Explain ResNet50 ImageNet classification using Partition explainer
This notebook demonstrates how to use SHAP to explain image classification models. In this example we are explaining the output of ResNet50 model for classifying images into 1000 ImageNet classes.
[1]:
import json
import numpy as np
from tensorflow.keras.applications.resnet50 import ResNet50, preprocess_input
import shap
Loading Model and Data
[2]:
# load pre-trained model and data
model = ResNet50(weights="imagenet")
X, y = shap.datasets.imagenet50()
[3]:
# getting ImageNet 1000 class names
url = "https://s3.amazonaws.com/deep-learning-models/image-models/imagenet_class_index.json"
with open(shap.datasets.cache(url)) as file:
class_names = [v[1] for v in json.load(file).values()]
# print("Number of ImageNet classes:", len(class_names))
# print("Class names:", class_names)
SHAP ResNet50 model explanation for images
Build a partition explainer with: - the model (a python function) - the masker (a python function) - output names (a list of names of the output classes)
A quick run with a few evaluations
[8]:
# python function to get model output; replace this function with your own model function.
def f(x):
tmp = x.copy()
preprocess_input(tmp)
return model(tmp)
# define a masker that is used to mask out partitions of the input image.
masker = shap.maskers.Image("inpaint_telea", X[0].shape)
# create an explainer with model and image masker
explainer = shap.Explainer(f, masker, output_names=class_names)
# here we explain two images using 500 evaluations of the underlying model to estimate the SHAP values
shap_values = explainer(X[1:3], max_evals=100, batch_size=50, outputs=shap.Explanation.argsort.flip[:4])
[ ]:
shap_values.shape
[11]:
np.tile(np.array(shap_values.output_names), 2)
[11]:
array(['American_egret', 'crane', 'little_blue_heron', 'flamingo',
'American_egret', 'crane', 'little_blue_heron', 'flamingo'],
dtype='<U17')
[7]:
np.array(shap_values.output_names).shape[0]
[7]:
4
[11]:
shap_values.output_dims
[11]:
(4,)
Explainer options:
Above image masker uses a blurring technique called “inpaint_telea”. There are alternate masking options available to experiment with such as “inpaint_ns” and “blur(kernel_xsize, kernel_xsize)”.
Recommended number of evaluations is 300-500 to get explanations with sufficient granularity for the super pixels. More the number of evaluations, more the granularity but also increases run-time.
Note:
outputs=shap.Explanation.argsort.flip[:4]has been used in the code above for getting SHAP values because we want to get the top 4 most probable classes for each image i.e. top 4 classes with decreasing probability. Hence, a flip argsort sliced by 4 has been used.
Visualizing SHAP values output
[6]:
# output with shap values
shap.image_plot(shap_values)
[24]:
shap_values[0].base_values.shape
[24]:
(1000,)
[31]:
shap_values[0].base_values
[31]:
array([3.98265023e-04, 4.48955398e-05, 5.20897702e-05, 2.04479627e-04,
5.67642674e-05, 1.37821451e-04, 1.43101468e-04, 2.34555133e-04,
1.65241177e-03, 8.71808079e-05, 2.73833750e-04, 1.03291946e-04,
1.03254220e-04, 2.66912321e-05, 9.38569065e-05, 4.04711056e-04,
3.30887298e-04, 8.17410037e-05, 4.58840441e-05, 7.76972214e-04,
9.79037606e-04, 1.82875578e-04, 6.26673936e-05, 4.15584684e-04,
1.75147157e-04, 2.43703653e-05, 1.90513470e-04, 4.63949953e-04,
3.72401183e-03, 2.16879227e-04, 4.97241551e-03, 2.01521367e-01,
4.61910939e-04, 3.70553578e-04, 1.67468010e-04, 3.31375678e-03,
1.03218982e-03, 1.10009874e-04, 4.62198048e-04, 2.72900041e-04,
1.63967852e-04, 1.55510512e-04, 7.61000774e-05, 8.27621043e-05,
1.24086940e-03, 5.44384784e-05, 1.88978534e-04, 1.65600795e-03,
4.68490616e-04, 2.02517796e-04, 2.24879230e-04, 1.53284360e-04,
7.28016312e-04, 2.18322930e-05, 9.70595356e-05, 2.40055524e-04,
8.75858896e-05, 8.05352684e-05, 5.24058286e-03, 4.47012542e-04,
1.37802403e-04, 5.53620885e-05, 1.67022704e-03, 2.84969807e-04,
8.81938729e-04, 1.68564642e-04, 1.31366367e-04, 9.56201809e-04,
2.78153806e-04, 7.04826234e-05, 1.15577168e-04, 5.03735617e-04,
3.23968381e-03, 2.96114665e-03, 2.07320263e-04, 1.06132834e-03,
2.17621040e-04, 7.06988329e-04, 1.62530720e-04, 2.02930160e-03,
8.79141953e-05, 1.91258034e-03, 1.58560916e-03, 1.76169444e-03,
6.13501586e-04, 1.01693309e-04, 1.22117897e-04, 2.70858320e-04,
3.35622499e-05, 1.05499045e-03, 4.14801470e-05, 1.82900709e-04,
