def movie_sentiment_dataset():
"""
Load and prepare movie sentiment data.
"""
movies = fetch_movie_sentiment()
data = movies.data
labels = movies.target
train, test, train_labels, test_labels = train_test_split(data,
labels,
test_size=.2,
random_state=0)
train_labels = np.array(train_labels)
vectorizer = CountVectorizer(min_df=1)
vectorizer.fit(train)
return {
'X_train': train,
'y_train': train_labels,
'X_test': test,
'y_test': test_labels,
'preprocessor': vectorizer,
'metadata': {
'name': 'movie_sentiment'
},
}
def test_anchor_text():
skmodel = SKLearnServer(MOVIE_MODEL_URI)
skmodel.load()
movies = fetch_movie_sentiment()
anchor_text = AnchorText(skmodel.predict, None)
np.random.seed(0)
explanation = anchor_text.explain(movies.data[4:5])
exp_json = json.loads(explanation.to_json())
assert exp_json["meta"]["name"] == "AnchorText"
def test_anchor_text():
os.environ.clear()
skmodel = SKLearnModel("adult", MOVIE_MODEL_URI)
skmodel.load()
predictor = Predictor(skmodel)
anchor_text = AnchorText(predictor.predict_fn, None)
movies = fetch_movie_sentiment()
np.random.seed(0)
explanation = anchor_text.explain(movies.data[4:5])
exp_json = json.loads(explanation.to_json())
print(exp_json["data"]["anchor"])
def test_anchor_text():
os.environ.clear()
skmodel = SKLearnServer(MOVIE_MODEL_URI)
skmodel.load()
movies = fetch_movie_sentiment()
anchor_text = AnchorText(skmodel.predict, None)
np.random.seed(0)
explanation = anchor_text.explain(movies.data[4:5])
exp_json = json.loads(explanation.to_json())
print(exp_json["data"]["anchor"])
def test_movie_sentiment(return_X_y):
data = fetch_movie_sentiment(return_X_y=return_X_y)
if return_X_y:
assert len(data) == 2
X, y = data
else:
assert len(data) == 3
X = data.data
y = data.target
assert len(X) == len(y)
assert len(set(y)) == MOVIE_CLASSES
def prepare_data(test_size):
# load data
X, y = fetch_movie_sentiment(return_X_y=True)
# prepare data
data = pd.DataFrame()
data['text'] = X
data['labels'] = y
if math.isclose(test_size, 0.0):
return data, None
else:
train, test = train_test_split(data, test_size=test_size)
return train, test
def test_movie_sentiment(return_X_y):
try:
data = fetch_movie_sentiment(return_X_y=return_X_y)
except RequestException:
pytest.skip('Movie sentiment dataset URL down')
if return_X_y:
assert len(data) == 2
X, y = data
else:
assert len(data) == 3
X = data.data
y = data.target
assert len(X) == len(y)
assert len(set(y)) == MOVIE_CLASSES
import numpy as np
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score
from sklearn.model_selection import train_test_split
from alibi.datasets import fetch_movie_sentiment
from sklearn.pipeline import Pipeline
import joblib
from alibi.explainers import AnchorText
import spacy
from alibi.utils.download import spacy_model
# load data
movies = fetch_movie_sentiment()
movies.keys()
data = movies.data
labels = movies.target
target_names = movies.target_names
# define train and test set
np.random.seed(0)
train, test, train_labels, test_labels = train_test_split(data,
labels,
test_size=.2,
random_state=42)
train, val, train_labels, val_labels = train_test_split(train,
train_labels,
test_size=.1,
random_state=42)
train_labels = np.array(train_labels)
test_labels = np.array(test_labels)
def test_anchor_text(predict_type, present, use_similarity_proba, use_unk,
threshold):
# load data and create train and test sets
movies = fetch_movie_sentiment()
data = movies.data
labels = movies.target
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_proba = .5
top_n = 500
words, positions, sample_fn = explainer.get_sample_fn(
text,
use_similarity_proba=use_similarity_proba,
use_unk=use_unk,
sample_proba=sample_proba,
top_n=top_n)
raw_data, data, labels = sample_fn(present, num_samples)
if use_similarity_proba and len(
present
) > 0: # check that words in present are in the proposed anchor
assert len(present) * data.shape[0] == data[:, present].sum()
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_similarity_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']