interpret

InterpretDL实现ResNet的可解释性分析

InterpretingResNetModelWithLRP(Layer-wiRelevancePropagation)

使⽤Layer-wiRelevancePropagation对ResNet和VGG⽹络进⾏可解释的分析。

StarandFork!

本项⽬的所有代码和tutorials均源于PaddlePaddle的仓库，欢迎⼤家Star和Fork！

InterpretDL:InterpretationofDeepLearningModelsbadonPaddlePaddle

InterpretDL,shortforinterpretationsofdeeplearningmodels,olkit

containsimplementationsofmanyinterpretationalgorithms,includingLIME,Grad-CAM,IntegratedGradientsandmore.

SomeSOTAandnewinterpretationalgorithmsarealsoimplemented.

InterpretDLisunderactiveconstructionandallcontributionsarewelcome!

克隆仓库

#下载⽐较慢附上了压缩包

#!gitclone/PaddlePaddle/

%cd~

!unzip-oqdata/data98915/

%cdInterpretDL/tutorials

/home/aistudio

/home/aistudio/InterpretDL/tutorials

#安装依赖

!pipinstallscikit-image

!all

Layer-wiRelevancePropagation(LRP)isanexplanationtechniqueapplicabletomodelsstructuredasneuralnetworks,

,videos,ortext.

LRPoperatesbypropagatingthepredictionbackwardsintheneuralnetwork,bymeansofpurpolydesignedlocal

propagationrules.

tice,state-of-the-artneuralnetworkssuch

asResNetmakeuofmorecomplexlayerssuchasconvolutionsandpooling.

Inthisca,LRPrulesaremoreconvenientlyimplementedbycastingtheoperationsofthefour-stepprocedureaboveas

perationsarereadilyavailableinneuralnetworkframeworkssuch

asPaddlePaddleandPytorch,andcanthereforebereudforthepurpoofimplementingLRP.

Here,wetaketheResNetpretrainednetworkforimageclassification.

LRPRulesforDeepRectifierNetworks

BasicRule(LRP-0)

Thisruleredistributesinproportionto:

EpsilonRule(LRP-)

f(x)

R=jRk

∑

aw∑0,jjjk

awjjk

k

ϵ

AfirstenhancementofthebasicLRP-0ruleconsistsofaddingasmallpositiveterminthedenominator:

Theroleofistoabsorbsomerelevancewhe

becomeslarger,picallyleadstoexplanationsthat

aresparrintermsofinputfeaturesandlessnoisy.

GammaRule(LRP-)

Anotherenhancementwhichweintroducehereisobtainedbyfavoringtheeffectofpositivecontributionsovernegative

contributions:

eas,negativecontributionsstartto

valenceofpositivecontributionshasalimitingeffectonhowlargepositiveandnegativerelevancecan

lpstodelivermorestableexplanations.

ImplementingLRPEfficiently

erthegenericrule

putationofthisropagationrulecanbedecompodinfoursteps:

KaTeXparerror:Nosuchenvironment:align*atposition8:begin{align*}forall_k:~z_k…

Thefirststepisaforwardpassonacopyofthelayerwheretheweightsandbiashavebeenappliedthemap,to

third

step,onenotesthatcanalsobeexpresdasthegradientcomputation:

whereisthevectoroflower-layeractivations,whereisafunctionofit,andwhereskisinsteadtreatedas

adientcanbecomputedviaautomaticdifferentiation,whichisavailableinmostneuralnetworkslibraries.

Inthisrepository,weurule:

Youcantpropervalueofandtogettheruleorruleifyouneed.

R=jRk

∑

ϵ+aw∑0,jjjk

awjjk

k

ϵ

ϵ

γ

R=jRk

∑

a⋅(w+γw)∑0,jjjkjk

+

a⋅(w+γw)jjkjk

+

k

γγ

R=jRk

∑

ϵ+aρ(w)∑0,jjjk

aρ(w)jjk

k

LRP−0/ϵ/γ

θ→ρ(θ)

ϵ

cj

c=j[∇(z(a)⋅s))]k

∑

kk

j

a=(a)jjzk

LRP−αβ

LRP−αβ

R=j(α−β)Rk

∑

(aw)∑0,jjjk

+

(aw)jjk

+

(aw)∑0,jjjk

−

(aw)jjk

−

k

αβLRP−0LRP−γ

importpaddle

importinterpretdlasit

Loadinputimage

fromPILimportImage

_simportread_image

img_path='asts/'

x=ray(read_image(img_path)[0])

x

Loadlrpmodel

Supportedlrpmodel

resnet18

resnet34

resnet50

resnet101

resnet152

vgg16

More…

Loadmodelresnet50

_modelimportresnet50

paddle_model=resnet50(pretrained=True)