5.84131012e-05, 1.04882440e-03, 2.19639414e-03, 1.43847588e-04,
8.00978305e-05, 2.61449284e-04, 5.75504149e-04, 1.17039803e-04,
3.48173926e-04, 9.26456341e-05, 2.51709687e-04, 4.78086266e-04,
2.59803119e-03, 1.59291131e-03, 1.66199796e-04, 1.56387687e-04,
1.11853424e-03, 1.68977969e-03, 2.42620936e-05, 1.18718512e-04,
3.65030882e-03, 5.25807380e-04, 2.50643864e-03, 4.63505945e-04,
3.17113445e-05, 2.52411381e-04, 8.77027574e-04, 5.45950279e-05,
8.71916054e-05, 4.65254387e-04, 9.42074548e-05, 3.14210588e-03,
6.98393560e-05, 1.40927164e-04, 2.92200624e-04, 1.27284331e-02,
1.96964378e-04, 3.43363092e-04, 6.04275789e-04, 2.46494092e-05,
4.61999938e-04, 2.57244479e-04, 4.37212363e-03, 1.17149972e-03,
4.89713420e-05, 3.27316113e-04, 4.47746861e-04, 2.02462077e-04,
8.61583801e-04, 1.34167858e-04, 3.71429429e-04, 6.87520718e-04,
1.40331360e-03, 1.22133445e-03, 1.66886639e-05, 1.42857461e-04,
1.14866011e-02, 3.36653134e-03, 3.29846021e-04, 4.72038257e-04,
1.00409007e-03, 1.29732513e-03, 3.61699349e-04, 5.30289579e-03,
5.18341956e-04, 4.70881874e-04, 6.44493266e-05, 1.05675921e-04,
2.06455775e-03, 2.58339016e-04, 4.76116576e-04, 1.20431083e-04,
1.24215968e-02, 2.86966406e-05, 7.48898237e-05, 1.71770240e-04,
1.53026136e-04, 2.45231204e-04, 1.14865397e-04, 3.22352367e-04,
3.37602425e-04, 6.14294258e-04, 9.70410256e-05, 1.59063807e-03,
5.08750509e-03, 2.67939991e-04, 1.43872213e-03, 1.19161239e-04,
2.46804469e-04, 2.98232539e-04, 4.77343885e-04, 4.77507783e-05,
4.32613585e-03, 1.01641439e-04, 6.29403366e-05, 3.28598602e-04,
4.87449579e-03, 1.87227430e-04, 5.53444610e-04, 5.19249064e-04,
1.10896956e-03, 3.43542802e-03, 1.74651941e-04, 2.22513816e-04,
2.95794592e-03, 2.09766687e-04, 4.75166926e-05, 2.88550509e-04,
6.87570893e-04, 2.06802375e-04, 1.71519205e-04, 2.70951336e-04,
9.43710693e-05, 1.12773397e-03, 9.82338935e-03, 3.32266139e-03,
2.19096444e-04, 2.17689230e-05, 1.61327654e-04, 1.06600019e-04,
1.96517725e-03, 5.96507452e-04, 7.03871483e-04, 1.59394997e-03,
2.64894590e-03, 1.60959898e-04, 6.20000577e-03, 6.26118344e-05,
1.64568322e-04, 9.34288691e-05, 1.06487481e-04, 2.47626798e-04,
7.09660351e-03, 1.91016507e-03, 7.00892007e-04, 2.34331994e-04,
2.19726609e-03, 1.90845807e-04, 4.88849764e-04, 5.34049235e-03,
1.82612066e-03, 1.67027392e-04, 1.59819072e-04, 5.94002078e-04,
1.34168877e-04, 9.71634930e-04, 2.19330061e-04, 2.32918101e-04,
2.32211058e-03, 8.56369548e-03, 1.04254432e-04, 1.94690569e-04,
1.75361261e-02, 3.74004652e-04, 2.08059035e-04, 1.98756400e-02,
3.35199817e-04, 7.62398820e-04, 4.94323089e-04, 9.44103158e-05,
7.72377680e-05, 2.18191394e-03, 9.12886608e-05, 1.49337298e-04,
5.48757566e-03, 2.28227451e-04, 2.91966484e-04, 3.75757751e-04,
3.29240400e-04, 4.76886606e-04, 1.14875911e-04, 2.97170162e-04,
5.42078109e-04, 1.10409739e-04, 9.61581129e-04, 2.99977983e-05,
7.14230409e-04, 2.44811999e-05, 1.56077331e-05, 8.35507381e-05,
1.11137240e-04, 2.19001100e-04, 1.94354588e-03, 3.41678562e-04,
2.43427046e-03, 3.74987503e-05, 6.12538948e-04, 3.07665410e-04,
9.51292086e-03, 1.96635447e-04, 2.98951869e-03, 1.51283562e-03,
1.11732305e-04, 3.72940012e-05, 1.21334312e-03, 4.69207989e-05,
1.40990349e-04, 3.28055816e-04, 6.17932164e-05, 9.55878990e-04,
7.26847284e-05, 2.80747423e-04, 7.74153834e-03, 2.19893031e-04,
1.46946550e-04, 8.18398767e-05, 9.75400544e-05, 3.98906486e-05,
9.15985182e-03, 6.38652884e-04, 8.55517923e-04, 5.58320142e-04,
2.84550682e-04, 6.79018209e-04, 1.22444949e-03, 1.38428388e-02,
9.44111234e-05, 9.13376061e-05, 1.90601801e-04, 2.07817706e-04,
9.78715834e-04, 5.39315515e-04, 9.03732842e-04, 2.02737399e-04,
2.44897928e-05, 1.55207381e-04, 7.26549304e-04, 2.10206723e-04,
7.31480657e-04, 3.42042244e-04, 2.21371403e-04, 4.17837378e-04,
6.34400232e-04, 7.87110184e-05, 1.63991455e-04, 1.89242372e-03,
2.18088622e-03, 2.73842219e-04, 1.27485720e-04, 6.10379575e-05,
6.02986023e-04, 1.50342930e-05, 3.14068719e-04, 1.06195104e-03,
2.15661494e-04, 8.36609805e-04, 5.64410839e-05, 6.65351326e-05,
2.11620267e-04, 1.86994192e-04, 1.15989812e-03, 7.73980646e-05,
7.99473783e-05, 6.08985603e-04, 1.25390972e-04, 5.11682360e-03,
5.26337171e-05, 2.94670346e-04, 1.78296075e-04, 1.97027315e-04,
2.16452434e-04, 3.26410867e-04, 1.02113234e-04, 1.52413466e-03,
3.93337468e-05, 6.38024008e-04, 2.47433345e-04, 3.88692890e-04,
3.44268419e-03, 2.77661893e-05, 1.42515026e-04, 1.01953614e-04,
