def explain_pred(input_data, model):
y_preds = []
y_probs = []
encoded_htmls = []
for i in input_data:
expl = eli5.explain_prediction(
model.steps[-1][1],
i,
model.steps[0][1],
target_names=['Compliant', 'Not Compliant'],
top=10)
html_explanation = format_as_html(expl,
force_weights=False,
show_feature_values=True).replace(
"\n", "").strip()
encoded_html = base64.b64encode(
bytes(html_explanation, encoding='utf-8'))
encoded_htmls.append(encoded_html)
expl_dict = format_as_dict(expl)
targets = expl_dict['targets'][0]
target = targets['target']
y_pred = 1 if target.startswith('N') else 0
y_prob = targets['proba']
if len(i.split()) < 3:
# one or two words can't be non-compliant
y_pred = 0
y_prob = 1.0
y_prob = f'{round(y_prob, 3) * 100}%'
y_preds.append(y_pred)
y_probs.append(y_prob)
inferences = np.column_stack((y_probs, y_preds, encoded_htmls))
return inferences
def test_format_as_dict():
assert format_as_dict(
Explanation(
estimator='some estimator',
targets=[
TargetExplanation('y',
feature_weights=FeatureWeights(pos=[
FeatureWeight('a', np.float32(13.0))
],
neg=[])),
],
)) == {
'estimator':
'some estimator',
'targets': [
{
'target': 'y',
'feature_weights': {
'pos': [{
'feature': 'a',
'weight': 13.0,
'std': None,
'value': None
}],
'pos_remaining':
0,
'neg': [],
'neg_remaining':
0,
},
'score': None,
'proba': None,
'weighted_spans': None,
'heatmap': None,
},
],
'decision_tree':
None,
'description':
None,
'error':
None,
'feature_importances':
None,
'highlight_spaces':
None,
'is_regression':
False,
'method':
None,
'transition_features':
None,
'image':
None,
}
def sklearn_predict(lines, current_app):
estimator = current_app.config['ESTIMATOR']
explanations = []
y_preds = []
y_probs = []
for line in lines:
line = ' '.join(t for t in line.split() if t not in stopwords)
if not line.strip():
continue
expl = eli5.explain_prediction(
estimator.steps[-1][1],
line,
estimator.steps[0][1],
target_names=['Compliant', 'Not Compliant'],
top=10)
html_explanation = format_as_html(expl,
force_weights=False,
show_feature_values=True).replace(
"\n", "").strip()
explanations.append(html_explanation)
expl_dict = format_as_dict(expl)
targets = expl_dict['targets'][0]
target = targets['target']
y_pred = 1 if target.startswith('N') else 0
y_prob = targets['proba']
if len(line.split()) < 3:
# one or two words can't be non-compliant
y_pred = 0
y_prob = 1.0
y_preds.append(y_pred)
y_probs.append(y_prob)
y_probs = [f'{round(y_prob, 3) * 100}%' for y_prob in y_probs]
data = zip(y_preds, lines, lines, y_probs, explanations)
results = [
dict(y_pred=y, line=l, clean_line=cl, y_prob=y_prob, expl=expl)
for y, l, cl, y_prob, expl in data
]
return results
def explain_review_prediction():
"""
Explain a specific prediction using the eli5 library
"""
data = request.get_json(force=True)
# Use the original documents, not the corrected ones
target_names = ['negative', 'neutral', 'positive', 'very_negative', 'very_positive']
clf, vocabulary = load_clf_and_vocabulary(data['classifier'], data['vocabModel'], data['tfIdf'], False)
vect = CountVectorizer(vocabulary=vocabulary)
vect._validate_vocabulary()
# reviews = load_files(dir_path + '/../../data/reviews/not_corrected')
# text_train, text_test, y_train, y_test = train_test_split(reviews.data, reviews.target, test_size=0.2, random_state=0)
# if data['tfIdf']:
# if data['vocabModel'] == 'unigram':
# vect = CountVectorizer(min_df=5, stop_words=stopwords, ngram_range=(1, 1)).fit(text_train)
# elif data['vocabModel'] == 'bigram':
# vect = CountVectorizer(min_df=5, stop_words=stopwords, ngram_range=(2, 2)).fit(text_train)
# elif data['vocabModel'] == 'trigram':
# vect = CountVectorizer(min_df=5, stop_words=stopwords, ngram_range=(3, 3)).fit(text_train)
# else:
# if data['vocabModel'] == 'unigram':
# vect = CountVectorizer(min_df=5, stop_words=stopwords, ngram_range=(1, 1)).fit(text_train)
# elif data['vocabModel'] == 'bigram':
# vect = CountVectorizer(min_df=5, stop_words=stopwords, ngram_range=(2, 2)).fit(text_train)
# elif data['vocabModel'] == 'trigram':
# vect = CountVectorizer(min_df=5, stop_words=stopwords, ngram_range=(3, 3)).fit(text_train)
if data['classifier'] == 'LR':
explanation = explain_prediction.explain_prediction_linear_classifier(clf, data['review'], vec=vect, top=10, target_names=target_names)
div = html.format_as_html(explanation, include_styles=False)
style = html.format_html_styles()
txt = text.format_as_text(explanation, show=eli5.formatters.fields.ALL, highlight_spaces=True, show_feature_values=True)
print(txt)
return jsonify({
'div': div,
'style': style
})
elif data['classifier'] == 'SVM' or data['classifier'] == 'MLP':
te = TextExplainer(n_samples=100, clf=LogisticRegression(solver='newton-cg'), vec=vect, random_state=0)
te.fit(data['review'], clf.predict_proba)
explanation = te.explain_prediction(top=10, target_names=target_names)
div = html.format_as_html(explanation, include_styles=False)
style = html.format_html_styles()
distorted_texts = []
for sample in te.samples_:
sample_explanation = explain_prediction.explain_prediction_linear_classifier(te.clf_, sample, vec=te.vec_, top=10, target_names=target_names)
dict_explanation = as_dict.format_as_dict(sample_explanation)
curr = {
'text': sample
}
for c in dict_explanation['targets']:
if c['target'] == 'negative':
curr['negative'] = c['proba']
elif c['target'] == 'neutral':
curr['neutral'] = c['proba']
elif c['target'] == 'positive':
curr['positive'] = c['proba']
elif c['target'] == 'very_negative':
curr['very_negative'] = c['proba']
elif c['target'] == 'very_positive':
curr['very_positive'] = c['proba']
distorted_texts.append(curr)
review_explanation = as_dict.format_as_dict(explanation)
probabilities = {}
for c in review_explanation['targets']:
if c['target'] == 'negative':
probabilities['negative'] = c['proba']
elif c['target'] == 'neutral':
probabilities['neutral'] = c['proba']
elif c['target'] == 'positive':
probabilities['positive'] = c['proba']
elif c['target'] == 'very_negative':
probabilities['very_negative'] = c['proba']
elif c['target'] == 'very_positive':
probabilities['very_positive'] = c['proba']
return jsonify({
'div': div,
'style': style,
'original_text': data['review'],
'probabilities': probabilities,
'distorted_texts': distorted_texts,
'metrics': te.metrics_
})