Keras LSTM for IMDB Sentiment Classification¶
This is simple example of how to explain a Keras LSTM model using DeepExplainer.
In [1]:
# This model training code is directly from:
# https://github.com/keras-team/keras/blob/master/examples/imdb_lstm.py
'''Trains an LSTM model on the IMDB sentiment classification task.
The dataset is actually too small for LSTM to be of any advantage
compared to simpler, much faster methods such as TF-IDF + LogReg.
# Notes
- RNNs are tricky. Choice of batch size is important,
choice of loss and optimizer is critical, etc.
Some configurations won't converge.
- LSTM loss decrease patterns during training can be quite different
from what you see with CNNs/MLPs/etc.
'''
from __future__ import print_function
from keras.preprocessing import sequence
from keras.models import Sequential
from keras.layers import Dense, Embedding
from keras.layers import LSTM
from keras.datasets import imdb
max_features = 20000
maxlen = 80 # cut texts after this number of words (among top max_features most common words)
batch_size = 32
print('Loading data...')
(x_train, y_train), (x_test, y_test) = imdb.load_data(num_words=max_features)
print(len(x_train), 'train sequences')
print(len(x_test), 'test sequences')
print('Pad sequences (samples x time)')
x_train = sequence.pad_sequences(x_train, maxlen=maxlen)
x_test = sequence.pad_sequences(x_test, maxlen=maxlen)
print('x_train shape:', x_train.shape)
print('x_test shape:', x_test.shape)
print('Build model...')
model = Sequential()
model.add(Embedding(max_features, 128))
model.add(LSTM(128, dropout=0.2, recurrent_dropout=0.2))
model.add(Dense(1, activation='sigmoid'))
# try using different optimizers and different optimizer configs
model.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy'])
print('Train...')
model.fit(x_train, y_train,
batch_size=batch_size,
epochs=15,
validation_data=(x_test, y_test))
score, acc = model.evaluate(x_test, y_test,
batch_size=batch_size)
print('Test score:', score)
print('Test accuracy:', acc)
Explain the model with DeepExplainer and visualize the first prediction¶
In [3]:
import shap
# we use the first 100 training examples as our background dataset to integrate over
explainer = shap.DeepExplainer(model, x_train[:100])
# explain the first 10 predictions
# explaining each prediction requires 2 * background dataset size runs
shap_values = explainer.shap_values(x_test[:10])
In [4]:
# init the JS visualization code
shap.initjs()
# transform the indexes to words
import numpy as np
words = imdb.get_word_index()
num2word = {}
for w in words.keys():
num2word[words[w]] = w
x_test_words = np.stack([np.array(list(map(lambda x: num2word.get(x, "NONE"), x_test[i]))) for i in range(10)])
# plot the explanation of the first prediction
# Note the model is "multi-output" because it is rank-2 but only has one column
shap.force_plot(explainer.expected_value[0], shap_values[0][0], x_test_words[0])
Out[4]:
Note that each sample is an IMDB review text document, represented as a sequence of words. This means "feature 0" is the first word in the review, which will be different for difference reviews. This means calling summary_plot will combine the importance of all the words by their position in the text. This is likely not what you want for a global measure of feature importance (which is why we have not called summary_plot here). If you do want a global summary of a word's importance you could pull apart the feature attribution values and group them by words.