5.43837762e-03, 8.35901155e-05, 1.35738961e-03, 4.04788414e-04,
1.58322830e-04, 1.16253315e-04, 1.26562698e-03, 1.86057645e-04,
5.57620137e-04, 1.01343765e-04, 2.88358642e-05, 4.74227796e-04,
3.16448684e-04, 4.19742246e-05, 1.08359178e-04, 1.04418854e-04,
2.03641539e-04, 4.04073711e-04, 7.33445268e-05, 3.12642660e-03,
1.77940441e-04, 1.04366243e-03, 2.02374431e-04, 5.83104083e-05,
2.26815441e-03, 7.83597527e-04, 8.78596678e-04, 2.20475224e-04,
3.20527288e-05, 8.91639676e-04, 2.62770463e-05, 2.32063496e-04,
1.86159959e-04, 2.52828421e-03, 1.35955503e-04, 3.13804485e-05,
4.37561609e-03, 6.02974789e-04, 5.17095978e-05, 1.81379146e-04,
1.14109635e-03, 8.31345329e-04, 2.78152613e-04, 6.97735872e-04,
1.05109788e-03, 8.61559995e-04, 8.07119440e-03, 3.48262605e-04,
2.55202758e-05, 5.52419981e-04, 6.68813955e-05, 3.29635768e-05,
7.73706532e-04, 3.64235049e-04, 1.51698780e-03, 2.81690009e-04,
4.75254655e-03, 6.69377157e-04, 9.54573101e-04, 1.08042186e-04,
5.68635260e-05, 4.24618163e-04, 2.58263346e-04, 1.45405869e-03,
1.02591919e-04, 1.55952512e-04, 1.85346958e-04, 7.03747282e-05,
2.41437432e-04, 2.35192638e-04, 1.24183003e-04, 1.20975601e-04,
7.09347078e-05, 6.99766242e-05, 7.71764317e-05, 1.89607745e-04,
7.93690444e-04, 1.37692579e-04, 7.09580781e-05, 5.91546879e-04,
2.90565134e-04, 2.89046846e-04, 5.19869197e-03, 4.70197119e-04,
1.15571762e-04, 9.94998263e-04, 2.57296371e-04, 4.94676409e-03,
5.00711962e-04, 2.99575167e-05, 1.86757010e-04, 6.52849994e-05,
6.04235218e-04, 4.68017897e-05, 3.87711974e-04, 4.72599531e-05,
6.97334204e-03, 2.53278966e-04, 2.60601024e-04, 8.47871488e-05,
3.23877830e-05, 1.19668525e-03, 5.75865619e-03, 4.11593704e-04,
1.53028246e-04, 1.79754803e-04, 6.80635159e-04, 1.39376134e-04,
6.17261903e-05, 2.87276634e-04, 8.66111950e-05, 1.57140807e-04,
9.42889383e-05, 2.58421525e-04, 5.66013405e-05, 2.35475280e-04,
1.04127292e-04, 6.85896521e-05, 7.02898469e-05, 1.98832509e-04,
1.16414631e-04, 3.62440856e-04, 4.04755992e-04, 1.61688542e-04,
8.00234993e-05, 5.06251745e-05, 1.01898695e-04, 1.85255765e-03,
2.10804399e-03, 3.95909999e-04, 1.04698156e-04, 1.23862468e-04,
2.01316288e-04, 2.67114752e-04, 1.09853136e-04, 1.04562007e-02,
1.44658901e-04, 1.38847245e-04, 1.86406018e-04, 2.91111821e-04,
2.58459095e-05, 1.76549394e-04, 5.91578195e-04, 1.80155170e-04,
1.16277879e-04, 6.64511754e-05, 1.91455300e-04, 9.07513968e-05,
1.45928192e-04, 1.78181485e-03, 1.08011079e-03, 1.49311381e-04,
9.82756392e-05, 2.88707030e-04, 2.67076190e-04, 9.56205360e-04,
9.14820994e-05, 1.44630976e-04, 2.71266763e-04, 2.20545771e-04,
3.51296447e-04, 5.65011360e-05, 5.39898028e-05, 3.32881791e-05,
4.18267970e-04, 8.53816455e-04, 1.39340307e-04, 2.61821784e-04,
1.01256774e-04, 6.20011197e-05, 8.18510260e-03, 5.84557885e-04,
3.23629247e-05, 1.49878833e-04, 1.95337387e-04, 1.71919542e-04,
2.35342843e-04, 2.49581831e-03, 1.60848489e-04, 1.09985948e-03,
4.44245707e-05, 1.16778909e-04, 1.92636347e-04, 3.99644450e-05,
6.28446651e-05, 3.98166245e-03, 1.55952876e-04, 2.17046312e-04,
1.09251858e-04, 2.49226135e-03, 1.54479500e-03, 1.25159859e-04,
9.80652621e-05, 7.80950242e-04, 1.25729610e-04, 2.39668138e-04,
3.19829152e-04, 2.43557079e-04, 6.90597994e-03, 1.81299256e-04,
7.97690518e-05, 3.59173209e-05, 7.77634486e-05, 5.12411294e-04,
6.92587346e-04, 9.71015688e-05, 5.37150423e-04, 8.37979460e-05,
1.03847982e-04, 1.64556390e-04, 1.62095050e-04, 7.50828767e-05,
6.85978739e-04, 1.59797564e-04, 1.71349748e-04, 7.46525475e-05,
5.57506050e-04, 7.15658753e-05, 5.18313493e-04, 1.54518042e-04,
7.34301575e-04, 7.41808981e-05, 9.81214471e-05, 1.08507685e-02,
2.20979142e-04, 7.11828982e-03, 1.17054806e-05, 8.21940892e-04,
2.02488227e-05, 6.03395638e-05, 2.55094492e-04, 3.03323497e-04,
5.15690008e-05, 1.75295805e-04, 4.68559418e-04, 1.89823142e-04,
5.47616291e-05, 1.30891727e-04, 2.46041076e-04, 4.26332932e-04,
3.16890655e-04, 4.96258763e-05, 8.59340711e-04, 2.02530529e-04,
4.18328382e-05, 5.38278109e-05, 7.10692402e-05, 1.22153884e-04,
4.86796489e-04, 1.24314654e-04, 7.03770143e-04, 1.04051956e-03,
2.68163021e-05, 6.27119932e-03, 2.30516074e-03, 1.38077943e-04,
4.67849924e-04, 1.64890280e-05, 1.47823477e-04, 9.48119268e-05,
7.49706873e-04, 2.70575456e-05, 2.05066081e-05, 1.48544926e-03,
7.35011272e-05, 1.63713237e-03, 1.29237270e-03, 6.70079084e-04,
3.67125933e-04, 3.29222501e-04, 9.32080293e-05, 1.03922932e-04,
4.79535338e-05, 7.76475863e-05, 1.59768169e-04, 2.87516066e-03,
