def test_compute_prediction(self):
model = Mljar(project=self.proj_title,
experiment=self.expt_title,
algorithms=['rfc'],
metric='logloss',
validation_kfolds=3,
tuning_mode='Normal',
single_algorithm_time_limit=1)
self.assertTrue(model is not None)
# fit models and wait till all models are trained
model.fit(X=self.X, y=self.y, dataset_title='My dataset')
# get project id
project_id = model.project.hid
# get model id
model_id = model.selected_algorithm.hid
dc = DatasetClient(project_id)
init_datasets_cnt = len(dc.get_datasets())
# compute predictions
pred = Mljar.compute_prediction(self.X, model_id, project_id)
# compute score
score = self.mse(pred, self.y)
self.assertTrue(score < 0.9)
# check if dataset was removed
self.assertEqual(init_datasets_cnt, len(dc.get_datasets()))
# run predictions again, but keep dataset
pred = Mljar.compute_prediction(self.X,
model_id,
project_id,
keep_dataset=True)
self.assertEqual(init_datasets_cnt + 1,
len(dc.get_datasets())) # should be one more
def test_usage_with_validation_dataset(self):
#Test usage with validation dataset.
model = Mljar(project=self.proj_title,
experiment=self.expt_title,
algorithms=['xgb'],
tuning_mode='Normal',
single_algorithm_time_limit=1)
self.assertTrue(model is not None)
# load validation data
df = pd.read_csv('tests/data/test_1_vald.csv')
cols = ['sepal length', 'sepal width', 'petal length', 'petal width']
target = 'class'
X_vald = df[cols]
y_vald = df[target]
# fit models and wait till all models are trained
model.fit(X=self.X,
y=self.y,
validation_data=(X_vald, y_vald),
wait_till_all_done=False)
# wait some time
time.sleep(80)
# run prediction
pred = model.predict(self.X)
# get MSE
score = self.mse(pred, self.y)
self.assertTrue(score < 0.9)
# check default validation
self.assertEqual(model.selected_algorithm.validation_scheme,
"With dataset")
def test_wrong_input_dim(self):
with self.assertRaises(IncorrectInputDataException) as context:
model = Mljar(project=self.proj_title, experiment=self.expt_title)
samples = 100
columns = 10
X = np.random.rand(samples, columns)
y = np.random.choice([0, 1], samples + 1, replace=True)
model.fit(X, y)
def mljar_compute(X_train, y_train, X_test, dataset_id, seed):
from mljar import Mljar
#os.environ['MLJAR_TOKEN'] = '' # set token here or in your env
if 'MLJAR_TOKEN' not in os.environ:
raise Exception('Missing MLJAR_TOKEN, please set it.')
mlj = Mljar('DatasetId_{0}'.format(dataset_id),
'Seed_{0}'.format(seed),
metric = 'logloss',
algorithms = ['xgb', 'lgb', 'rfc', 'etc', 'mlp'],
tuning_mode = 'Sport',
create_ensemble = True,
single_algorithm_time_limit = 10)
mlj.fit(X_train, y_train)
response = mlj.predict(X_test)
return response
def test_usage_with_defaults(self):
'''
Test usage with defaults.
'''
model = Mljar(project=self.proj_title, experiment=self.expt_title)
self.assertTrue(model is not None)
# fit models and wait till all models are trained
model.fit(X=self.X, y=self.y, wait_till_all_done=False)
# wait some time
time.sleep(120) # wait a little longer - there are a lot of models
# run prediction
pred = model.predict(self.X)
# get MSE
score = self.mse(pred, self.y)
self.assertTrue(score < 0.5)
# check default validation
self.assertEqual(model.selected_algorithm.validation_scheme,
"5-fold CV, Shuffle, Stratify")
def test_basic_usage(self):
'''
Test the most common usage.
