def compute_anchor(text):
print("Initialize Anchor")
explainer_anchor_cnn = AnchorText(nlp, predict_cnn)
explanation_anchor_cnn = explainer_anchor_cnn.explain(text, desired_label=1, threshold=0.75,
use_similarity_proba=False, use_unk=True, sample_proba=0.5, beam_size=1, tau=0.15)
del explainer_anchor_cnn
return explanation_anchor_cnn
def test_anchor_text(predict_type, present, use_proba, use_unk, threshold):
# load data and create train and test sets
data, labels = movie_sentiment()
train, test, train_labels, test_labels = train_test_split(data, labels, test_size=.2, random_state=0)
train_labels = np.array(train_labels)
# apply CountVectorizer
vectorizer = CountVectorizer(min_df=1)
vectorizer.fit(train)
# train Logistic Regression model
clf = LogisticRegression()
clf.fit(vectorizer.transform(train), train_labels)
# define predict function
if predict_type == 'proba':
predict_fn = lambda x: clf.predict_proba(vectorizer.transform(x))
elif predict_type == 'class':
predict_fn = lambda x: clf.predict(vectorizer.transform(x))
# test explainer initialization
explainer = AnchorText(nlp, predict_fn)
assert explainer.predict_fn(['book']).shape == (1,)
# test sampling function
text = 'This is a good book .'
num_samples = 100
sample_prob_unk = .5
top_n = 500
words, positions, sample_fn = explainer.get_sample_fn(text, use_proba=use_proba, use_unk=use_unk)
raw_data, data, labels = sample_fn(present, num_samples, sample_prob_unk=sample_prob_unk, top_n=top_n)
if use_proba: # check that words in present are in the proposed anchor
assert len(present) * data.shape[0] == data[:, :-1].sum() # exclude '.'
if use_unk:
# get list of unique words
all_words = []
for i in range(raw_data.shape[0]):
all_words.append(raw_data[i][0].split())
all_words = [word for word_list in all_words for word in word_list]
# unique words = words in text + UNK
assert len(np.unique(all_words)) == len(text.split()) + 1
# check nb of UNKs
assert data.shape[0] * data.shape[1] - data.sum() == Counter(all_words)['UNK']
# test explanation
explanation = explainer.explain(text, threshold=threshold, use_proba=use_proba, use_unk=use_unk)
assert explanation['precision'] >= threshold
# check if sampled sentences are not cut short
keys = ['covered', 'covered_true', 'covered_false']
for i in range(len(explanation['raw']['feature'])):
example_dict = explanation['raw']['examples'][i]
for k in keys:
for example in example_dict[k]:
assert example[0][-1] in ['.', 'K']
def test_anchor_text(lr_classifier, text, n_punctuation_marks, n_unique_words,
predict_type, anchor, use_similarity_proba, use_unk,
threshold):
# test parameters
num_samples = 100
sample_proba = .5
top_n = 500
temperature = 1.