2.80911103e-03, 7.61366333e-04, 6.05647219e-04, 1.54264562e-04,
2.49425339e-05, 1.76321148e-04, 1.35979441e-03, 2.27449345e-04,
1.73271954e-04, 1.28269999e-03, 7.98516776e-05, 8.44012073e-04,
1.77738373e-04, 2.57999836e-05, 1.64547542e-04, 5.57913503e-04,
1.26685749e-03, 3.86202475e-04, 8.87230679e-04, 7.48189050e-04,
2.13030304e-04, 3.55858996e-04, 8.33153354e-06, 2.95320933e-04,
3.16009100e-05, 2.53893842e-04, 1.18534744e-03, 1.37425945e-04,
8.79636675e-04, 5.70305798e-04, 1.32686793e-04, 2.63076770e-04,
2.84897833e-04, 4.98322945e-04, 1.60985219e-05, 1.27237683e-04,
6.39092177e-04, 2.34655710e-03, 2.54381361e-04, 2.92099867e-05,
8.70297808e-05, 7.05914063e-05, 1.59767864e-04, 8.34528691e-05,
6.28060734e-05, 1.51177555e-05, 1.69857871e-04, 3.63609492e-04,
4.09535496e-05, 6.42370665e-04, 1.83896496e-04, 3.21555126e-05,
9.33920406e-03, 1.51379456e-04, 8.91392294e-04, 7.34808709e-05,
7.63745839e-03, 9.67634551e-05, 3.08589311e-03, 1.76709506e-03,
8.15805033e-05, 2.84980168e-04, 1.37901196e-04, 2.92042911e-04,
5.23895316e-04, 1.24685976e-04, 5.17050910e-04, 2.32969309e-04,
3.13187120e-05, 4.23460988e-05, 3.14876524e-04, 3.53547439e-05,
1.10661949e-03, 1.50967354e-03, 2.42085895e-04, 1.70345920e-05,
1.50748165e-04, 1.34549962e-04, 5.18803572e-05, 9.82871279e-04,
1.07740611e-03, 5.10502650e-05, 9.50180674e-06, 4.26059996e-05,
2.65619456e-04, 7.56132140e-05, 7.09234373e-05, 3.56290082e-04,
2.77788815e-04, 2.11004211e-04, 3.53443623e-03, 4.85317978e-05,
1.16430230e-04, 7.01543409e-04, 1.44020942e-05, 5.91689895e-04,
6.82438840e-05, 6.52912931e-05, 2.92932855e-05, 1.66858081e-04,
5.85521979e-04, 4.82931762e-04, 3.54350952e-04, 8.34171474e-03,
1.62974946e-04, 2.67933850e-04, 1.30400702e-04, 3.10731193e-05,
2.82113149e-04, 6.14081291e-05, 2.89048231e-03, 3.98092088e-04,
1.53062156e-05, 1.85977016e-03, 4.84216580e-04, 1.83015407e-04,
1.53889210e-04, 1.32637697e-05, 4.98024783e-05, 9.33829069e-05,
1.08729248e-04, 1.96910370e-03, 8.92132593e-05, 7.47793456e-05,
3.03913461e-04, 5.80130472e-05, 1.10250810e-04, 3.30982637e-03,
5.04739233e-04, 9.20615275e-05, 7.00237288e-04, 8.91018353e-05,
1.07366790e-03, 2.60756264e-04, 2.11966879e-04, 8.20988615e-04,
2.19961730e-04, 2.33830389e-04, 3.15418845e-04, 4.79362061e-05,
2.69673445e-04, 5.04744712e-05, 5.64143920e-05, 5.33246894e-05,
6.90017478e-05, 1.36698000e-04, 2.83982721e-04, 8.47348856e-05,
5.74015139e-05, 1.15665382e-04, 4.05571511e-04, 4.68011480e-04,
4.67822865e-05, 3.00760876e-04, 1.67407416e-04, 3.02288918e-05,
1.11583337e-04, 1.59645104e-04, 1.73546228e-04, 6.30670751e-04,
1.02046807e-03, 2.51042400e-03, 1.16524985e-04, 7.50074469e-05,
8.18201734e-05, 1.62838216e-04, 2.80639011e-04, 2.22368413e-04,
1.39261465e-04, 7.51803600e-05, 3.63202620e-04, 4.17022275e-05,
3.47620167e-04, 5.77659339e-05, 4.18579002e-05, 2.42096503e-05,
1.12250843e-03, 2.84655980e-04, 1.09178177e-03, 2.29647441e-04,
8.82061548e-04, 2.46612704e-04, 2.03358708e-04, 1.37021038e-04,
2.18374873e-04, 2.04092226e-04, 2.34132949e-05, 3.30167168e-05,
9.79309698e-05, 5.24312258e-04, 5.23222424e-03, 4.28866944e-04,
4.27278283e-04, 4.20961696e-05, 7.91574595e-04, 1.49085918e-05,
6.50748750e-03, 1.34461065e-04, 2.10423168e-04, 2.58454558e-04,
6.88945584e-05, 2.52605474e-04, 1.82618242e-05, 7.12760302e-05,
2.55153718e-04, 2.36239048e-05, 1.01488891e-04, 1.98719892e-04,
3.77983743e-05, 7.20363343e-04, 9.62147082e-04, 1.28874002e-04,
9.41858088e-05, 3.23168957e-03, 1.23727470e-04, 1.92188854e-05,
3.54324962e-04, 3.18146122e-05, 7.00732926e-04, 6.85932173e-05,
2.01224640e-04, 6.98182499e-03, 4.59520757e-04, 9.72801761e-04,
5.03679446e-04, 4.76087164e-03, 3.34953744e-04, 1.01442929e-04,
9.72617563e-05, 1.41746947e-04, 7.29511899e-04, 3.49429640e-04,
2.15930704e-05, 6.75361225e-05, 3.15851881e-04, 6.48199712e-05,
3.88102460e-04, 1.08534540e-03, 3.40335624e-04, 7.11025714e-05,
2.24293879e-04, 1.50417589e-04, 2.75202565e-05, 5.46244482e-05,
4.36935152e-05, 2.22663773e-04, 2.66235329e-05, 5.41452435e-04,
3.01329768e-03, 6.68641151e-05, 2.13318184e-04, 4.83280281e-04,
3.25042219e-03, 2.68050027e-03, 2.30311023e-04, 8.69394498e-05,
2.22041956e-04, 3.05501686e-04, 2.05413511e-04, 1.23969074e-02,
1.59675707e-04, 8.81001033e-05, 1.13639137e-04, 5.50219047e-05,
7.88093021e-05, 2.22792616e-04, 4.43840865e-04, 1.43152502e-04,
1.66197169e-05, 3.35655146e-04, 5.74437363e-05, 1.59689171e-05,
