def test_unmarked_date_column():
np.random.seed(0)
df_twitter_train, df_twitter_test = utils.get_twitter_sentiment_multilabel_classification_dataset(
)
column_descriptions = {
'airline_sentiment': 'output',
'airline': 'categorical',
'text': 'nlp',
'tweet_location': 'categorical',
'user_timezone': 'categorical'
# tweet_created is our date column. We want to test that "forgetting" to mark it as a date column throws a user warning before throwing the error
# , 'tweet_created': 'date'
}
ml_predictor = Predictor(type_of_estimator='classifier',
column_descriptions=column_descriptions)
with warnings.catch_warnings(record=True) as w:
try:
ml_predictor.train(df_twitter_train)
except TypeError as e:
pass
print('Here are the caught warnings:')
print(w)
assert len(w) == 1
def test_nlp_multilabel_classification():
np.random.seed(0)
df_twitter_train, df_twitter_test = utils.get_twitter_sentiment_multilabel_classification_dataset(
)
column_descriptions = {
'airline_sentiment': 'output',
'airline': 'categorical',
'text': 'nlp',
'tweet_location': 'categorical',
'user_timezone': 'categorical',
'tweet_created': 'date'
}
ml_predictor = Predictor(type_of_estimator='classifier',
column_descriptions=column_descriptions)
ml_predictor.train(df_twitter_train)
test_score = ml_predictor.score(df_twitter_test,
df_twitter_test.airline_sentiment,
verbose=0)
# Make sure our score is good, but not unreasonably good
print('test_score')
print(test_score)
assert 0.67 < test_score < 0.79
def test_multilabel_classification_predict_on_Predictor_instance():
np.random.seed(0)
df_twitter_train, df_twitter_test = utils.get_twitter_sentiment_multilabel_classification_dataset()
# Note that this does not take 'text' into account, intentionally (as that takes a while longer to train)
ml_predictor = utils.train_basic_multilabel_classifier(df_twitter_train)
predictions = ml_predictor.predict(df_twitter_test)
test_score = accuracy_score(predictions, df_twitter_test.airline_sentiment)
# Make sure our score is good, but not unreasonably good
print('test_score')
print(test_score)
assert 0.72 < test_score < 0.77
def test_multilabel_classification_predict_on_Predictor_instance():
np.random.seed(0)
df_twitter_train, df_twitter_test = utils.get_twitter_sentiment_multilabel_classification_dataset()
ml_predictor = utils.train_basic_multilabel_classifier(df_twitter_train)
predictions = ml_predictor.predict(df_twitter_test)
test_score = accuracy_score(predictions, df_twitter_test.airline_sentiment)
# Right now we're getting a score of -.205
# Make sure our score is good, but not unreasonably good
print('test_score')
print(test_score)
assert 0.67 < test_score < 0.79
def test_multilabel_classification():
np.random.seed(0)
df_twitter_train, df_twitter_test = utils.get_twitter_sentiment_multilabel_classification_dataset(
)
ml_predictor = utils.train_basic_multilabel_classifier(df_twitter_train)
test_score = ml_predictor.score(df_twitter_test,
df_twitter_test.airline_sentiment,
verbose=0)
# Right now we're getting a score of -.205
# Make sure our score is good, but not unreasonably good
print('test_score')
print(test_score)
assert 0.67 < test_score < 0.79
def test_saving_trained_pipeline_multilabel_classification():
np.random.seed(0)
df_twitter_train, df_twitter_test = utils.get_twitter_sentiment_multilabel_classification_dataset(
)
ml_predictor = utils.train_basic_multilabel_classifier(df_twitter_train)
file_name = ml_predictor.save(str(random.random()))
with open(file_name, 'rb') as read_file:
saved_ml_pipeline = dill.load(read_file)
os.remove(file_name)
test_score = saved_ml_pipeline.score(df_twitter_test,
df_twitter_test.airline_sentiment)
# Right now we're getting a score of -.205
# Make sure our score is good, but not unreasonably good
print('test_score')
print(test_score)
assert 0.67 < test_score < 0.79
def test_getting_single_predictions_nlp_date_multilabel_classification(
model_name=None):
# auto_ml does not support multilabel classification for deep learning at the moment
if model_name == 'DeepLearningClassifier':
return
np.random.seed(0)
df_twitter_train, df_twitter_test = utils.get_twitter_sentiment_multilabel_classification_dataset(
)
column_descriptions = {
'airline_sentiment': 'output',
'airline': 'categorical',
'text': 'nlp',
'tweet_location': 'categorical',
'user_timezone': 'categorical',
'tweet_created': 'date'
}
ml_predictor = Predictor(type_of_estimator='classifier',
column_descriptions=column_descriptions)
ml_predictor.train(df_twitter_train, model_names=model_name)
file_name = ml_predictor.save(str(random.random()))