n_covered_ex = 5 # number of examples where the anchor applies to be returned
# fit and initialise predictor
clf, preprocessor = lr_classifier
predictor = predict_fcn(predict_type, clf, preproc=preprocessor)
# test explainer initialization
explainer = AnchorText(nlp, predictor)
assert explainer.predictor(['book']).shape == (1, )
# setup explainer
perturb_opts = {
'use_similarity_proba': use_similarity_proba,
'sample_proba': sample_proba,
'temperature': temperature,
}
explainer.n_covered_ex = n_covered_ex
explainer.set_words_and_pos(text)
explainer.set_sampler_perturbation(use_unk, perturb_opts, top_n)
explainer.set_data_type(use_unk)
if predict_type == 'proba':
label = np.argmax(predictor([text])[0])
elif predict_type == 'class':
label = predictor([text])[0]
explainer.instance_label = label
assert isinstance(explainer.dtype, str)
assert len(explainer.punctuation) == n_punctuation_marks
assert len(explainer.words) == len(explainer.positions)
# test sampler
cov_true, cov_false, labels, data, coverage, _ = explainer.sampler(
(0, anchor), num_samples)
if not anchor:
assert coverage == -1
if use_similarity_proba and len(
anchor
) > 0: # check that words in present are in the proposed anchor
assert len(anchor) * data.shape[0] == data[:, anchor].sum()
if use_unk:
# get list of unique words
all_words = explainer.words
# unique words = words in text + UNK
assert len(np.unique(all_words)) == n_unique_words
# test explanation
explanation = explainer.explain(
text,
use_unk=use_unk,
threshold=threshold,
use_similarity_proba=use_similarity_proba,
)
assert explanation.precision >= threshold
assert explanation.raw['prediction'].item() == label
assert explanation.meta.keys() == DEFAULT_META_ANCHOR.keys()
assert explanation.data.keys() == DEFAULT_DATA_ANCHOR.keys()
# check if sampled sentences are not cut short
keys = ['covered_true', 'covered_false']
for i in range(len(explanation.raw['feature'])):
example_dict = explanation.raw['examples'][i]
for k in keys:
for example in example_dict[k]:
# check that we have perturbed the sentences
if use_unk:
assert 'UNK' in example or example.replace(
' ', '') == text.replace(' ', '')
else:
assert 'UNK' not in example
assert example[-1] in ['.', 'K']
val_labels = np.array(val_labels)
# define and train an cnn model
vectorizer = CountVectorizer(min_df=1)
clf = LogisticRegression(solver='liblinear')
pipeline = Pipeline([('preprocess', vectorizer), ('clf', clf)])
print('Training ...')
pipeline.fit(train, train_labels)
print('Training done!')
preds_train = pipeline.predict(train)
preds_val = pipeline.predict(val)
preds_test = pipeline.predict(test)
print('Train accuracy', accuracy_score(train_labels, preds_train))
print('Validation accuracy', accuracy_score(val_labels, preds_val))
print('Test accuracy', accuracy_score(test_labels, preds_test))
print("Saving Model to model.joblib")
joblib.dump(pipeline, "model.joblib")
print("Creating Anchor Text explainer")
spacy_language_model = 'en_core_web_md'
spacy_model(model=spacy_language_model)
nlp = spacy.load(spacy_language_model)
anchors_text = AnchorText(nlp=nlp, predictor=lambda x: pipeline.predict(x))
# Test explanations locally
expl = anchors_text.explain("the actors are very bad")
print(expl)
def atext_explainer(lr_classifier, english_spacy_model, movie_sentiment_data):
predictor = predict_fcn(predict_type='class',
clf=lr_classifier,
preproc=movie_sentiment_data['preprocessor'])
atext = AnchorText(nlp=english_spacy_model, predictor=predictor)
return atext
# define and train an cnn model
vectorizer = CountVectorizer(min_df=1)
clf = LogisticRegression(solver='liblinear')
pipeline = Pipeline([('preprocess', vectorizer), ('clf', clf)])
print('Training ...')
pipeline.fit(train, train_labels)
print('Training done!')
preds_train = pipeline.predict(train)
preds_val = pipeline.predict(val)
preds_test = pipeline.predict(test)
print('Train accuracy', accuracy_score(train_labels, preds_train))
print('Validation accuracy', accuracy_score(val_labels, preds_val))
print('Test accuracy', accuracy_score(test_labels, preds_test))
print("Saving Model to model.joblib")
joblib.dump(pipeline, "model.joblib")
print("Creating Anchor Text explainer")
spacy_language_model = 'en_core_web_md'
spacy_model(model=spacy_language_model)
nlp = spacy.load(spacy_language_model)
predict_fn = lambda x: pipeline.predict(x)
anchors_text = AnchorText(nlp, predict_fn)
# Test explanations locally
expl = anchors_text.explain("the actors are very bad")
print(expl)