6.05816713e-05, 3.97622352e-04, 3.04213725e-03, 1.58328272e-04,
1.24343760e-05, 9.89821274e-04, 2.06074947e-05, 7.68100668e-04,
3.63771833e-04, 1.09289220e-04, 9.56191216e-05, 2.01528255e-05,
2.48080294e-04, 1.70826330e-04, 3.17801314e-05, 8.31752113e-05,
1.45166923e-04, 3.32538475e-05, 4.35978654e-05, 2.31716942e-04,
9.36048309e-05, 1.46375887e-05, 5.24211209e-05, 4.74207263e-05,
5.80011001e-05, 1.25948922e-04, 6.51396252e-03, 8.12801300e-04,
5.31415171e-05, 1.57267496e-03, 6.40222663e-03, 7.39321840e-05,
4.70565539e-03, 1.65742051e-04, 1.82069867e-04, 8.12231665e-05,
2.21989758e-05, 6.43523454e-05, 1.68223705e-04, 2.12641185e-04,
4.13967442e-04, 6.86593557e-05, 3.99616620e-05, 3.24723987e-05,
5.34793035e-05, 1.45994454e-05, 8.81493106e-05, 1.56545924e-04,
2.27955333e-03, 4.21397563e-04, 4.95772947e-05, 2.61580717e-04,
1.35582286e-05, 3.30929557e-04, 2.16315282e-04, 6.19240600e-05,
2.29309764e-04, 1.89436760e-05, 9.95220253e-05, 1.30594799e-05,
2.31584621e-04, 2.58677243e-03, 3.26315960e-04, 1.45622762e-04,
1.95064305e-04, 7.66519734e-05, 1.00914447e-03, 1.00372148e-04,
6.48575078e-05, 1.86936624e-04, 1.26099811e-04, 5.64772126e-05,
3.31675867e-04, 2.09970924e-04, 1.60944095e-04, 5.30265497e-05,
1.47023078e-04, 7.85566168e-04, 7.93441359e-05, 1.78703503e-03,
4.36751638e-04, 1.18882686e-04, 1.46988168e-04, 6.06477435e-04,
6.12383519e-05, 8.20263958e-05, 4.62990356e-05, 4.65407902e-05])
[26]:
shap_values.output_names
[26]:
['American_egret',
'crane',
'little_blue_heron',
'flamingo',
'peacock',
'goose',
'white_stork',
'spoonbill',
'black_stork',
'pelican',
'albatross',
'limpkin',
'bustard',
'drake',
'bittern',
'king_penguin',
'black_swan',
'eel',
'jay',
'bee_eater',
'weasel',
'lakeside',
'magpie',
'red-breasted_merganser',
'vulture',
'crayfish',
'ostrich',
'water_ouzel',
'hook',
'Italian_greyhound',
'vine_snake',
'nematode',
'paper_towel',
'redshank',
'gazelle',
'balance_beam',
'dragonfly',
'dingo',
'chambered_nautilus',
'coil',
'macaw',
'dowitcher',
'brain_coral',
'ice_bear',
'toilet_tissue',
'mushroom',
'European_gallinule',
'damselfly',
'bubble',
'sea_snake',
'hognose_snake',
'oystercatcher',
'fountain_pen',
'coral_reef',
'green_snake',
'Siamese_cat',
'sundial',
'ruddy_turnstone',
'paintbrush',
'umbrella',
'American_chameleon',
'coral_fungus',
'rubber_eraser',
'gar',
'ant',
'American_coot',
'fox_squirrel',
'snail',
'Ibizan_hound',
'earthstar',
'jacamar',
'hummingbird',
'quill',
'chain',
'red-backed_sandpiper',
'pole',
'sandbar',
'goblet',
'fountain',
'knot',
'bighorn',
'Granny_Smith',
'chime',
'coffee_mug',
'bald_eagle',
'hen-of-the-woods',
'barrow',
'banana',
'anemone_fish',
'syringe',
'indigo_bunting',
'slug',
'starfish',
'nipple',
'remote_control',
'bathing_cap',
'ram',
'crossword_puzzle',
'pencil_sharpener',
'hornbill',
'African_chameleon',
'sea_urchin',
'goldfinch',
'thunder_snake',
'monitor',
'ping-pong_ball',
'bath_towel',
'sulphur-crested_cockatoo',
'ringneck_snake',
'horizontal_bar',
'chainlink_fence',
'Eskimo_dog',
'rule',
'cowboy_hat',
'tick',
'parallel_bars',
'African_elephant',
'whippet',
'meerkat',
'boa_constrictor',
'lorikeet',
'padlock',
'impala',
'thimble',
'bonnet',
'sea_slug',
'sidewinder',
'golf_ball',
'park_bench',
'bucket',
'horned_viper',
'quail',
'bannister',
'ballplayer',
'loudspeaker',
'Band_Aid',
'Labrador_retriever',
'Indian_elephant',
'binoculars',
'ptarmigan',
'kite',
'plastic_bag',
'proboscis_monkey',
'scorpion',
'iPod',
'lotion',
'bolete',
'ear',
'dishwasher',
'lampshade',
'birdhouse',
'flagpole',
'ski',
'sunscreen',
'polecat',
'letter_opener',
'washbasin',
'Egyptian_cat',
'bee',
'green_lizard',
'cup',
'handkerchief',
'French_horn',
'rapeseed',
'bassoon',
'grocery_store',
'canoe',
'bib',
'toucan',
'sea_lion',
'swimming_trunks',
'bell_pepper',
'sombrero',
'orange',
'parking_meter',
'lens_cap',
'screw',
'leatherback_turtle',
'dumbbell',
'barbell',
'tub',
'Weimaraner',
'black-footed_ferret',
'gibbon',
'pick',
'Saluki',
'mailbox',
'Petri_dish',
'whistle',
'macaque',
'paddle',
'gown',
'bearskin',
'ladle',
'toy_terrier',
'titi',
'web_site',
'orangutan',
'worm_fence',
'partridge',
'sock',
'dalmatian',
'feather_boa',
'Siberian_husky',
'siamang',
'velvet',
'digital_clock',
'hatchet',
'measuring_cup',
'stingray',
'stinkhorn',
'volleyball',
'bow',
'electric_fan',
'notebook',
'flute',
'television',
'butternut_squash',
'banded_gecko',
'sea_anemone',
'maillot',
