇€瑕佷笂鍗佷嚎鐨凢LOPs锛堟瘡绉掓诞鐐硅绠楁鏁帮級锛屼篃灏遍檺鍒朵簡浠栦滑鍦ㄧЩ鍔ㄦ墜鏈哄钩鍙版垨绉诲姩鏈哄櫒浜哄钩鍙扮殑搴旂敤銆備负浜嗚В鍐宠繖涓€闂锛屾瀯寤簂ight weight CNN妯″瀷[10-19]鎴愪负浜嗕竴涓潪甯告椿璺冪殑鐮旂┒鏂瑰悜銆傝繎骞存潵鍦╨ight weight CNN妯″瀷鏋勫缓鏂归潰鐨勪竴浜涙垚鏋滃凡缁忚〃鏄庯紝涓€浜涚粡杩囩簿蹇冭璁＄殑灏忔ā鍨嬶紝涔熻兘澶熷湪鍥惧儚鍒嗙被闂濡侷mageNet涓婂彇寰梥tate-of-the-art鐨勬€ц兘锛岃€屽湪妯″瀷鍙傛暟浠ュ強璁＄畻鏁堢巼鏂归潰閮芥湁澶у箙搴︾殑鍑忓皯涓庢彁楂樸€備负浜嗚兘澶熷湪瀹屾垚鍥惧儚璇嗗埆浠诲姟锛堝鑿滃搧璇嗗埆绛夛級鏃舵ā鍨嬬殑閫夋嫨涓庤璁℃柟闈㈡彁渚涚浉鍏崇殑鍙傝€冭祫鏂欙紝鏈枃棣栧厛绠€鍗曠殑瀵规瀯寤簂ight weight CNN妯″瀷鐨勬柟娉曡繘琛屼簡浠嬬粛锛岀户鑰岄€夋嫨浜?涓ā鍨嬪苟瀵瑰畠浠殑璁捐鎬濊矾鍜屾€ц兘杩涜浜嗗垎鏋愪笌鎬荤粨銆?/p>
2 Light Weight CNN妯″瀷
鍦ㄦ瀯寤簂ight weight CNN妯″瀷鏂归潰涓昏鏈変袱涓ぇ鐨勬柟鍚? 瀵逛簬Light Weight CNN妯″瀷鐨勮瘎浼伴櫎鍑嗙‘鐜囦箣澶栵紝鎴戜滑杩樻瘮杈冨叧娉ㄦā鍨嬪弬鏁板帇缂╃殑姣旂巼浠ュ強妯″瀷鐨勮绠楁晥鐜囥€傚洜姝わ紝鍦ㄩ€夋嫨妯″瀷鏃堕渶瑕佹敞鎰忎竴涓嬶紝涓€浜涙柟娉曞湪妯″瀷鍙傛暟鍘嬬缉鏂归潰鐨勬瘮鐜囧拰妯″瀷璁＄畻鏂归潰鐨勬晥鐜囦箣闂寸殑鏉冭 銆?br/>锛?锛? All Convolution Net [10] 杩欓噷鏈変竴涓€煎緱娉ㄦ剰鐨勫湴鏂规槸锛孉ll Convolution Net鍜?Fully Convolution Net [20]瀹冧滑鐨勪腑鏂囩炕璇戦兘鍙仛鈥滃叏鍗风Н缃戠粶鈥濓紝浣嗘槸浠栦滑鏄湁鍖哄埆鐨勶紝鍖哄埆鍦ㄤ簬鍓嶈€呮姏寮冧簡姹犲寲灞傚拰鍏ㄨ繛鎺ュ眰锛岃€屽悗鑰呬繚鐣欎簡姹犲寲灞傝€屼涪寮冧簡鍏ㄨ繛鎺ュ眰銆?/p>
锛?锛? SqueezeNet[11] 鍩轰簬涓婅堪鐨勪笁涓璁″師鍒欙紝SqueezeNet鍦↖mageNet涓婁笌AlexNet浠ュ強鍏朵粬鐨勪竴浜涙ā鍨嬪帇缂╂柟娉曡繘琛屼簡瀵规瘮锛屽湪鍑嗙‘鐜囧樊涓嶅鐨勬儏鍐典笅锛孲queezeNet妯″瀷鍙傛暟鏁伴噺鏄捐憲闄嶄綆浜嗭紝濡備笅鍥炬墍绀恒€?/p>
浣嗘槸锛孲queeze Net骞舵病鏈夊湪妯″瀷鍙傛暟鏁伴噺涓庢ā鍨嬭绠楁晥鐜囦箣闂村仛寰堝ソ鐨則rade-off锛岃涓嬪浘锛屼簡瑙ｆ洿澶氬彲璁块棶https://pjreddie.com/darknet/tiny-darknet/ 銆傛帴涓嬫潵锛屾垜浠皢鍒嗘瀽涓や釜鍦ㄥ弬鏁版暟閲忎笌璁＄畻鏁堢巼闂存湁姣旇緝濂界殑trade-off鐨勬ā鍨婱obileNet[12]鍜孲huffleNet[13]銆?/p>
