class Preprocessor:
def __init__(self):
self.io_utils = IOUtility()
self.settings = self.io_utils.get_settings()
self.profile = self.io_utils.get_profile()
self.enc = None
def load_encoder(self):
"""
Loads OneHotEncoder object if present in cache provided 'cache_encoder' is True in settings.
If any of the above condition is false, it creates a new encoder object, trains it and caches it.
"""
encoder_cache_file_name = "encoder.cache"
if self.settings['cache']['cache_encoder']:
self.enc = self.io_utils.load_from_cache(encoder_cache_file_name)
if self.enc is None:
logging.debug("Encoder is not cached or cache is disabled")
self.enc = self.__train_encoder(encoder_cache_file_name)
else:
logging.debug("Successfully loaded encoder from cache")
return self.enc
def encode(self, x):
return self.enc.transform(x)
def __train_encoder(self, encoder_cache_file_name):
"""
Trains and caches the encoder in mini batch mode
:param encoder_cache_file_name:
:return: trained encoder object
"""
model = self.io_utils.get_model()
_use_cols = misc_utils.get_cols_to_load(labels=False, load_int_cols=False, load_cat_cols=True,
ignore_features=model['ignore_features'])
train_chunks = self.io_utils.load_train_set(use_cols=_use_cols, converters=misc_utils.get_converters(_use_cols),
chunk_size=1500000)
enc = OneHotEncoder()
start = datetime.now()
logging.debug("Training Encoder...")
for i, chunk in enumerate(train_chunks):
enc.partial_fit(chunk.as_matrix())
logging.debug("Trained encoder with chunk %d\t%s" % (i+1, str(datetime.now() - start)))
logging.debug("Finished training encoder in \t%s" % str(datetime.now() - start))
self.io_utils.save_to_cache(enc, encoder_cache_file_name)
return enc
class Classifier:
def __init__(self, preprocessor):
self.preprocessor = preprocessor
self.io_utils = IOUtility()
self.settings = self.io_utils.get_settings()
self.models = {
"SGDC": linear_model.SGDClassifier,
"BNB": naive_bayes.BernoulliNB,
"MNB": naive_bayes.MultinomialNB,
"RF": ensemble.RandomForestClassifier,
"ET": ensemble.ExtraTreesClassifier,
"ADA": ensemble.AdaBoostClassifier,
"GB": ensemble.GradientBoostingClassifier}
def run(self):
model, model_name, model_settings = self.__load_model()
self.train(model, model_settings['ignore_features'])
log_loss_score = self.predict(model, model_settings['ignore_features'])
logging.info("Log loss score for model '%s' is '%f'" % (model_name, log_loss_score))
def train(self, model, ignore_features):
logging.debug("Loading train set...")
_use_cols = misc_utils.get_cols_to_load(labels=True, load_int_cols=True, load_cat_cols=True,
ignore_features=ignore_features)
train_df = self.io_utils.load_train_set(use_cols=_use_cols, converters=misc_utils.get_converters(_use_cols),
chunk_size=self.settings['chunk_size'])
start = datetime.now()
logging.debug("Training Model...")
if self.settings['chunk_size'] is not None:
for i, chunk in enumerate(train_df):
logging.debug("Processing chunk %d\t%s" % (i + 1, str(datetime.now() - start)))
self.__process_train_chunk(chunk, model, ignore_features)
else:
self.__process_train_chunk(train_df, model, ignore_features)
logging.debug("Finished training in \t%s" % str(datetime.now() - start))
def predict(self, model, ignore_features):
logging.debug("Loading test set...")
_use_cols = misc_utils.get_cols_to_load(labels=True, load_int_cols=True, load_cat_cols=True,
ignore_features=ignore_features)
test_df = self.io_utils.load_validation_set(use_cols=_use_cols,
converters=misc_utils.get_converters(_use_cols),
chunk_size=self.settings['chunk_size'])
start = datetime.now()
logging.debug("Predicting probabilities...")
actual_labels = []
predicted_labels = []
if self.settings['chunk_size'] is not None:
for i, chunk in enumerate(test_df):
logging.debug("Processing chunk %d\t%s" % (i + 1, str(datetime.now() - start)))
y, predictions = self.__process_test_chunk(chunk, model, ignore_features)
actual_labels += y.tolist()
predicted_labels += self.__cap_predictions(predictions)
else:
y, predictions = self.__process_test_chunk(test_df, model, ignore_features)
actual_labels += y.tolist()
predicted_labels += self.__cap_predictions(predictions)
logging.debug("Finished predicting in \t%s" % str(datetime.now() - start))
return self.evaluate_model(actual_labels, predicted_labels)
def __process_test_chunk(self, chunk, model, ignore_features):
y = chunk['Label']
X = self.__preprocess(chunk, ignore_features)
predictions = model.predict_proba(X)
return y, predictions
def __process_train_chunk(self, chunk, model, ignore_features):
y = chunk['Label']
X = self.__preprocess(chunk, ignore_features)
model.partial_fit(X, y, classes=[0, 1])
def __preprocess(self, chunk, ignore_features):
int_features = sparse.coo_matrix(chunk[misc_utils.get_integer_cols(ignore_features=ignore_features)]).tocsr()
cat_features = chunk[misc_utils.get_categorical_cols(ignore_features=ignore_features)].as_matrix()
cat_features = self.preprocessor.encode(cat_features)
X = sparse.hstack((int_features, cat_features))
return X
def __load_model(self):
model_name = self.settings['model']
model_id = model_name.split('_')[0]
model = self.models[model_id](random_state=SEED)
model_settings = self.io_utils.get_model(model_name)
model_params = model_settings["params"]
model.set_params(**model_params)
return model, model_name, model_settings
@staticmethod
def evaluate_model(actual_labels, predicted_labels):
def log_loss(act, pred):
epsilon = 1e-15
pred = sp.maximum(epsilon, pred)
pred = sp.minimum(1-epsilon, pred)
ll = sum(act*sp.log(pred) + sp.subtract(1, act)*sp.log(sp.subtract(1, pred)))
ll = ll * -1.0/len(act)
return ll
return log_loss(actual_labels, predicted_labels)
@staticmethod
def __cap_predictions(preds):
return_val = []
for p in preds:
prob = p[1]
if prob > 0.98: prob = 0.98
elif prob < 0.02: prob = 0.02
return_val.append(prob)
return return_val