def train(self, trainImages, k):
"""
Function: Train
-------------
Given a set of training images, and a number of centroids to learn,
k, calculate any information you will need in order to make
predictions in the testing phase.
"""
assert not self.trained
# We are going to use the Feature Learner you programmed
# in the first two parts of the assignment.
self.featureLearner = FeatureLearner(k)
# First run your K-means function. After, you will be
# able to use the extractFeatures method that you wrote.
self.featureLearner.runKmeans(trainImages)
# Initialize weight vector
numPatches = (util.image_x / util.patch_dim)**2
self.theta = 1e-2 * np.random.randn(k * numPatches + 1)
### YOUR CODE HERE ###
# Extract features
X = [
np.array(self.featureLearner.extractFeatures(image))
for image in trainImages
]
X = np.vstack(X).transpose() # featdim by num samples
X = np.vstack([np.ones(X.shape[1]), X])
# label array
Y = np.array([image.getLabel() for image in trainImages])
# run gradient descent on learned features
for i in range(self.maxIter):
# logistic probabilities
probs = 1 / (1 + np.exp(-self.theta.dot(X)))
cost = -(np.sum(np.log(probs[Y==1]))+ \
np.sum(np.log(1-probs[Y==0])))
if i % 100 == 0:
print "Logistic Regression: Cost on iteration %d is %f" % (
i, cost)
grad = np.sum((probs - Y) * X, axis=1) # calculates gradient
self.theta = self.theta - self.alpha * grad
### END CODE ###
self.trained = True
def train(self, trainImages, k):
assert not self.trained
# We are going to use the Feature Learner you programmed
# in the first two parts of the assignment.
self.featureLearner = FeatureLearner(k)
# First run your k-means function. After, you will be
# able to use the extractFeatures method that you wrote.
self.featureLearner.runKmeans(trainImages)
### YOUR CODE HERE ###
# Initialize the weight vector from a zero-mean gaussian with 0.01 standard deviation.
numPatches = len(trainImages[0].getPatches())
numParams = k * numPatches
self.theta = np.random.randn(numParams) * 0.01
# Initialize bookkeeping variables
numImages = len(trainImages)
iterations = 0
oldTheta = None
imageFeatures = [self.featureLearner.extractFeatures(trainImages[i]) for i in range(numImages)]
while not self.shouldStop(oldTheta, self.theta, iterations):
oldTheta = self.theta
iterations += 1
self.theta -= self.alpha * sum([(self.logisticFn(imageFeatures[i]) - trainImages[i].getLabel()) * imageFeatures[i] for i in range(numImages)])
self.trained = True
class Classifier(object):
# Constructor
# -----------
# Called when the classifier is first created.
def __init__(self):
# DONT CHANGE THIS USED FOR GRADING
self.trained = False
self.alpha = 1e-5 # learning rate
self.maxIter = 5000 # max num iterations
self.featureLearner = None
self.theta = None # parameter vector for logistic regression
# Function: Train
# -------------
# Given a set of training images, and a number of centroids
# to learn k,
# calculate any information you will need in order to make
# predictions in the testing phase. This function will be
# called only once. Your training must feature select!
def train(self, trainImages, k):
assert not self.trained
# We are going to use the Feature Learner you programmed
# in the first two parts of the assignment.
self.featureLearner = FeatureLearner(k)
# First run your k-means function. After, you will be
# able to use the extractFeatures method that you wrote.
self.featureLearner.runKmeans(trainImages)
### YOUR CODE HERE ###
# Initialize the weight vector from a zero-mean gaussian with 0.01 standard deviation.
numPatches = len(trainImages[0].getPatches())
numParams = k * numPatches
self.theta = np.random.randn(numParams) * 0.01
# Initialize bookkeeping variables
numImages = len(trainImages)
iterations = 0
oldTheta = None
imageFeatures = [self.featureLearner.extractFeatures(trainImages[i]) for i in range(numImages)]
while not self.shouldStop(oldTheta, self.theta, iterations):
oldTheta = self.theta
iterations += 1
self.theta -= self.alpha * sum([(self.logisticFn(imageFeatures[i]) - trainImages[i].getLabel()) * imageFeatures[i] for i in range(numImages)])
self.trained = True
def shouldStop(self, oldTheta, theta, iterations):
return iterations >= self.maxIter
def logisticFn(self, features):
return 1.0/(1.0 + math.exp(-np.inner(self.theta, features)))
# Function: Test
# -------------
# Given a set of testing images
# calculate a list of predictions for those images. You
# may assume that the train function has already been called.
# This function will be called multiple times.
def test(self, testImages):
assert self.trained
# populate this list with best guess for each image
predictions = []
for i in range(len(testImages)):
prediction = np.inner(self.featureLearner.extractFeatures(testImages[i]), self.theta)
if prediction > 0.0:
predictions.append(1)
else:
if prediction < 0.0:
predictions.append(0)
else:
predictions.append(random.randrange(0, 2, 1))
### YOUR CODE HERE ###
return predictions
class Classifier(object):
def __init__(self):
"""
Constructor
-----------
Called when the classifier is first created.
DONT CHANGE THIS USED FOR GRADING
"""
self.trained = False
self.alpha = 1e-5
self.maxIter = 5000
self.featureLearner = None
self.theta = None
def train(self, trainImages, k):
"""
Function: Train
-------------
Given a set of training images, and a number of centroids to learn,
k, calculate any information you will need in order to make
predictions in the testing phase.
"""
assert not self.trained
# We are going to use the Feature Learner you programmed
# in the first two parts of the assignment.
self.featureLearner = FeatureLearner(k)
# First run your K-means function. After, you will be
# able to use the extractFeatures method that you wrote.
self.featureLearner.runKmeans(trainImages)
# Initialize weight vector
numPatches = (util.image_x / util.patch_dim)**2
self.theta = 1e-2 * np.random.randn(k * numPatches + 1)
### YOUR CODE HERE ###
# Extract features
X = [
np.array(self.featureLearner.extractFeatures(image))
for image in trainImages
]
X = np.vstack(X).transpose() # featdim by num samples
X = np.vstack([np.ones(X.shape[1]), X])
# label array
Y = np.array([image.getLabel() for image in trainImages])
# run gradient descent on learned features
for i in range(self.maxIter):
# logistic probabilities
probs = 1 / (1 + np.exp(-self.theta.dot(X)))
cost = -(np.sum(np.log(probs[Y==1]))+ \
np.sum(np.log(1-probs[Y==0])))
if i % 100 == 0:
print "Logistic Regression: Cost on iteration %d is %f" % (
i, cost)
grad = np.sum((probs - Y) * X, axis=1) # calculates gradient
self.theta = self.theta - self.alpha * grad
### END CODE ###
self.trained = True
def test(self, testImages):
"""
Function: Test
-------------
Given a set of testing images make a prediction for each
image. You may assume that the train function has already been
called.
"""
assert self.trained
# populate this list with best guess for each image
predictions = []
### YOUR CODE HERE ###
# Extract features
X = [
self.featureLearner.extractFeatures(image) for image in testImages
]
X = np.vstack(X).transpose() # featdim by num samples
X = np.vstack([np.ones(X.shape[1]), X])
# probabilities
probs = 1 / (1 + np.exp(-self.theta.dot(X)))
predictions = np.zeros(probs.shape)
predictions[probs > 0.5] = 1
predictions = predictions.tolist()
### END CODE ###
return predictions