锛?锛? MobileNet[12] MobileNet 鐨勭綉缁滅粨鏋勫叡 28 灞傦紝娌℃湁閲囩敤姹犲寲鐨勬柟寮忚繘琛屼笅閲囨牱锛岃€屾槸鍒╃敤 depth-wise convolution 鐨勬椂鍊欏皢姝ラ暱璁剧疆涓?2锛岃揪鍒颁笅閲囨牱鐨勭洰鐨勩€傜浉杈冧簬 GoogLeNet锛屽湪鍚屼竴涓噺绾х殑鍙傛暟鎯呭喌涓嬶紝浣嗘槸鍦ㄨ繍绠楅噺涓婂嵈灏忎簬 GoogLeNet 涓€涓噺绾э紝鍚屾椂涔熶繚璇佷簡杈冮珮鐨勫噯纭巼锛岃涓嬪浘銆傝繖浜涢兘鏄緱鐩婁簬鍖呭惈depth-wise convolution浠ュ強point-wise convolution鐨刣epth-wise separable convolution銆?/p>
锛?锛? ShuffleNet[13] ShuffleNet璁烘枃涓噰鐢ㄤ簡 Complexity (MFLOPs) 鎸囨爣锛屽湪鐩稿悓鐨?Complexity (MFLOPs) 涓嬶紝瀵筍huffleNet 鍜屽悇涓綉缁滀互鍙婄壒鍒殑涓?MobileNet 杩涜浜嗗姣旓紝鐢变簬 ShuffleNet 鐩歌緝浜?MobileNet 灏戜簡point-wise convolution锛屾墍浠ユ晥鐜囧ぇ澶х殑鎻愰珮浜嗭紝鍚屾椂涔熷彇寰梥tate-of-the-art鐨勫噯纭巼锛岃涓嬪浘銆?/p>
浠庝笂杩?绉嶈交閲忕骇鐨勭綉缁滄ā鍨嬬殑瀹為獙姣旇緝绉嶅彲浠ョ湅鍑猴紝鍩轰簬涓嶅悓鐨勫嵎绉绠楁柟寮忔瀯閫犳柊鐨勭綉缁滅粨鏋勮繖涓柟鍚戞槸闈炲父鏈夋綔鍔涚殑锛岃€屼笖鐜伴樁娈靛凡鏈夌殑涓€浜涙柟娉曚笌涓€浜沜omplex CNN妯″瀷鐩告瘮锛屼篃鍩烘湰鑳藉杈惧埌state-of-the-art鐨勬€ц兘銆傛澶栵紝浠庝笂杩?绉嶇綉缁滅粨鏋勭殑璁捐涓紝鎴戜滑涔熷彲浠ュ緱鍒板涓嬬殑涓€浜涜璁ight Weight CNN妯″瀷鐨勭粡楠岋細 澶囨敞: 浜嗚В鏇村璇疯闂細http://edu.51cto.com/course/14030.html 銆?/p>
[1] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton, 鈥淚mageNet Classification with Deep Convolutional Neural Networks,鈥?in Advances in Neural Information Processing Systems (NIPS), pages 1097鈥?105, 2012
(1) 鍩轰簬涓嶅悓鐨勫嵎绉绠楁柟寮忔瀯閫犳柊鐨勭綉缁滅粨鏋?/strong>
濡侫ll Convolution Net[10]锛孲queezeNet[11]锛孧obileNet[12]浠ュ強ShuffleNet[13]绛夈€?br/>(2)鍦ㄥ凡璁粌濂界殑妯″瀷涓婂仛瑁佸壀[14-19]
Pruning锛氫粠鏉冮噸锛坵eight锛夊眰闈㈡垨浠庯紙kernel & channel锛夊眰闈㈠妯″瀷杩涜淇壀锛汣ompressing锛氬鏉冮噸鍏变韩锛坈lustering based weight sharing锛夋垨瀵规潈閲嶉噰鐢℉uffman缂栫爜绛夛紱Low-bit representing 锛氬鏉冮噸閲忓寲锛坬uantization锛夊皢娴偣鍨嬭浆鎹㈠埌[0~255]锛屾垨鑰呭皢缃戠粶浜屽€硷紙binary锛夊寲绛夈€?