# if model_name == 'DeepLearningClassifier':
# from auto_ml.utils_models import load_keras_model
# saved_ml_pipeline = load_keras_model(file_name)
# else:
# with open(file_name, 'rb') as read_file:
# saved_ml_pipeline = dill.load(read_file)
saved_ml_pipeline = load_ml_model(file_name)
os.remove(file_name)
try:
keras_file_name = file_name[:-5] + '_keras_deep_learning_model.h5'
os.remove(keras_file_name)
except:
pass
df_twitter_test_dictionaries = df_twitter_test.to_dict('records')
# 1. make sure the accuracy is the same
predictions = []
for row in df_twitter_test_dictionaries:
predictions.append(saved_ml_pipeline.predict(row))
print('predictions')
print(predictions)
first_score = accuracy_score(df_twitter_test.airline_sentiment,
predictions)
print('first_score')
print(first_score)
# Make sure our score is good, but not unreasonably good
lower_bound = 0.73
# if model_name == 'LGBMClassifier':
# lower_bound = 0.655
assert lower_bound < first_score < 0.79
# 2. make sure the speed is reasonable (do it a few extra times)
data_length = len(df_twitter_test_dictionaries)
start_time = datetime.datetime.now()
for idx in range(1000):
row_num = idx % data_length
saved_ml_pipeline.predict(df_twitter_test_dictionaries[row_num])
end_time = datetime.datetime.now()
duration = end_time - start_time
print('duration.total_seconds()')
print(duration.total_seconds())
# It's very difficult to set a benchmark for speed that will work across all machines.
# On my 2013 bottom of the line 15" MacBook Pro, this runs in about 0.8 seconds for 1000 predictions
# That's about 1 millisecond per prediction
# Assuming we might be running on a test box that's pretty weak, multiply by 3
# Also make sure we're not running unreasonably quickly
# time_upper_bound = 3
# if model_name == 'XGBClassifier':
# time_upper_bound = 4
assert 0.2 < duration.total_seconds() < 15
# 3. make sure we're not modifying the dictionaries (the score is the same after running a few experiments as it is the first time)
predictions = []
for row in df_twitter_test_dictionaries:
predictions.append(saved_ml_pipeline.predict(row))
print('predictions')
print(predictions)
print('df_twitter_test_dictionaries')
print(df_twitter_test_dictionaries)
second_score = accuracy_score(df_twitter_test.airline_sentiment,
predictions)
print('second_score')
print(second_score)
# Make sure our score is good, but not unreasonably good
assert lower_bound < second_score < 0.79
def test_getting_single_predictions_multilabel_classification_with_dates():
np.random.seed(0)
df_twitter_train, df_twitter_test = utils.get_twitter_sentiment_multilabel_classification_dataset(
)
ml_predictor = utils.train_basic_multilabel_classifier(df_twitter_train)
file_name = ml_predictor.save(str(random.random()))
with open(file_name, 'rb') as read_file:
saved_ml_pipeline = dill.load(read_file)
os.remove(file_name)
df_twitter_test_dictionaries = df_twitter_test.to_dict('records')
# 1. make sure the accuracy is the same
predictions = []
for row in df_twitter_test_dictionaries:
predictions.append(saved_ml_pipeline.predict(row))
print('predictions')
print(predictions)
first_score = accuracy_score(df_twitter_test.airline_sentiment,
predictions)
print('first_score')
print(first_score)
# Make sure our score is good, but not unreasonably good
assert 0.67 < first_score < 0.79
# 2. make sure the speed is reasonable (do it a few extra times)
data_length = len(df_twitter_test_dictionaries)
start_time = datetime.datetime.now()
for idx in range(1000):
row_num = idx % data_length
saved_ml_pipeline.predict(df_twitter_test_dictionaries[row_num])
end_time = datetime.datetime.now()
duration = end_time - start_time
print('duration.total_seconds()')
print(duration.total_seconds())
# It's very difficult to set a benchmark for speed that will work across all machines.
# On my 2013 bottom of the line 15" MacBook Pro, this runs in about 0.8 seconds for 1000 predictions
# That's about 1 millisecond per prediction
# Assuming we might be running on a test box that's pretty weak, multiply by 3
# Also make sure we're not running unreasonably quickly
assert 0.2 < duration.total_seconds() < 3
# 3. make sure we're not modifying the dictionaries (the score is the same after running a few experiments as it is the first time)
predictions = []
for row in df_twitter_test_dictionaries:
predictions.append(saved_ml_pipeline.predict(row))
print('predictions')
print(predictions)
print('df_twitter_test_dictionaries')
print(df_twitter_test_dictionaries)
second_score = accuracy_score(df_twitter_test.airline_sentiment,
predictions)
print('second_score')
print(second_score)
# Make sure our score is good, but not unreasonably good
assert 0.67 < second_score < 0.79
