Shap: KeyError: 'ClipByValue' during keras regression

[dynamik1703]

Hey Scott,

thanks for your great work.

I'm using keras for regression. My data set has 5 continuous features.

Following error occurs when shap is executed:

~AppDataLocalcondacondaenvstensorflowlibsite-packagesshapexplainersdeep.py in phi_symbolic(self, i)
194 reg = tf_ops._gradient_registry._registry # hack our way in to the registry (TODO: find an API for this)
195 for n in DeepExplainer.nonlinearities:
--> 196 self.orig_grads[n] = reg[n]["type"]
197 reg[n]["type"] = self.custom_grad
198

KeyError: 'ClipByValue'

What could be the reason?

Best,
Roman

[SteveJae]

Same error in example notebook

Hello Scott,
I get the same error when running the sample notebook for the DeepExplainer for MNIST:
(shap 0.21.0, Keras 2.2.2, TensorFlow 1.9)

KeyError Traceback (most recent call last)
in ()
11 # ...or pass tensors directly
12 # e = shap.DeepExplainer((model.layers[0].input, model.layers[-1].output), background)
---> 13 shap_values = e.shap_values(x_test[1:5])

~/virtualenvs/pyenv36/lib/python3.6/site-packages/shap/explainers/deep.py in shap_values(self, X, ranked_outputs, output_rank_order)
282 # run attribution computation graph
283 feature_ind = model_output_ranks[j,i]
--> 284 sample_phis = self.run(self.phi_symbolic(feature_ind), self.model_inputs, joint_input)
285
286 # assign the attributions to the right part of the output arrays

~/virtualenvs/pyenv36/lib/python3.6/site-packages/shap/explainers/deep.py in phi_symbolic(self, i)
194 reg = tf_ops._gradient_registry._registry # hack our way in to the registry (TODO: find an API for this)
195 for n in DeepExplainer.nonlinearities:
--> 196 self.orig_grads[n] = reg[n]["type"]
197 reg[n]["type"] = self.custom_grad
198

KeyError: 'ClipByValue'

Best,
Steve

[dynamik1703]

My config:
Keras==2.2.0
tensorflow-gpu==1.9.0
shap==0.21.0

Best,
Roman

[AllardJM]

Same error:

import shap
from keras.models import Sequential
from keras.layers import Dense, Activation
from sklearn.datasets.california_housing import fetch_california_housing
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler

cal_housing = fetch_california_housing()

ss=StandardScaler()
X=ss.fit_transform(cal_housing.data)


# split 80/20 train-test
X_train, X_test, y_train, y_test = train_test_split(X,
                                                        cal_housing.target,
                                                        test_size=0.2,
                                                      random_state=1)
names = cal_housing.feature_names

model = Sequential()
model.add(Dense(32, input_dim=X_train.shape[1]))
model.add(Activation('relu'))
model.add(Dense(16))
model.add(Activation('relu'))
model.add(Dense(1, activation="linear"))
model.compile(optimizer='rmsprop',
              loss='mse')
model.fit(X_train,y_train,batch_size=8,epochs=10)

background = X_train
# explain predictions of the model on three images
e = shap.DeepExplainer(model, background)


shap_values = e.shap_values(X_test)

---

KeyError Traceback (most recent call last)
in ()
4
5
----> 6 shap_values = e.shap_values(X_test)

~/anaconda3/lib/python3.6/site-packages/shap/explainers/deep.py in shap_values(self, X, ranked_outputs, output_rank_order)
282 # run attribution computation graph
283 feature_ind = model_output_ranks[j,i]
--> 284 sample_phis = self.run(self.phi_symbolic(feature_ind), self.model_inputs, joint_input)
285
286 # assign the attributions to the right part of the output arrays

~/anaconda3/lib/python3.6/site-packages/shap/explainers/deep.py in phi_symbolic(self, i)
194 reg = tf_ops._gradient_registry._registry # hack our way in to the registry (TODO: find an API for this)
195 for n in DeepExplainer.nonlinearities:
--> 196 self.orig_grads[n] = reg[n]["type"]
197 reg[n]["type"] = self.custom_grad
198

KeyError: 'ClipByValue'

[slundberg]

Just got to this issue... I can't seem to reproduce the problem using v0.22.2 which includes some updates to DeepExplainer. Could one of you check and see if the problem still happens for you in 0.22?

[IFV]

Same problem using v0.22.2. Ps: keras classification

KeyError Traceback (most recent call last)
in ()
4 # explain predictions of the model on four images
5 e = shap.DeepExplainer(classifier, background)
----> 6 shap_values = e.shap_values(X_test)

~Anaconda3libsite-packagesshapexplainersdeep.py in shap_values(self, X, ranked_outputs, output_rank_order)
302 # run attribution computation graph
303 feature_ind = model_output_ranks[j,i]
--> 304 sample_phis = self.run(self.phi_symbolic(feature_ind), self.model_inputs, joint_input)
305
306 # assign the attributions to the right part of the output arrays

~Anaconda3libsite-packagesshapexplainersdeep.py in phi_symbolic(self, i)
214 reg = tf_ops._gradient_registry._registry # hack our way in to the registry (TODO: find an API for this)
215 for n in DeepExplainer.nonlinearities:
--> 216 self.orig_grads[n] = reg[n]["type"]
217 reg[n]["type"] = self.custom_grad
218

KeyError: 'ClipByValue'

[slundberg]

Looks like ClipByValue was removed from the gradient registry in TF 1.9, which caused this problem. I pushed a fix that should prevent these issues in the future which will go out with v0.22.3