鎴戜滑杩欓噷閫夋嫨锛屼粠鍩轰簬涓嶅悓鐨勫嵎绉绠楁柟寮忔瀯閫犳柊鐨勭綉缁滅粨鏋勮繖涓柟鍚戜笅锛屾寫閫変竴浜涘氨杩戠殑鎴愭灉鏉ヨ繘琛屽垎鏋愪笌鎬荤粨銆傝繖鏍烽€夋嫨鐨勭悊鐢辨槸鍥犱负锛屼笂杩?杩欎釜鏂瑰悜鏄В鍐抽棶棰樼殑涓€绉嶆洿涓烘湰鐪熺殑鏂瑰紡銆傚畠鏄竴绉嶆洿涓洪潬杩戝簳灞傜殑瑙ｅ喅闂鐨勬€濊矾锛岃€屼笂杩?杩欑鏂瑰紡鏄彲浠ヨ璁や负鏄互1涓哄熀纭€鑰岃繘琛岀殑鐩稿叧鎷撳睍銆?/p>
3 鍩轰簬涓嶅悓鐨勫嵎绉绠楁柟寮忔瀯閫犵殑Light Weight CNN妯″瀷
澶ч噺鐨勫嵎绉缁忕綉缁滃潎閲囩敤浜嗙浉浼肩殑璁捐鏂瑰紡锛屽嵆浣跨敤鍙彉鐨勫嵎绉眰銆佹睜鍖栧眰銆佸啀鍔犱竴浜涘皬鏁伴噺鐨勫叏杩炴帴灞傘€傚湪濡備綍澧炲己杩欑璁捐鏂瑰紡涓嬬殑缃戠粶妯″瀷鐨勬€ц兘涓婏紝寰堝浜鸿繘琛屼簡涓嶅悓鐨勬帰绱細1锛変娇鐢ㄦ洿澶嶆潅鐨勬縺娲诲嚱鏁帮紝浣跨敤鏀瑰杽鐨勬鍒欏寲锛坮egularization锛夛紝浠ュ強鍒╃敤鏍囩淇℃伅杩涜layer-wise鐨勯璁粌 锛?锛変娇鐢ㄤ笉鍚岀殑CNN鏋舵瀯 銆侫ll Convolution Net鐨勪綔鑰呴€氳繃鍦ㄤ笉鍚岀殑鏁版嵁闆嗕笂杩涜瀹為獙鍚庯紝鍙戠幇浠呬粎浣跨敤鍗风Н灞傜殑缃戠粶缁撴瀯骞朵笉浼氬鐗╀綋妫€娴嬬殑鎬ц兘浜х敓褰卞搷锛屾睜鍖栧眰鐨勫瓨鍦ㄥ苟闈炲繀瑕侊紝鍙互浣跨敤姝ラ暱杈冨ぇ鐨勫嵎绉眰杩涜鏇夸唬銆傚洜姝わ紝All Convolution Net鎶涘純浜嗕互寰€CNN缃戠粶涓殑姹犲寲灞傚拰鍏ㄨ繛鎺ュ眰锛岄€氳繃浣跨敤姝ラ暱鏇村ぇ鐨勫嵎绉眰鏉ヤ唬鏇挎睜鍖栦互鍙婁娇鐢ㄥ嵎绉牳涓?鐨勫嵎绉眰鏉ヤ唬鏇垮叏杩炴帴灞傘€?br/>瀵逛簬姹犲寲灞傜殑浣滅敤鐜板湪杩樺緢闅剧粰鍑烘瘮杈冨畬鏁寸殑瑙ｉ噴锛屼竴鑸亣瀹氭睜鍖栧眰鍙互閫氳繃濡備笅涓変釜鏂归潰鏉ュCNN鐨勬€ц兘浜х敓甯姪锛?锛塸-norm锛坧鑼冩暟锛変娇CNN鐨勮〃绀烘洿鍏蜂笉鍙樻€э紙invariance锛夛紱2锛夐檷缁翠娇楂樺眰鑳藉瑕嗙洊杈撳叆灞傜殑鏇村閮ㄥ垎锛坮eceptive field锛夛紱3锛夋睜鍖栫殑feature-wise鐗规€ц兘澶熶娇寰椾紭鍖栨洿涓哄鏄撱€傚亣璁剧2锛夌偣锛屽嵆闄嶇淮瀵逛笌CNN鐨勬€ц兘鎻愬崌鑷冲叧閲嶈锛岄偅涔堟垜浠彲浠ラ€氳繃浣跨敤濡備笅涓ょ鏂规硶锛屾潵浠ｆ浛姹犲寲灞傚彇寰楃浉浼肩殑闄嶇淮鏁堟灉锛?锛夌洿鎺ョЩ闄ゆ睜鍖栧眰骞跺澶у嵎绉眰鐨勬闀匡紱2锛変娇鐢ㄦ闀垮ぇ浜?