predict.py

import numpy as np
from numpy import random as rand
import pickle

# DO NOT CHANGE THE NAME OF THIS METHOD OR ITS INPUT OUTPUT BEHAVIOR

# PLEASE BE CAREFUL THAT ERROR CLASS NUMBERS START FROM 1 AND NOT 0. THUS, THE FIFTY ERROR CLASSES ARE
# NUMBERED AS 1 2 ... 50 AND NOT THE USUAL 0 1 ... 49. PLEASE ALSO NOTE THAT ERROR CLASSES 33, 36, 38
# NEVER APPEAR IN THE TRAINING SET NOR WILL THEY EVER APPEAR IN THE SECRET TEST SET (THEY ARE TOO RARE)

# Input Convention
# X: n x d matrix in csr_matrix format containing d-dim (sparse) bag-of-words features for n test data points
# k: the number of compiler error class guesses to be returned for each test data point in ranked order

# Output Convention
# The method must return an n x k numpy nd-array (not numpy matrix or scipy matrix) of classes with the i-th row 
# containing k error classes which it thinks are most likely to be the correct error class for the i-th test point.
# Class numbers must be returned in ranked order i.e. the label yPred[i][0] must be the best guess for the error class
# for the i-th data point followed by yPred[i][1] and so on.

# CAUTION: Make sure that you return (yPred below) an n x k numpy nd-array and not a numpy/scipy/sparse matrix
# Thus, the returned matrix will always be a dense matrix. The evaluation code may misbehave and give unexpected
# results if an nd-array is not returned. Please be careful that classes are numbered from 1 to 50 and not 0 to 49.

def findErrorClass( X, k ):
	# Find out how many data points we have
	n = X.shape[0]
	filename = "model.sav"
	ld_model = pickle.load(open(filename, 'rb'))
	y_proba = ld_model.predict_proba(X)
	add3col = np.zeros(X.shape[0])
	y_proba = np.insert(y_proba,32,add3col, axis=1)
	y_proba = np.insert(y_proba,35,add3col, axis=1)
	y_proba = np.insert(y_proba,37,add3col, axis=1)
	y_pred = np.zeros((X.shape[0],k))
	for i in range(X.shape[0]):
		top5 = np.argsort(y_proba[i])[-k:][::-1]+1
		y_pred[i] = top5
	return y_pred