def improve_predictions(
probability_predictions_file='../models/pred_ml_proba.pkl',
out_file='../models/pred_ml_improved.pkl',
use_infer_topology=True):
print "> IMPROVING PREDICTIONS\n--- Forcing at least one label (most likely)"
print "Loading probability predictions"
y_pred_proba = joblib.load(probability_predictions_file)
#Because we use a one-versus-rest classifier, there may be documents without any labels
#We deal with this by adding the most likely labels
y_pred_improved = np.zeros(y_pred_proba.shape, dtype=np.int_)
print "Converting to binary predictions"
y_pred = np.where(y_pred_proba >= 0.5, 1, 0)
for i, (prediction,
prediction_proba) in enumerate(tqdm(zip(y_pred, y_pred_proba))):
if sum(prediction) == 0:
most_likely_label_index = np.argmax(prediction_proba)
y_pred_improved[i, most_likely_label_index] = 1
y_pred_improved[i] += prediction
print np.sum(np.subtract(y_pred_improved, y_pred)), "labels added"
if use_infer_topology:
print "> IMPROVING PREDICTIONS\n--- Topology rules"
print "Loading train set y-values"
y_train, filenames_train = dataset.load_train_y()
rules = infer_topology.infer_topology_rules(y_train)
y_pred_improved = infer_topology.apply_topology_rules(
rules, y_pred_improved)
print "Saving to file"
joblib.dump(y_pred_improved, out_file)
print "Done!\n---"
def improve_predictions(
probability_predictions_file="../models/pred_ml_proba.pkl",
out_file="../models/pred_ml_improved.pkl",
use_infer_topology=True,
):
print "> IMPROVING PREDICTIONS\n--- Forcing at least one label (most likely)"
print "Loading probability predictions"
y_pred_proba = joblib.load(probability_predictions_file)
# Because we use a one-versus-rest classifier, there may be documents without any labels
# We deal with this by adding the most likely labels
y_pred_improved = np.zeros(y_pred_proba.shape, dtype=np.int_)
print "Converting to binary predictions"
y_pred = np.where(y_pred_proba >= 0.5, 1, 0)
for i, (prediction, prediction_proba) in enumerate(tqdm(zip(y_pred, y_pred_proba))):
if sum(prediction) == 0:
most_likely_label_index = np.argmax(prediction_proba)
y_pred_improved[i, most_likely_label_index] = 1
y_pred_improved[i] += prediction
print np.sum(np.subtract(y_pred_improved, y_pred)), "labels added"
if use_infer_topology:
print "> IMPROVING PREDICTIONS\n--- Topology rules"
print "Loading train set y-values"
y_train, filenames_train = dataset.load_train_y()
rules = infer_topology.infer_topology_rules(y_train)
y_pred_improved = infer_topology.apply_topology_rules(rules, y_pred_improved)
print "Saving to file"
joblib.dump(y_pred_improved, out_file)
print "Done!\n---"
print np.sum(np.subtract(predictions_improved,predictions)), "labels added"
return predictions_improved
def print_topology(rules):
label_list = dataset.load_labels()
root_nodes = set([y for x,y in rules])
tree = {root:[] for root in root_nodes}
for rule in rules:
from_label, to_label = rule
tree[to_label].append(from_label)
for parent, children in tree.iteritems():
print label_list[parent]
for child in children:
print " >", label_list[child]
if __name__ == "__main__":
with open(DATA_FOLDER+'labels_int.p', 'r') as f:
y_dict = pickle.load(f)
rules = infer_topology_rules(y_dict.values(), verbose=False)
print_topology(rules)
y, _ = dataset.load_train_y()
rules = infer_topology_rules(y, verbose=False)
print_topology(rules)