鐨勫嵎绉眰鏉ヤ唬鏇挎睜鍖栧眰銆傜1锛夌鏂规硶绛変环浜庢睜鍖栨搷浣滀絾浠呰€冭檻浜嗛《閮ㄥ乏渚х殑鐗瑰緛鍝嶅簲锛屽洜鑰屽彲鑳戒細闄嶄綆妫€娴嬬殑鍑嗙‘鐜囥€傜浜岀鍒欐病鏈夎繖绉嶇己闄枫€?br/>鏈€缁圓ll Convolution Net妯″瀷鐩歌緝浜庝互寰€CNN妯″瀷鐨勪笉鍚屼箣澶勬湁锛?锛変娇鐢╯tride澶т簬1鐨勫嵎绉眰浠ｆ浛浠ュ線CNN涓殑姹犲寲灞傦紙涓嬮噰鏍峰眰锛夛紱2锛変娇鐢╢ilter澶у皬涓?*1鐨勫嵎绉眰浠ｆ浛鍏ㄨ繛鎺ュ眰锛堝噺灏戝弬鏁板拰璁＄畻閲忥級銆傚叾鍦╟ifar-10鍜宑ifar-100涓婂彇寰椾簡寰堝ソ鐨勬晥鏋滐紝瑙佷笅鍥俱€?br/>
SqueezeNet閬靛惊浜嗕笁涓璁″師鍒欙細1锛変娇鐢ㄦ洿灏忕殑11鐨勫嵎绉牳鏉ヤ唬鏇?3鐨勫嵎绉牳銆傜被浼间簬GoogleNet[5]浠ュ強ResNet[6]璁捐涓殑33鍗风Н鏉ヤ唬鏇緼lexNet涓殑77鍗风Н鐨勬€濊矾锛孲queezeNet杩涗竴姝ョ敤11鍗风Н浠ｆ浛浜?3鍗风Н銆備负浜嗕繚璇佷笉褰卞搷璇嗗埆鐨勭簿搴︼紝鍥犳鍦⊿queezeNet涓苟涓嶆槸瀹屽叏鐨勪唬鏇匡紝鑰屾槸杩涜浜嗛儴鍒嗙殑鏇挎崲銆?锛夊噺灏戣緭鍏?3鍗风Н鐨勭壒寰佸浘鏁伴噺銆傚鏋滄槸conv1鍒癱onv2杩欐牱鐨勭洿鎺ヨ繛鎺ワ紝瀹為檯涓婃槸涓嶅お濂藉噺灏戣緭鍏ュ埌conv2鐨勭壒寰佸浘鐨勬暟閲忕殑锛屽湪SqueezeNet涓€氳繃鍒╃敤11鍗风Н鐨勫崌缁存垨闄嶇淮鐨勪綔鐢紝閲囩敤11鍗风Н鐢熸垚鏂扮殑鏁伴噺鐨勭壒寰佸浘锛岀劧鍚庡皢浠栦滑鎺ュ叆鍒?3鐨勫嵎绉紝杩涜€岃揪鍒扮洰鏍囥€傚湪SqueezeNet涓皢杩欑鎬濊矾灏佽鎴愪簡涓€涓狥ire Modeule锛?瑙佷笅鍥俱€?锛夊噺灏憄ooling銆傚湪锛?锛変腑鐨凙ll Convolution Net锛孏oogleNet浠ュ強ResNet涓兘鍙戠幇锛岄€傚綋鐨勫噺灏憄ooling鎿嶄綔鑳藉寰楀埌姣旇緝濂界殑鏁堟灉锛屽洜姝ゅ湪SqueezeNet涓彧鏄繘琛屼簡3娆ax pooling鍜?娆lobal pooling銆?/p>
MobileNet鎻愬嚭浜嗕竴绉峝epth-wise separable convolution鍗风Н鐨勬柟寮忔潵浠ｆ浛浼犵粺鍗风Н鐨勬柟寮忥紝 depth-wise separable convolution鍖呭惈涓ょ鎿嶄綔锛?锛塪epth-wise convolution锛?