'''
model = Mljar(project=self.proj_title,
experiment=self.expt_title,
algorithms=['xgb'],
metric='logloss',
validation_kfolds=3,
tuning_mode='Normal')
self.assertTrue(model is not None)
# fit models and wait till all models are trained
model.fit(X=self.X, y=self.y)
# run prediction
pred = model.predict(self.X)
# get MSE
score = self.mse(pred, self.y)
self.assertTrue(score < 0.1)
def main():
# login to numerai to get token
print 'Login into Numer.ai'
api = numerapi.NumerAPI(NUMERAI_USER, NUMERAI_PASS)
# get datasets
print 'Get dataset'
if not os.path.isfile(TRAIN_FNAME):
api.download_current_dataset(dest_path='.', unzip=True)
# read datasets
train = pd.read_csv(TRAIN_FNAME)
test = pd.read_csv(TEST_FNAME)
print 'Numer.ai data downloaded'
print 'Train shape', train.shape, 'test shape', test.shape
X_train = train[train.columns[:50]]
y_train = train['target']
X_test = test
print 'Create MLJAR project and experiment'
models = Mljar(
project='Auto-trading',
experiment="Raw data",
metric='logloss',
validation_kfolds=5, # we will use 5-fold CV with stratify and shuffle
validation_shuffle=True,
validation_stratify=True,
algorithms=['xgb', 'lgb',
'mlp'], # select Xgboost, LightGBM and Neural Network
tuning_mode=
'Normal', # number of models to be checked for each algorithm
single_algorithm_time_limit=5) # 5 minutes for training single model
print 'Train models:'
# fit models - that's all, only one line of code ;)
models.fit(X_train, y_train)
# get predictions on test data
predictions = models.predict(X_test)
# save predictions to file
predictions.to_csv(PREDICTIONS_FNAME, index=False)
result = api.upload_prediction(PREDICTIONS_FNAME)
print 'Your score:', result['submission']['accuracy_score']
def test_retrive_models(self):
#Test scenario, when user create project, fit models, and try to once
#again run project. In this case, there will be no additional computations,
#all models will be simply retrived from existing project.
model = Mljar(project=self.proj_title,
experiment=self.expt_title,
algorithms=['xgb'],
metric='logloss',
validation_kfolds=3,
tuning_mode='Normal',
single_algorithm_time_limit=1)
self.assertTrue(model is not None)
# fit models and wait till all models are trained
model.fit(X=self.X, y=self.y)
# run prediction
pred = model.predict(self.X)
# get MSE
score = self.mse(pred, self.y)
self.assertTrue(score < 0.1)
# re-use already trained models
# call fit but models are already trained
# should be retrived - this should not be longer than 3 minutes
start_time = time.time()
model.fit(X=self.X, y=self.y)
end_time = time.time()
self.assertTrue(end_time - start_time < 3 * 60)
# check prediction
pred = model.predict(self.X)
# get MSE
score_2 = self.mse(pred, self.y)
self.assertTrue(score_2 < 0.1)
# scores should be the same
self.assertTrue(np.abs(score - score_2) < 1e-3)
# re-use project
start_time = time.time()
model_2 = Mljar(project=self.proj_title,
experiment=self.expt_title,
algorithms=['xgb'],
metric='logloss',
validation_kfolds=3,
tuning_mode='Normal',
single_algorithm_time_limit=1)
self.assertTrue(model_2 is not None)
# re-use trained models
model_2.fit(X=self.X, y=self.y)
end_time = time.time()
# it should not take longer than 5 minutes
self.assertTrue(end_time - start_time < 5 * 60)
# run prediction
pred = model_2.predict(self.X)
# get MSE
score_3 = self.mse(pred, self.y)
self.assertTrue(score_3 < 0.1)
# scores should be the same
self.assertTrue(np.abs(score - score_3) < 1e-3)
def test_usage_with_train_split(self):
'''
Test usage with train split.
'''
model = Mljar(project=self.proj_title,
experiment=self.expt_title,
validation_train_split=0.8,
algorithms=['xgb'],
tuning_mode='Normal')
self.assertTrue(model is not None)
# fit models and wait till all models are trained
model.fit(X=self.X, y=self.y, wait_till_all_done=False)
# wait some time
time.sleep(60)
# run prediction
pred = model.predict(self.X)
# get MSE
score = self.mse(pred, self.y)
self.assertTrue(score < 0.5)
# check default validation
self.assertEqual(model.selected_algorithm.validation_scheme,
"Split 80/20, Shuffle, Stratify")
def test_non_wait_fit(self):
'''
Test the non wait fit.
