def learn(self, trainExamples, validationExamples, loss, lossGradient,
options):
self.weights = util.Counter()
random.seed(42)
initStepSize = options.initStepSize
stepSizeReduction = options.stepSizeReduction
regularization = options.regularization
# You should go over the training data numRounds times.
# Each round, go through all the examples in some random order and update
# the weights with respect to the gradient.
for round in range(0, options.numRounds):
random.shuffle(trainExamples)
numUpdates = 0 # Should be incremented with each example and determines the step size.
# Loop over the training examples and update the weights based on loss and regularization.
# If your code runs slowly, try to explicitly write out the dot products
# in the code here (e.g., "for key,value in counter: counter[key] += ---"
# rather than "counter * other_vector")
for x, y in trainExamples:
numUpdates += 1
stepSize = initStepSize / (numUpdates**stepSizeReduction)
lossTerm = lossGradient(self.featureExtractor(x), y,
self.weights) * stepSize
if regularization != 0:
regTerm = self.weights * (regularization /
len(trainExamples))
self.weights = self.weights - lossTerm - regTerm
else:
self.weights = self.weights - lossTerm
# Compute the objective function.
# Here, we have split the objective function into two components:
# the training loss, and the regularization penalty.
# The objective function is the sum of these two values
trainLoss = 0 # Training loss
regularizationPenalty = 0 # L2 Regularization penalty
for x, y in trainExamples:
trainLoss += loss(self.featureExtractor(x), y, self.weights)
regularizationPenalty += 0.5 * (self.weights * self.weights)
self.objective = trainLoss + regularizationPenalty
# See how well we're doing on our actual goal (error rate).
trainError = util.getClassificationErrorRate(
trainExamples, self.predict, 'train', options.verbose,
self.featureExtractor, self.weights)
validationError = util.getClassificationErrorRate(
validationExamples, self.predict, 'validation',
options.verbose, self.featureExtractor, self.weights)
print "Round %s/%s: objective = %.2f = %.2f + %.2f, train error = %.4f, validation error = %.4f" % (
round + 1, options.numRounds, self.objective, trainLoss,
regularizationPenalty, trainError, validationError)
# Print out feature weights
out = open('weights', 'w')
for f, v in sorted(self.weights.items(), key=lambda x: -x[1]):
print >> out, f + "\t" + str(v)
out.close()
def learn(self, trainExamples, validationExamples, loss, lossGradient, options):
self.weights = util.Counter()
random.seed(42)
initStepSize = options.initStepSize
stepSizeReduction = options.stepSizeReduction
regularization = options.regularization
# You should go over the training data numRounds times.
# Each round, go through all the examples in some random order and update
# the weights with respect to the gradient.
for round in range(0, options.numRounds):
random.shuffle(trainExamples)
numUpdates = 0 # Should be incremented with each example and determines the step size.
# Loop over the training examples and update the weights based on loss and regularization.
# If your code runs slowly, try to explicitly write out the dot products
# in the code here (e.g., "for key,value in counter: counter[key] += ---"
# rather than "counter * other_vector")
for x, y in trainExamples:
numUpdates += 1
stepSize = initStepSize/(numUpdates**stepSizeReduction)
lossTerm = lossGradient(self.featureExtractor(x), y, self.weights)*stepSize
if regularization != 0:
regTerm = self.weights*(regularization/len(trainExamples))
self.weights = self.weights - lossTerm - regTerm
else:
self.weights = self.weights - lossTerm
# Compute the objective function.
# Here, we have split the objective function into two components:
# the training loss, and the regularization penalty.
# The objective function is the sum of these two values
trainLoss = 0 # Training loss
regularizationPenalty = 0 # L2 Regularization penalty
for x, y in trainExamples:
trainLoss += loss(self.featureExtractor(x), y, self.weights)
regularizationPenalty += 0.5*(self.weights*self.weights)
self.objective = trainLoss + regularizationPenalty
# See how well we're doing on our actual goal (error rate).
trainError = util.getClassificationErrorRate(trainExamples, self.predict, 'train', options.verbose, self.featureExtractor, self.weights)
validationError = util.getClassificationErrorRate(validationExamples, self.predict, 'validation', options.verbose, self.featureExtractor, self.weights)
print "Round %s/%s: objective = %.2f = %.2f + %.2f, train error = %.4f, validation error = %.4f" % (round+1, options.numRounds, self.objective, trainLoss, regularizationPenalty, trainError, validationError)
# Print out feature weights
out = open('weights', 'w')
for f, v in sorted(self.weights.items(), key=lambda x: -x[1]):
print >>out, f + "\t" + str(v)
out.close()
def learn(self, trainExamples, validationExamples, loss, lossGradient,
options):
self.weights = util.Counter()
random.seed(42)
# You should go over the training data numRounds times.
# Each round, go through all the examples in some random order and update
# the weights with respect to the gradient.
for r in xrange(0, options.numRounds):
random.shuffle(trainExamples)
numUpdates = 0 # Should be incremented with each example and determines the step size.