锛塸oint-wise convolution銆傞鍏堟彁鍑篸epth-wise convolution杩欑鍗风Н鏂瑰紡鐨勬槸璁烘枃[21]锛岃繖绉嶅嵎绉柟寮忕殑濂藉鏄噺灏戝弬鏁版暟閲忕殑鍚屾椂涔熸彁鍗囦簡璁＄畻鏁堢巼銆備笉鍚屼簬浼犵粺鐨勫嵎绉绠楋紝灏嗕竴涓嵎绉牳鍦ㄦ墍鏈夌殑杈撳叆閫氶亾涓婂仛鍗风Н鎿嶄綔锛屽湪depth-wise convolution涓竴涓嵎绉牳鍙湪涓€涓緭鍏ラ€氶亾涓婅繘琛屽嵎绉€備絾鏄痙epth-wise convolution鍙槸瀵硅緭鍏ラ€氶亾杩涜鍗曠嫭鐨勭壒寰佽绠楋紝骞舵病鏈夊皢浠栦滑缁撳悎璧锋潵璁＄畻鐗瑰緛銆傚洜姝わ紝涓轰簡浜х敓涓€浜涜仈鍚堢壒寰侊紝澧炲姞浜嗕竴灞傜綉缁滈噰鐢?*1鍗风Н鏉ュdepth-wise convolution鐢熸垚鐨勭壒寰佷箣闂磋繘琛屼竴涓嚎鎬х粍鍚堬紝杩欏氨鏄痯oit-wise convolution銆?/p>
ShuffleNet鍙互琚湅浣滄槸浠嶮obileNet鍙戝睍鑰屾潵鐨勩€傞鍏堬紝ShuffleNet鐨勪綔鑰呭彂鐜皃oint-wise convolution杩欑鎿嶄綔瀹為檯涓婃槸闈炲父鑰楁椂鐨勶紝涓轰簡鑳藉楂樻晥鐨勫湪杈撳叆鐗瑰緛鍥鹃棿寤虹珛淇℃伅娴侀€氾紝浠栦滑鍩轰簬group convolution 鎻愬嚭浜嗕竴绉峜hannel shuffle鐨勬搷浣溿€傞€氳繃鍒╃敤group convolution鍜宑hannel shuffle杩欎袱涓搷浣滄潵璁捐鍗风Н绁炵粡缃戠粶妯″瀷锛屽湪鍑忓皯浜嗗弬鏁扮殑鍚屾椂涔熻兘澶熸湁鏁堟彁楂樿绠楁晥鐜囥€?br/>Group convolution 鏄竴缁勫嵎绉牳璐熻矗涓€缁勮緭鍏ラ€氶亾锛屼篃灏辨槸姣忕粍閫氶亾鍙涓€缁勫嵎绉牳鍗风Н銆傞偅depth-wise convolution 鍏跺疄鏄彲浠ョ湅鎴愭槸涓€涓壒娈婄殑 group convolution锛屽嵆姣忎竴涓€氶亾鏄竴缁勩€侴roup convolution鏃╁湪AlexNet涓氨宸茬粡鍑虹幇浜嗭紝褰撴椂鏄洜涓虹‖浠堕檺鍒惰€岄噰鐢ㄥ垎缁勫嵎绉殑銆備箣鍚庡湪 2016 骞寸殑 ResNeXt [9] 涓紝涔熻〃鏄庨噰鐢?group convolution 鍙互鑾峰緱楂樻晥鐨勭綉缁溿€傜敱浜庡熀浜庝笉鍚岀粍璁＄畻鍑烘潵鐨勭壒寰佸浘涔嬮棿鏄病鏈変俊鎭祦閫氱殑锛岃繖浼氬奖鍝嶇綉缁滅殑琛ㄨ揪锛屼负浜嗗缓绔嬬粍闂寸殑淇℃伅娴侀€氾紝channel