'pretzel',
'alligator_lizard',
'agama',
'tennis_ball',
'trombone',
'pitcher',
'stage',
'crutch',
'soccer_ball',
'hard_disc',
'pedestal',
'analog_clock',
'golden_retriever',
'wing',
'honeycomb',
'shopping_cart',
'green_mamba',
'basenji',
'rugby_ball',
'platypus',
'can_opener',
"yellow_lady's_slipper",
'lab_coat',
'oboe',
'rock_python',
'Bedlington_terrier',
'matchstick',
'mink',
'crane',
'wooden_spoon',
'trimaran',
'mongoose',
'bolo_tie',
'robin',
'church',
'table_lamp',
'running_shoe',
'African_grey',
'seashore',
'hand-held_computer',
'mortarboard',
'mixing_bowl',
'conch',
'lemon',
'ruffed_grouse',
'tripod',
'ice_lolly',
'drumstick',
'butcher_shop',
'toyshop',
'chickadee',
'borzoi',
'boathouse',
'strainer',
'black_grouse',
'loupe',
'lynx',
'pinwheel',
'rocking_chair',
'stopwatch',
'Arctic_fox',
'cellular_telephone',
'neck_brace',
'bagel',
'stole',
'brambling',
'collie',
'walking_stick',
'hare',
'hammerhead',
'window_screen',
'suspension_bridge',
'schooner',
'hair_slide',
'doormat',
'badger',
'house_finch',
'hog',
'dock',
'hair_spray',
'racket',
'Indian_cobra',
'night_snake',
'corn',
'rain_barrel',
'sunglass',
'nail',
'beagle',
'vizsla',
'beaver',
'groenendael',
'bloodhound',
'electric_ray',
'piggy_bank',
'diaper',
'maillot',
'fiddler_crab',
'fly',
'otter',
'mouse',
'computer_keyboard',
'printer',
'Boston_bull',
'desktop_computer',
'typewriter_keyboard',
'Doberman',
'beaker',
'mousetrap',
'carton',
'swing',
'convertible',
'red_fox',
'breakwater',
'mask',
'beacon',
'water_snake',
'sunglasses',
'barbershop',
'dough',
'English_foxhound',
'coucal',
'screen',
'jeep',
'ringlet',
'reel',
'squirrel_monkey',
'water_bottle',
'American_Staffordshire_terrier',
'obelisk',
'grasshopper',
'wallaby',
'airliner',
'wig',
'Irish_setter',
'centipede',
'boxer',
'street_sign',
'cash_machine',
'jean',
'plunger',
'minivan',
'mosque',
'cabbage_butterfly',
'microphone',
'lion',
'eggnog',
'espresso',
'car_mirror',
'pineapple',
'toilet_seat',
'snorkel',
'common_newt',
'miniature_poodle',
'flat-coated_retriever',
'file',
'spaghetti_squash',
'standard_poodle',
'beach_wagon',
'junco',
'fur_coat',
'croquet_ball',
'tricycle',
'combination_lock',
'mobile_home',
'wall_clock',
'Great_Dane',
'zebra',
'face_powder',
'daisy',
'balloon',
'Cardigan',
'pillow',
'washer',
'spindle',
'schipperke',
'redbone',
'harmonica',
'miniature_pinscher',
'ice_cream',
'revolver',
'swab',
'echidna',
'seat_belt',
'power_drill',
'window_shade',
'kuvasz',
'bow_tie',
'rock_beauty',
'hammer',
'screwdriver',
'prayer_rug',
'admiral',
'water_jug',
'cardigan',
'hen',
'strawberry',
'warplane',
'perfume',
'buckeye',
'spotlight',
'gong',
'wine_bottle',
'guenon',
'armadillo',
'photocopier',
'pickelhaube',
'ski_mask',
'bikini',
'coyote',
'bulbul',
'jinrikisha',
'langur',
'chain_mail',
'loggerhead',
'shield',
'gorilla',
'hay',
'Greater_Swiss_Mountain_dog',
'Pomeranian',
'torch',
'puffer',
'potpie',
'flatworm',
'tobacco_shop',
'picket_fence',
'digital_watch',
'unicycle',
'sarong',
'dugong',
'desk',
'space_shuttle',
'hoopskirt',
'patio',
'dhole',
'tray',
'speedboat',
'Rottweiler',
'triceratops',
'Komodo_dragon',
'punching_bag',
'Chihuahua',
'medicine_chest',
'tabby',
'wild_boar',
'shoji',
'corkscrew',
'custard_apple',
'llama',
'hip',
'marmoset',
'sliding_door',
'red_wolf',
'cicada',
'mashed_potato',
'hermit_crab',
'pier',
'sweatshirt',
'tusker',
'Arabian_camel',
'white_wolf',
'plow',
'soup_bowl',
'kimono',
'malinois',
'maze',
'hamper',
'promontory',
'bassinet',
'cock',
'diamondback',
'knee_pad',
'baseball',
'oscilloscope',
'crash_helmet',
'trifle',
'frilled_lizard',
'space_bar',
'golfcart',
'lighter',
'crib',
'crate',
'jellyfish',
'cardoon',
'agaric',
'prairie_chicken',
'beer_bottle',
'lawn_mower',
'cucumber',
'Irish_terrier',
'tiger_beetle',
'plate',
'football_helmet',
'academic_gown',
'radio',
'wool',
'German_shepherd',
'quilt',
'leafhopper',
'violin',
"carpenter's_kit",
'ox',
'Norfolk_terrier',
'scuba_diver',
'dome',
'pickup',
'cliff',
'spiny_lobster',
'shovel',
'CD_player',
'West_Highland_white_terrier',
'military_uniform',
'fireboat',
'basset',
'lipstick',
'ashcan',
'briard',
'skunk',
'purse',
'Great_Pyrenees',
'axolotl',
'projectile',
'vacuum',
'ocarina',
'groom',
'shower_curtain',
'whiptail',
'lionfish',
'spider_web',
'binder',
'Border_collie',
'pot',
'palace',
'isopod',
'sax',
'sewing_machine',
'teddy',
'lycaenid',
'capuchin',
'brown_bear',
'indri',
'kelpie',
