def explain_model_weights():
"""
Explain the weights/parameters of a certain model
"""
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'], data['corrected'])
vect = CountVectorizer(vocabulary=vocabulary)
vect._validate_vocabulary()
if data['classifier'] == 'LR':
explanation = explain_weights.explain_linear_classifier_weights(clf, vec=vect, target_names=target_names)
div = html.format_as_html(explanation, include_styles=False)
style = html.format_html_styles()
return jsonify({
'div': div,
'style': style
})
def get_lime_analyze_as_html(self,
query,
indicies,
max_len=50,
max_replace=0.7,
top_targets=None):
query = query.replace(' ', '')
for i in indicies:
if i >= len(query):
raise ValueError()
def build_query_info_html(query, index):
_html_format = ''
_html_format += '<p>' + 'query: ' + query + '</p><br>'
_html_format += '<p>' + 'index: ' + str(index) + '</p><br>'
_html_format += '<p>' + 'char: ' + query[index] + '</p><br>'
return _html_format
explainers = self._lime_analyze(query,
indicies,
max_len,
max_replace,
top_targets=top_targets)
html_format = ''
html_format += '<div>'
for i, e in zip(indicies, explainers):
html_format += '<table class="table"><tr>'
html_format += '<td><div>' + build_query_info_html(
query, i) + '</div></td>'
html_format += '<td><div>' + format_as_html(e).replace(
'\n', '') + '</div></td>'
html_format += '</tr></table>'
html_format += '<hr>'
html_format += '</div>'
return html_format
def _repr_html_(self):
""" HTML formatting for the notebook.
"""
from eli5.formatters import fields
from eli5.formatters.html import format_as_html
return format_as_html(self, force_weights=False, show=fields.WEIGHTS)
def get_prediction_message(self, text):
class_prediction, eli5_prediction = self.predict(text)
prediction_message = self.classes_dict[class_prediction]
return prediction_message, format_as_html(eli5_prediction)
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_
})