shuffle鎿嶄綔灏辫寮曞叆杩涙潵銆侰hannel shuffle鎿嶄綔瀹為檯涓婃槸瀵筭roup convolution鎿嶄綔璁＄畻鍑烘潵鐨勭壒寰佸浘缁勮繘琛屼竴涓噸缁勶紝涔熷氨鏄粠鍚勪釜鐗瑰緛鍥剧粍涓殢鏈虹殑閫夋嫨涓€涓壒寰佸浘锛岃繘鑰屾瀯鎴愭柊鐨勭壒寰佸浘缁勩€傛瘡涓€缁勮繖绉嶆柊鐨勭壒寰佸浘缁勪腑閮藉寘鍚簡鏃х殑鐗瑰緛鍥剧粍涓殑淇℃伅鐨勶紝涔熷氨鏄湪鏃х殑鐗瑰緛鍥剧粍闂村缓绔嬩簡淇℃伅娴侀€氥€傚悓鏃讹紝channel shuffle鎿嶄綔涔熼伩鍏嶄簡MobileNet涓殑point-wise convolution鐨勮绠楅噺銆侰hannel shuffle鎿嶄綔鐨勭暐缂╁浘濡備笅鎵€绀恒€?/p>
4 鎬荤粨
1锛? 鍑忓皯姹犲寲鎿嶄綔锛?br/>2锛? 浣跨敤杈冨皬灏哄害鐨勫嵎绉牳锛?br/>3锛? 閲囩敤depth-wise convolution鏄璁¤交閲忓寲妯″瀷鐨勯潪甯搁噸瑕佺殑鎶€鏈紝浣嗘槸瑕佹敞鎰忚В鍐充俊鎭笉娴侀€氱殑闂銆?br/>鏈€鍚庯紝鐢变簬姘村钩鏈夐檺锛屽嚭閿欎箣澶勶紝杩樿澶氬鎸囨暀锛?/p>
Reference
[2] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman, 鈥淩eturn of the Devil in the Details: Delving Deep into Convolutional Nets,鈥?in British Machine Vision Conference (BMVC), 2014.
[3] Karen Simonyan and Andrew Zisserman, 鈥淰ery Deep Convolutional Networks for Large-Scale Image Recognition,鈥?in CoRR, 2014. URL http://arxiv.org/abs/1409.1556.
[4] Artem Babenko, Anton Slesarev, Alexandr Chigorin, and Victor Lempitsky, 鈥淣eural Codes for Image Retrieval,鈥?in European Conference on Computer Vision (ECCV), volume 8689, pages 584鈥?99, 2014.