'''
model = Mljar(project=self.proj_title,
experiment=self.expt_title,
algorithms=['xgb'],
metric='logloss',
validation_kfolds=3,
tuning_mode='Normal',
single_algorithm_time_limit=1)
self.assertTrue(model is not None)
# fit models, just start computation and do not wait
start_time = time.time()
model.fit(X=self.X, y=self.y, wait_till_all_done=False)
end_time = time.time()
# time to initialize models should not be greater than 5 minutes
self.assertTrue(end_time - start_time < 5 * 60)
# run prediction
# good model is not guaranteed
# but there should be at least one
max_trys = 50
pred = None
while True:
pred = model.predict(self.X)
if pred is None:
# there is no model ready, please wait
time.sleep(10)
else:
break
max_trys -= 1
if max_trys <= 0:
break
self.assertTrue(pred is not None)
# get MSE
score = self.mse(pred, self.y)
print 'Score', score
self.assertTrue(score < 0.99)
# experiment='Ex 1.2', # for n_samples = 100, and 2 classes: 0, 1
# experiment='Ex 1.3', # for n_samples = 100, and 2 classes: 0, 2
# experiment='Ex 1.4', # for n_samples = 100, and 2 classes: 0, 2, and just re-run anew for probabilities
experiment='Ex 1.5', # as 1.4, but with validation_train_split=0.95
metric='auc',
algorithms=['rfc'],
validation_kfolds=None,
validation_shuffle=False,
validation_stratify=True,
validation_train_split=0.95,
tuning_mode='Normal', # Used Sport for experiments 1-3
create_ensemble=False,
single_algorithm_time_limit='1')
print("fit")
clf_mlj.fit(X, y) #,dataset_title="Ones and zeros")
print("predict")
pred_proba_mlj = clf_mlj.predict(X)
pred_proba_mlj = pred_proba_mlj.squeeze().values
print("pred_proba_mlj", pred_proba_mlj) # shows values = 0 or 1
pred_mlj = [2 if x == 1 else 0 for x in pred_proba_mlj]
print("prediction mljar == actual", (pred_mlj == y).all()) # returns True
# mljar_fit_params = {'max_features': 0.5, 'min_samples_split': 50, 'criterion': "gini", 'min_samples_leaf': 1}
# mljar_fit_params = {'max_features': 0.7, 'min_samples_split': 4, 'criterion': "entropy", 'min_samples_leaf': 2}
mljar_fit_params = clf_mlj.selected_algorithm.params['model_params'][
'fit_params']
print("mljar_fit_params", mljar_fit_params)
########################
target_col = df.columns[14]
# Let's define MLJAR project
mljar = Mljar(
project=
'UCI-Adult', # project's title that we will use to find it among our projects in MLJAR
experiment='Try tree methods', # experiment's title
validation_kfolds=5, # we will use 5-fold CV with stratify and shuffle
validation_shuffle=True,
validation_stratify=True,
metric='auc', # we will use Area Under ROC Curve (AUC) as a metric,
tuning_mode='Normal', # select tuning mode
algorithms=[
'rfc', 'xgb', 'lgb'
], # we want to tune Random Forest, LightGBM and Xgboost models
create_ensemble=True, # create ensemble of all models,
single_algorithm_time_limit=
5 # time limit for single algorithm training in minutes
)
# Run models prediction
mljar.fit(df[input_cols], df[target_col])
# Print out the most useful algorithm
print str(mljar.selected_algorithm)
# Run prediction on train dataset just for example
pred = mljar.predict(df[input_cols])
print 'Please go to your mljar account and check details of all models'
# experiment='Ex 2.2', # manual through website
# experiment='Ex 2.3', # validation_train_split=0.95
experiment=
'Ex 2.4', # cache the data locally at the same time as uploading to be able to replicate results
metric='auc',
algorithms=['rfc'],
validation_kfolds=None,
validation_shuffle=False,
validation_stratify=True,
validation_train_split=0.95,
tuning_mode='Normal',
create_ensemble=False,
single_algorithm_time_limit='1')
print("fit")
clf_mlj.fit(X, y) #,dataset_title="sklearn make_blobs")
# from mljar.client.project import ProjectClient
# clf_mlj.project = ProjectClient().create_project_if_not_exists(clf_mlj.project_title, clf_mlj.project_task)
# Until https://github.com/mljar/mljar-api-python/issues/2
# gets fixed,
# manually get the result to use in the prediction
# by browsing mljar.com
from mljar.client.result import ResultClient
client = ResultClient(clf_mlj.project.hid)
results = client.get_results(clf_mlj.experiment.hid)
# len(results) # returns 75
# results = client.get_results(None)
# len(results) # also returns 75
rid = raw_input(