#paddle_model

100%|██████████|151272/151272[00:03<00:00,47730.41it/s]

LoadLRPinterperter

lrp=nterpreter(paddle_model,u_cuda=Fal)

Gettheinterpretationresult

Gettheinterpretationresultfromthemaximumproblabel

R=ret(

img_path,

label=None,

visual=True,

save_path=None)

bullmastiff(243):0.429

Brabancongriffon(262):0.048

tiger,Pantheratigr(292):0.044

boxer(242):0.038

doormat,welcomemat(539):0.029

tabby,tabbycat(281):0.026

tigercat(282):0.023

pug,pug-dog(254):0.022

Egyptiancat(285):0.013

bloodhound,sleuthho(163):0.012

Ulabel:243,bullmastiff

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/cbook/__init__.py:2349:DeprecationWarning:UsingorimportingtheABC

sfrom'collections'insteadoffrom''isdeprecated,andin3.8itwillstopworking

ifisinstance(obj,or):

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/cbook/__init__.py:2366:DeprecationWarning:UsingorimportingtheABC

sfrom'collections'insteadoffrom''isdeprecated,andin3.8itwillstopworking

returnlist(data)ifisinstance(data,gView)eldata

tlabeltotigercat(282)

R=ret(

img_path,

label=282,

visual=True,

save_path=None)

bullmastiff(243):0.429

Brabancongriffon(262):0.048

tiger,Pantheratigr(292):0.044

boxer(242):0.038

doormat,welcomemat(539):0.029

tabby,tabbycat(281):0.026

tigercat(282):0.023

pug,pug-dog(254):0.022

Egyptiancat(285):0.013

bloodhound,sleuthho(163):0.012

Ulabel:282,tigercat

Changemodeltoresnet101

_modelimportresnet101

paddle_model=resnet101(pretrained=True)

100%|██████████|263160/263160[00:04<00:00,63094.35it/s]

lrp=nterpreter(paddle_model,u_cuda=Fal)

R=ret(

img_path,

label=None,

visual=True,

save_path=None)

tigercat(282):0.483

tabby,tabbycat(281):0.212

tiger,Pantheratigr(292):0.171

Egyptiancat(285):0.083

bullmastiff(243):0.017

doormat,welcomemat(539):0.007

boxer(242):0.003

spaceheater(811):0.003

lynx,catamount(287):0.002

radiator(753):0.001

Ulabel:282,tigercat

tlabeltobullmastiff(243)

R=ret(

img_path,

label=243,

visual=True,

save_path=None)

tigercat(282):0.483

tabby,tabbycat(281):0.212

tiger,Pantheratigr(292):0.171

Egyptiancat(285):0.083

bullmastiff(243):0.017

doormat,welcomemat(539):0.007

boxer(242):0.003

spaceheater(811):0.003

lynx,catamount(287):0.002

radiator(753):0.001

Ulabel:243,bullmastiff

Changemodeltovgg16

_modelimportvgg16

paddle_model=vgg16(pretrained=True)

100%|██████████|817517/817517[00:13<00:00,62384.04it/s]

lrp=nterpreter(paddle_model,u_cuda=Fal)

R=ret(

img_path,

label=None,

visual=True,

save_path=None)

bullmastiff(243):0.620

tiger,Pantheratigr(292):0.306

GreatDane(246):0.071

pug,pug-dog(254):0.003

boxer(242):0.000

Leonberg(255):0.000

bloodhound,sleuthho(163):0.000

tigercat(282):0.000

Germanshepherd,Ger(235):0.000

Brabancongriffon(262):0.000

Ulabel:243,bullmastiff

tlabeltotigercat(282)

R=ret(

img_path,

label=282,

visual=True,

save_path=None)

bullmastiff(243):0.620

tiger,Pantheratigr(292):0.306

GreatDane(246):0.071

pug,pug-dog(254):0.003

boxer(242):0.000

Leonberg(255):0.000

bloodhound,sleuthho(163):0.000

tigercat(282):0.000

Germanshepherd,Ger(235):0.000

Brabancongriffon(262):0.000

Ulabel:282,tigercat

关于作者

袁廷翼

西安交通⼤学2020级研究⽣控制科学与⼯程

百度⼤数据实验室实习⽣

感兴趣的⽅向为：机器⼈,ml,cv,⽹络可解释性

AIstudio主页：

Wechat：yty___

有问题欢迎交流，喜欢可以三连⼀下！