[5] Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, Erhan D, Vanhoucke V, Rabinovich A, 鈥淕oing deeper with convolutions,鈥?in CVPR, 7-12 June 2015, doi: 10.1109/CVPR.2015.7298594.
[6] He K, Zhang X, Ren S, Sun J, 鈥淒eep residual learning for image recognition,鈥?in Computer Vision and Pattern Recognition, 2016.
[7] C. Szegedy, S. Ioffe, V. Vanhoucke, and A. Alemi, 鈥淚nceptionv4, inception-resnet and the impact of residual connections on learning,鈥?arXiv preprint arXiv:1602.07261, 2016.
[8] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, 鈥淩ethinking the inception architecture for computer vision,鈥?in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2818鈥?826, 2016.
[9] S. Xie, R. Girshick, P. Doll鈥瞐r, Z. Tu, and K. He, 鈥淎ggregated residual transformations for deep neural networks,鈥?arXiv preprint arXiv:1611.05431, 2016.
[10] J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. A. Riedmiller, 鈥淪triving for simplicity: The all convolutional net,鈥?in CoRR, vol. abs/1412.6806, 2014.
[11] F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally, and K. Keutzer, 鈥淪queezenet: Alexnet-level accuracy with 50x fewer parameters and< 0.5 mb model size,鈥?in ICLR, 2017.
[12] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, 鈥淢obilenets: Efficient convolutional neural networks for mobile vision applications,鈥?in CVPR, 2017.
[13] Zhang X, Zhou X, Lin M, and Sun J, 鈥淪hufflenet: An extremely efficient convolutional neural network for mobile devices,鈥?in CVPR 2017. arXiv preprint arXiv:1707.01083
[14] M. Jaderberg, A. Vedaldi, and A. Zisserman, 鈥淪peeding up convolutional neural networks with low rank expansions,鈥?arXiv preprint arXiv:1405.3866, 2014.
[15] X. Zhang, J. Zou, X. Ming, K. He, and J. Sun, 鈥淓fficient and accurate approximations of nonlinear convolutional networks,鈥?in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1984鈥?992, 2015.
[16] J. Wu, C. Leng, Y. Wang, Q. Hu, and J. Cheng, 鈥淨uantized convolutional neural networks for mobile devices,鈥?arXiv preprint arXiv:1512.06473, 2015.
[17] S. Han, H. Mao, and W. J. Dally, 鈥淒eep compression: Compressing deep neural network with pruning, trained quantization and huffman coding,鈥?in CoRR, abs/1510.00149, 2, 2015.
[18] M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi, 鈥淴nornet: Imagenet classification using binary convolutional neural networks,鈥?in European Conference on Computer Vision, pages 525鈥?42. Springer, 2016.
[19] Luo JH, Wu J, and Lin W, 鈥淭hinet: A filter level pruning method for deep neural network compression,鈥?in ICCV 2017.
[20] Evan Shelhamer, Jonathan Long, Trevor Darrell, 鈥淔ully Convolutional Networks for Semantic Segmentation,鈥?in CVPR, 2015.
[21] L. Sifre, 鈥淩igid-motion scattering for image classification,鈥?hD thesis, Ph. D. thesis, 2014.
[22] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, 鈥淚magenet: A large-scale hierarchical image database,鈥?in Proc. CVPR, 2009.
以上是关于Light Weight CNN妯″瀷鐨勫垎鏋愪笌鎬荤粨的主要内容，如果未能解决你的问题，请参考以下文章
Elasticsearch鐨勫垎鏋愯繃绋?鍐呯疆瀛楃杩囨护鍣ㄣ€佸垎鏋愬櫒銆佸垎璇嶅櫒銆佸垎璇嶈繃婊ゅ櫒锛堢湡鏄彉鎬佸鍟婏紒缇庢粙婊嬶級
缁濆湴姹傜敓娓告垙鏁版嵁鍙鍖栧垎鏋愶細濡備綍绋崇ǔ鍚冨埌楦★紵
Spark + Hadoop,鍩轰簬WIFI鎺㈤拡鐨勫ぇ鏁版嵁鍒嗘瀽绯荤粺