# Loop over the training examples and update the weights based on loss and regularization.
# If your code runs slowly, try to explicitly write out the dot products
# in the code here (e.g., "for key,value in counter: counter[key] += ---"
# rather than "counter * other_vector")
for x, y in trainExamples:
numUpdates += 1
"*** YOUR CODE HERE (around 7 lines of code expected) ***"
featureVector = self.featureExtractor(x)
lossGrad = lossGradient(featureVector, y, self.weights)
stepSize = options.initStepSize / (numUpdates**
options.stepSizeReduction)
if (options.regularization > 0):
c = float(options.regularization) / len(trainExamples)
for fKey in self.weights.iterkeys():
if self.weights[fKey] != 0 or lossGrad[fKey] != 0:
self.weights[fKey] -= stepSize * (
lossGrad[fKey] + c * self.weights[fKey])
else:
for fKey, fLoss in lossGrad.iteritems():
self.weights[fKey] -= stepSize * fLoss
# Compute the objective function.
# Here, we have split the objective function into two components:
# the training loss, and the regularization penalty.
# The objective function is the sum of these two values
trainLoss = 0 # Training loss
regularizationPenalty = 0 # L2 Regularization penalty
"*** YOUR CODE HERE (around 5 lines of code expected) ***"
for x, y in trainExamples:
featureVector = self.featureExtractor(x)
trainLoss += loss(featureVector, y, self.weights)
for weight in self.weights.itervalues():
if weight > 0:
regularizationPenalty += weight**2
regularizationPenalty *= 0.5 * options.regularization
self.objective = trainLoss + regularizationPenalty
# See how well we're doing on our actual goal (error rate).
trainError = util.getClassificationErrorRate(
trainExamples, self.predict, 'train', options.verbose,
self.featureExtractor, self.weights)
validationError = util.getClassificationErrorRate(
validationExamples, self.predict, 'validation',
options.verbose, self.featureExtractor, self.weights)
print "Round %s/%s: objective = %.2f = %.2f + %.2f, train error = %.4f, validation error = %.4f" % (
r + 1, options.numRounds, self.objective, trainLoss,
regularizationPenalty, trainError, validationError)
# Print out feature weights
out = open('weights', 'w')
for f, v in sorted(self.weights.items(), key=lambda x: -x[1]):
print >> out, f + "\t" + str(v)
out.close()
def learn(self, trainExamples, validationExamples, trainKickingExamples, validationKickingExamples, loss, lossGradient, options):
self.weights = util.Counter()
random.seed(42)
# You should go over the training data numRounds times.
# Each round, go through all the examples in some random order and update
# the weights with respect to the gradient.
for round in range(0, options.numRounds):
random.shuffle(trainExamples)
numUpdates = 0 # Should be incremented with each example and determines the step size.
trainingSize = len(trainExamples)
# Loop over the training examples and update the weights based on loss and regularization.
# If your code runs slowly, try to explicitly write out the dot products
# in the code here (e.g., "for key,value in counter: counter[key] += ---"
# rather than "counter * other_vector")
for x, y in trainExamples:
numUpdates += 1
stepSize = options.initStepSize / (numUpdates**options.stepSizeReduction)
featx = self.featureExtractor(x)
gradient = lossGradient(featx, y, self.weights)
if (gradient.totalCount() != 0):
updater = (gradient * stepSize)
for key in updater:
self.weights[key] -= updater[key]
if (options.regularization != 0):
for key in gradient:
self.weights[key] -= stepSize * (self.weights[key] * (options.regularization / trainingSize))