'ballpoint',
'shower_cap',
'cab',
'gas_pump',
'artichoke',
'book_jacket',
'plane',
'Madagascar_cat',
'chimpanzee',
'colobus',
'Staffordshire_bullterrier',
'Rhodesian_ridgeback',
'pirate',
'giant_panda',
'ladybug',
'great_grey_owl',
'cradle',
'malamute',
'clumber',
'head_cabbage',
'necklace',
'pay-phone',
'switch',
'broccoli',
'wok',
'spatula',
'backpack',
'soap_dispenser',
'shoe_shop',
'scoreboard',
'recreational_vehicle',
'cocker_spaniel',
'upright',
'totem_pole',
'sturgeon',
'papillon',
'barrel',
'terrapin',
'dishrag',
'cauliflower',
'Samoyed',
'geyser',
'beer_glass',
'grille',
'patas',
'bathtub',
'dung_beetle',
'Christmas_stocking',
'sorrel',
'grey_fox',
'mountain_tent',
'laptop',
'missile',
'parachute',
'restaurant',
'guillotine',
'cloak',
'planetarium',
'hartebeest',
'passenger_car',
'Airedale',
'kit_fox',
'studio_couch',
'Leonberg',
'jackfruit',
'cornet',
'Australian_terrier',
'rifle',
'rhinoceros_beetle',
'wolf_spider',
'barracouta',
'gondola',
'Gordon_setter',
'viaduct',
'mitten',
'Shih-Tzu',
'shopping_basket',
'panpipe',
'packet',
'Maltese_dog',
'grand_piano',
'Angora',
'hamster',
'ground_beetle',
'wallet',
'French_loaf',
'sandal',
'drum',
'carbonara',
'magnetic_compass',
'Gila_monster',
'greenhouse',
'catamaran',
'gasmask',
'cricket',
'odometer',
'iron',
'weevil',
'espresso_maker',
'mantis',
'cairn',
'stethoscope',
'castle',
'puck',
'fire_screen',
'monarch',
'scale',
'reflex_camera',
'gyromitra',
'spotted_salamander',
'muzzle',
'pop_bottle',
'oxcart',
'wood_rabbit',
'bell_cote',
'safety_pin',
'chiton',
'howler_monkey',
'sulphur_butterfly',
'miniskirt',
'tailed_frog',
'jaguar',
'miniature_schnauzer',
'bakery',
'bicycle-built-for-two',
'sloth_bear',
'porcupine',
'hand_blower',
'abacus',
'horse_cart',
'abaya',
'stupa',
'lacewing',
'pencil_box',
'oil_filter',
'Kerry_blue_terrier',
'pill_bottle',
'suit',
'leopard',
'solar_dish',
'lesser_panda',
'barber_chair',
'safe',
'pomegranate',
'Walker_hound',
'Dutch_oven',
'garter_snake',
'Pekinese',
'volcano',
'pajama',
'folding_chair',
'cello',
'tiger_cat',
'lumbermill',
'vault',
'chocolate_sauce',
'soft-coated_wheaten_terrier',
'chain_saw',
'cockroach',
'stone_wall',
'trolleybus',
'carousel',
'alp',
'moped',
'comic_book',
'tile_roof',
'acorn',
'water_tower',
'cocktail_shaker',
'mosquito_net',
'English_setter',
'prison',
'library',
'microwave',
'three-toed_sloth',
'maypole',
'coho',
'Afghan_hound',
'bottlecap',
'liner',
'eft',
'hourglass',
'motor_scooter',
'harp',
'Yorkshire_terrier',
'mountain_bike',
'consomme',
'cinema',
'joystick',
"jack-o'-lantern",
'bookshop',
'space_heater',
'accordion',
'disk_brake',
'toy_poodle',
'sea_cucumber',
'tarantula',
'Saint_Bernard',
'dial_telephone',
'buckle',
'basketball',
'fire_engine',
'teapot',
'paddlewheel',
'cougar',
'vase',
'Brabancon_griffon',
'menu',
'container_ship',
'bookcase',
'jersey',
'triumphal_arch',
'minibus',
'long-horned_beetle',
'marmot',
'Irish_water_spaniel',
'home_theater',
'Norwich_terrier',
'trailer_truck',
'guacamole',
'king_snake',
'Norwegian_elkhound',
'standard_schnauzer',
'Appenzeller',
'chiffonier',
'go-kart',
'Pembroke',
'bull_mastiff',
'bullet_train',
'acoustic_guitar',
'guinea_pig',
'Sussex_spaniel',
'wreck',
'bluetick',
'forklift',
'Blenheim_spaniel',
'killer_whale',
'coffeepot',
'pool_table',
'airship',
'vestment',
'Chesapeake_Bay_retriever',
'Scotch_terrier',
'Mexican_hairless',
'jigsaw_puzzle',
'American_alligator',
'tractor',
'rock_crab',
'Persian_cat',
'French_bulldog',
'black_widow',
'school_bus',
'African_crocodile',
'bobsled',
'radiator',
'chest',
'projector',
'African_hunting_dog',
'electric_guitar',
'grey_whale',
'Newfoundland',
'tiger_shark',
'steel_arch_bridge',
'cowboy_boot',
'zucchini',
'Welsh_springer_spaniel',
'barn',
'traffic_light',
'bulletproof_vest',
'turnstile',
"potter's_wheel",
'spider_monkey',
'goldfish',
'confectionery',
'whiskey_jug',
'Brittany_spaniel',
'burrito',
'ibex',
'trilobite',
'steel_drum',
'Border_terrier',
'hotdog',
'cheetah',
'mud_turtle',
'dining_table',
'cannon',
'electric_locomotive',
'plate_rack',
'toaster',
'Scottish_deerhound',
'megalith',
'milk_can',
'racer',
'Shetland_sheepdog',
'oxygen_mask',
'dogsled',
'cassette',
'entertainment_center',
'thresher',
'candle',
'saltshaker',
'tape_player',
'brassiere',
'envelope',
'fig',
'EntleBucher',
'harvestman',
'Old_English_sheepdog',
'Sealyham_terrier',
'barometer',
'thatch',
'streetcar',
