Fails to display in notebook - Object of type 'ndarray' is not JSON serializable

Hello I have a deep learning classifier using Keras (TensorFlow), I am able to get the explanation using basic text but note in the pretty UI What works

from lime.lime_text import LimeTextExplainer
explainer = LimeTextExplainer(class_names=TEXT_LABELS)

def new_predict(texts):
  _seq = tk.texts_to_sequences(texts)
  _text_data = pad_sequences(_seq, maxlen=MAX_LEN)
  return single_channel_kim_model.predict(_text_data)

exp = explainer.explain_instance('forget reservations thank great company i have cancelled flighted flight once again thank you',
                                 new_predict,
                                 num_features=7,
                                 labels=[0, 2]
                                )

print ('Explanation for class %s' % TEXT_LABELS[0])
print ('\n'.join(map(str, exp.as_list(label=0))))
print ()
print ('Explanation for class %s' % TEXT_LABELS[2])
print ('\n'.join(map(str, exp.as_list(label=2))))

When I use exp.show_in_notebook(text='forget reservations thank great company i have cancelled flighted flight once again thank you', labels=(0,)) It throws TypeError: Object of type ‘ndarray’ is not JSON serializable Thanks

[paramita-c]

I was getting the same error while trying to use LIME in Titanic data set (which had a mix of numeric and categorical variables) LimeTabularExplainer expects all the numeric attributes to be as float64 format in order to be displayed using show_in_notebook method. So convert all the int64 features (including the target variable) to np.float64 type before starting to build and interpret models using LIME. I have used LIME to interpret my models for Titanic data set, entire code (data cleaning, preprocessing, feature engineering, interpreting with LIME) can be found in my github account

[engmubarak48]

I have fixed the issue, I don’t think it was about versions. you just have to make sure your data type before you feed it to your the LimeTabularExplainer. In my case, categorical_names in LimeTabularExplainer was int64 [0, 1], instead of [no, yes]. I realize this parameter should be a Unicode data type. So .astype(‘U’) helped.