# Compute the objective function.
# Here, we have split the objective function into two components:
# the training loss, and the regularization penalty.
# The objective function is the sum of these two values
trainLoss = 0 # Training loss
regularizationPenalty = 0 # L2 Regularization penalty
for x, y in trainExamples:
featx = self.featureExtractor(x)
trainLoss += loss(featx, y, self.weights)
if (options.regularization != 0):
for key in self.weights:
regularizationPenalty += (self.weights[key]**2)
regularizationPenalty *= (options.regularization / 2)
self.objective = trainLoss + regularizationPenalty
# See how well we're doing on our actual goal (error rate).
trainError = util.getClassificationErrorRate(trainExamples, self.predict, trainKickingExamples, 'train', options.verbose, self.featureExtractor, self.weights)
if options.single == 'no':
validationError = util.getClassificationErrorRate(validationExamples, self.predict, validationKickingExamples, 'validation', options.verbose, self.featureExtractor, self.weights)
print "Round %s/%s: objective = %.2f = %.2f + %.2f, train error = %.4f, validation error = %.4f" % (round+1, options.numRounds, self.objective, trainLoss, regularizationPenalty, trainError, validationError)
# Print out feature weights
out = open('weights', 'w')
for f, v in sorted(self.weights.items(), key=lambda x: -x[1]):
print >>out, f + "\t" + str(v)
out.close()
if options.single == 'yes' and len(validationExamples) != 0:
print (self.weights * self.featureExtractor(validationExamples[0][0]))
if self.predict(validationExamples[0][0]) == 1:
print "pass"
else: print "run"
elif options.single == 'yes' and len(validationKickingExamples) != 0:
quarter, time, down, dist, yrdline, play, awyscr, homescr, epb, epa, home, team = validationKickingExamples[0][0].split(",")
prediction = "field goal"
minute = time.split(":")
scorediff = 0
if home == "True":
scorediff = int(homescr) - int(awyscr)
else:
scorediff = int(awyscr) - int(homescr)
if int(minute[0]) <= 3 and (quarter == "4"):
if (scorediff < -3) and (scorediff >= -16):
prediction = "pass"
elif (scorediff >= -3) and scorediff < 0 and (team in yrdline):
prediction = "pass"
else:
if team in yrdline:
prediction = "punt"
else:
yrdinfo = yrdline.split(" ")
first = yrdinfo[0]
if first == "50": prediction = "punt"
else:
if int(yrdinfo[1]) >= 38: prediction = "punt"
print prediction
def learn(self, trainExamples, validationExamples, loss, lossGradient, options):
self.weights = util.Counter()
random.seed(42)
# You should go over the training data numRounds times.
# Each round, go through all the examples in some random order and update
# the weights with respect to the gradient.
for r in xrange(0, options.numRounds):
random.shuffle(trainExamples)
numUpdates = 0 # Should be incremented with each example and determines the step size.
# Loop over the training examples and update the weights based on loss and regularization.
# If your code runs slowly, try to explicitly write out the dot products
# in the code here (e.g., "for key,value in counter: counter[key] += ---"
# rather than "counter * other_vector")
for x, y in trainExamples:
numUpdates += 1
"*** YOUR CODE HERE (around 7 lines of code expected) ***"
featureVector = self.featureExtractor(x)
lossGrad = lossGradient(featureVector, y, self.weights)
stepSize = options.initStepSize / (numUpdates ** options.stepSizeReduction)
if options.regularization > 0:
c = float(options.regularization) / len(trainExamples)
for fKey in self.weights.iterkeys():
if self.weights[fKey] != 0 or lossGrad[fKey] != 0:
self.weights[fKey] -= stepSize * (lossGrad[fKey] + c * self.weights[fKey])
else:
for fKey, fLoss in lossGrad.iteritems():
self.weights[fKey] -= stepSize * fLoss
# Compute the objective function.
# Here, we have split the objective function into two components:
# the training loss, and the regularization penalty.
# The objective function is the sum of these two values
trainLoss = 0 # Training loss
regularizationPenalty = 0 # L2 Regularization penalty
"*** YOUR CODE HERE (around 5 lines of code expected) ***"
for x, y in trainExamples:
featureVector = self.featureExtractor(x)
trainLoss += loss(featureVector, y, self.weights)
for weight in self.weights.itervalues():
if weight > 0:
regularizationPenalty += weight ** 2
regularizationPenalty *= 0.5 * options.regularization
self.objective = trainLoss + regularizationPenalty
# See how well we're doing on our actual goal (error rate).
trainError = util.getClassificationErrorRate(
trainExamples, self.predict, "train", options.verbose, self.featureExtractor, self.weights
)
validationError = util.getClassificationErrorRate(
validationExamples, self.predict, "validation", options.verbose, self.featureExtractor, self.weights
)
print "Round %s/%s: objective = %.2f = %.2f + %.2f, train error = %.4f, validation error = %.4f" % (
r + 1,
options.numRounds,
self.objective,
trainLoss,
regularizationPenalty,
trainError,
validationError,
)
# Print out feature weights
out = open("weights", "w")
for f, v in sorted(self.weights.items(), key=lambda x: -x[1]):
print >> out, f + "\t" + str(v)
out.close()