'Crock_Pot',
'ambulance',
'submarine',
'red_wine',
'car_wheel',
'common_iguana',
'vending_machine',
'pug',
'acorn_squash',
'water_buffalo',
'baboon',
'European_fire_salamander',
'brass',
'cuirass',
'maraca',
'waffle_iron',
'black-and-tan_coonhound',
'Loafer',
'broom',
'slide_rule',
'hyena',
'chow',
'sleeping_bag',
'mailbag',
'banjo',
'cleaver',
'yawl',
'aircraft_carrier',
'tiger',
'timber_wolf',
'slot',
'Polaroid_camera',
'Japanese_spaniel',
'snow_leopard',
'box_turtle',
'Dandie_Dinmont',
'poncho',
'giant_schnauzer',
'English_springer',
'rotisserie',
'refrigerator',
'Irish_wolfhound',
'koala',
'modem',
'silky_terrier',
'stove',
'stretcher',
'Tibetan_terrier',
'cliff_dwelling',
'Dungeness_crab',
'leaf_beetle',
'American_lobster',
'steam_locomotive',
'bullfrog',
'cassette_player',
'Model_T',
'Tibetan_mastiff',
'wardrobe',
'sports_car',
'apiary',
'valley',
'half_track',
'great_white_shark',
'black_and_gold_garden_spider',
'theater_curtain',
'tree_frog',
'snowmobile',
'Windsor_tie',
'assault_rifle',
'wombat',
'frying_pan',
'Bernese_mountain_dog',
'Lhasa',
'freight_car',
'altar',
'harvester',
'garbage_truck',
'Bouvier_des_Flandres',
'throne',
'yurt',
'garden_spider',
'four-poster',
'tow_truck',
'pizza',
'meat_loaf',
'china_cabinet',
'scabbard',
'marimba',
'tank',
'clog',
'warthog',
'wire-haired_fox_terrier',
'tench',
'organ',
'curly-coated_retriever',
'king_crab',
'komondor',
'police_van',
'apron',
'barn_spider',
'radio_telescope',
'hippopotamus',
'Lakeland_terrier',
'American_black_bear',
'mortar',
'hot_pot',
'German_short-haired_pointer',
'monastery',
'otterhound',
'dam',
'overskirt',
'trench_coat',
'affenpinscher',
'bison',
'holster',
'breastplate',
'keeshond',
'caldron',
'limousine',
'moving_van',
'manhole_cover',
'drilling_platform',
'amphibian',
'snowplow',
'cheeseburger',
'lifeboat']
[ ]:
[20]:
# output with shap values
shap.image_plot(shap_values[0])
---------------------------------------------------------------------------
Exception Traceback (most recent call last)
<ipython-input-20-b98c3ef2fd6b> in <module>
1 # output with shap values
----> 2 shap.image_plot(shap_values[0])
~/projects/shap/shap/plots/_image.py in image(shap_values, pixel_values, labels, width, aspect, hspace, labelpad, show)
57 shap_values = [shap_exp.values[..., i] for i in range(shap_exp.values.shape[-1])]
58 else:
---> 59 raise Exception("Number of outputs needs to have support added!! (probably a simple fix)")
60 if pixel_values is None:
61 pixel_values = shap_exp.data
Exception: Number of outputs needs to have support added!! (probably a simple fix)
Interpretation of SHAP output explanation:
In the first example, given bird image is classified as an American Egret with next probable classes being a Crane, Heron and Flamingo. It is the “bump” over the bird’s neck that causes it to be classified as an American Egret vs a Crane, Heron or a Flamingo. You can see the neck region of the bird appropriately highlighted in red super pixels.
In the second example, it is the shape of the boat which causes it to be classified as a speedboat instead of a fountain, lifeboat or snowplow (appropriately highlighted in red super pixels).
A longer run with many evaluations
By increasing the max_evals parameter we let SHAP execute the original model more times and so get a more finely detailed explaination. We also use the blur kernel here, both to demonstrate it, and because it is much faster than inpainting. Note that this will take a while if you are not using a modern GPU on your system.
[6]:
# python function to get model output; replace this function with your own model function.
def f(x):
tmp = x.copy()
preprocess_input(tmp)
return model(tmp)
# define a masker that is used to mask out partitions of the input image.
masker_blur = shap.maskers.Image("blur(128,128)", X[0].shape)
# create an explainer with model and image masker
explainer_blur = shap.Explainer(f, masker_blur, output_names=class_names)
# here we explain two images using 500 evaluations of the underlying model to estimate the SHAP values
shap_values_fine = explainer_blur(X[1:3], max_evals=5000, batch_size=50, outputs=shap.Explanation.argsort.flip[:4])
Partition explainer: 3it [00:17, 5.94s/it]
[7]:
# output with shap values
shap.image_plot(shap_values_fine)
Have an idea for more helpful examples? Pull requests that add to this documentation notebook are encouraged!