GradientExplainer only works on copies of models?

[delacylab]

I am trying to use shap with an LSTM and data with 4 time periods so my datasets typically look like (num_rows,4,num_features). Because DeepExplainer is not compatible with Tensorflow 2.x, I have been using GradientExplainer but ran into this bizarre problem where calling the Explainer with my model gives me a warning but calling it on a copy of the model works properly. clf = Sequential() clf.add(Bidirectional(LSTM(layer_size, activation='tanh', return_sequences=True))) clf.add(Bidirectional(LSTM(layer_size, activation='tanh'))) clf.add(Dense(2, activation = 'softmax')) clf.compile(loss="categorical_crossentropy", optimizer = tfa.optimizers.AdamW(weight_decay=.01), metrics=['accuracy']) y_cat = tf.keras.utils.to_categorical(y, num_classes=2) early_stopping_monitor = EarlyStopping(min_delta = 0.01, monitor='loss',patience=10) clf.fit(X,y_cat, epochs=200, callbacks=[early_stopping_monitor]) e = shap.GradientExplainer(clf, X)

If I then later call: shap_values, shap_indexes = e.shap_values(rep) I get this warning: WARNING:tensorflow:Layers in a Sequential model should only have a single input tensor, but we receive a <class 'list'> input: [<tf.Tensor: shape=(138, 4, 187)...

I do not get the warning if I change the final line in the first code segment to: e = shap.GradientExplainer(copy.deepcopy(clf), X) Has anyone else run into this? Is shap thread-safe or is there something else going on? This is the only way I have got GradientExplainer to work without warnings that worry me,

[CloseChoice]

Thanks for reporting the issue. Could you please provide a minimal reproducible example for this? So, specifying how to get X and y in your code or using a dummy dataset? Otherwise it is hard for us to look deeper into this.

[github-actions[bot]]

This issue has been inactive for two years, so it's been automatically marked as 'stale'.

We value your input! If this issue is still relevant, please leave a comment below. This will remove the 'stale' label and keep it open.

If there's no activity in the next 90 days the issue will be closed.