1
1.1
1.2
Assignment4 November 29, 2017
CSE 252A Computer Vision I Fall 2017
Assignment 4
Problem 1: Install Tensorflow 2 pts
Follow the directions on https:www.tensorflow.orginstall to install Tensorflow on your com
puter.
Note: You will not need GPU support for this assignment so dont worry if you dont have one.
Furthermore, installing with GPU support is often more difficult to configure so it is suggested that you install the CPU only version. However, if you have a GPU and would like to install GPU support feel free to do so at your own risk :
Note: On windows, Tensorflow is only supported in python3, so you will need to install python3 for this assignment.
Run the following cell to verify your instalation.
In : import tensorflow as tf
hello tf.constantHello, TensorFlow!
sess tf.Session
printsess.runhello
1.3 Problem 2: Downloading CIFAR10 1 pts
Download the CIFAR10 dataset http:www.cs.toronto.edukrizcifar.html. You will need the python version: http:www.cs.toronto.edukrizcifar10python.tar.gz
Extract the data to .data Once extracted run the following cell to view a few example images.
In 1: import numpy as np
unpickles raw data files
def unpicklefile:
import pickle
import sys
with openfile, rb as fo:
if sys.versioninfo0 3:
dict pickle.loadfo
else:
dict pickle.loadfo, encodingbytes
1
return dict
loads data from a single file
def getBatchfile:
dict unpicklefile
data dictbdata.reshape1,3,32,32.transpose0,2,3,1
labels np.asarraydictblabels, dtypenp.int64
return data,labels
loads all training and testing data
def getDatapath.data:
classes s.decodeUTF8 for s in unpicklepathbatches.metablabelnames
trainData, trainLabels ,
for i in range5:
data, labels getBatchpathdatabatchdi1
trainData.appenddata
trainLabels.appendlabels
trainData np.concatenatetrainData
trainLabels np.concatenatetrainLabels
testData, testLabels getBatchpathtestbatch
return classes, trainData, trainLabels, testData, testLabels
training and testing data that will be used in the following problems
classes, trainData, trainLabels, testData, testLabels getData
display some example images
import matplotlib.pyplot as plt
matplotlib inline
plt.figurefigsize14, 6
for i in range14:
plt.subplot2,7,i1
plt.imshowtrainDatai
plt.titleclassestrainLabelsi
plt.show
print train shape: strtrainData.shape , strtrainLabels.shape
print test shape : strtestData.shape , strtestLabels.shape
2
train shape: 50000, 32, 32, 3, 50000,
test shape : 10000, 32, 32, 3, 10000,
Below are some helper functions that will be used in the following problems.
In : a generator for batches of data
yields data batchsize, 3, 32, 32 and labels batchsize if shuffle, it will load batches in a random order
def DataBatchdata, label, batchsize, shuffleTrue:
n data.shape0
if shuffle:
index np.random.permutationn
else:
index np.arangen
for i in rangeintnp.ceilnbatchsize:
inds indexibatchsize : minn,i1batchsize
yield datainds, labelinds
tests the accuracy of a classifier
def testtestData, testLabels, classifier:
batchsize50
correct0.
for data,label in DataBatchtestData,testLabels,batchsize:
prediction classifierdata
print prediction
correct np.sumpredictionlabel
return correcttestData.shape0100
a sample classifier
given an input it outputs a random class class RandomClassifier:
3
def initself, classes10:
self.classesclasses
def callself, x:
return np.random.randintself.classes, sizex.shape0
randomClassifier RandomClassifier
print Random classifier accuracy: ftesttestData, testLabels, randomClassifier
1.4 Problem 3: Confusion Matirx 5 pts
Here you will implement a test script that computes the confussion matrix for a classifier. The matrix should be nxn where n is the number of classes. Entry Mi,j should contain the number of times an image of class i was classified as class j. M should be normalized such that each row sums to 1.
Hint: see the function test above for reference.
In : def confusiontestData, testLabels, classifier: your code here
return M
def VisualizeConfussionM:
plt.figurefigsize14, 6
plt.imshowM, vmin0, vmax1 plt.xticksnp.arangelenclasses, classes, rotationvertical plt.yticksnp.arangelenclasses, classes
plt.show
M confusiontestData, testLabels, randomClassifier
VisualizeConfussionM
1.5 Problem 4: KNearest Neighbors KNN 5 pts
Here you will implemnet a simple knn classifer. The distance metric is euclidian in pixel space. k refers to the number of neighbors involved in voting on the class.
Hint: you may want to use: sklearn.neighbors.KNeighborsClassifier
In : from sklearn.neighbors import KNeighborsClassifier
class KNNClassifer:
def initself, k3:
k is the number of neighbors involved in voting your code here
def trainself, trainData, trainLabels: your code here
def callself, x:
this method should take a batch of images batchsize, 32, 32, 3 and return a
4
predictions should be int64 values in the range 0,9 corrisponding to the cla
your code here
test your classifier with only the first 100 training examples use this while debuggi note you should get around 1020 accuracy
knnClassiferX KNNClassifer
knnClassiferX.traintrainData:100, trainLabels:100
print KNN classifier accuracy: ftesttestData, testLabels, knnClassiferX
In : test your classifier with all the training examples This may take a while note you should get around 30 accuracy
knnClassifer KNNClassifer
knnClassifer.traintrainData, trainLabels
print KNN classifier accuracy: ftesttestData, testLabels, knnClassifer
display confusion matrix for your KNN classifier with all the training examples
M confusiontestData, testLabels, knnClassifer
VisualizeConfussionM
1.6 Problem 5: Principal Component Analysis PCA KNearest Neighbors KNN 5 pts
Here you will implemnet a simple knn classifer in PCA space. You should implement PCA your self using svd you may not use sklearn.decomposition.PCA or any other package that directly implements PCA transofrmations
Hint: Dont forget to apply the same normalization at test time.
Note: you should get similar accuracy to above, but it should run faster.
In : from sklearn.decomposition import PCA
class PCAKNNClassifer:
def initself, components25, k3: your code here
def trainself, trainData, trainLabels: your code here
def callself, x: your code here
test your classifier with only the first 100 training examples use this while debuggi
pcaknnClassiferX PCAKNNClassifer
pcaknnClassiferX.traintrainData:100, trainLabels:100
print PCAKNN classifier accuracy: ftesttestData, testLabels, pcaknnClassiferX
5
In : test your classifier with all the training examples This may take a few minutes pcaknnClassifer PCAKNNClassifer
pcaknnClassifer.traintrainData, trainLabels
print KNN classifier accuracy: ftesttestData, testLabels, pcaknnClassifer
display the confusion matrix
M confusiontestData, testLabels, pcaknnClassifer
VisualizeConfussionM
1.7 Deep learning
Below is some helper code to train your deep networks
Hint: see https:www.tensorflow.orggetstartedmnistpros or
https:www.tensorflow.orggetstartedmnistbeginners for reference
In : base class for your Tensorflow networks. It implements the training loop train and p You will need to implement the init function to define the networks structures in class TFClassifier:
def initself:
pass
def trainself, trainData, trainLabels, epochs1, batchsize50:
self.prediction tf.argmaxself.y,1
self.crossentropy tf.reducemeantf.nn.sparsesoftmaxcrossentropywithlogi
self.trainstep tf.train.AdamOptimizer1e4.minimizeself.crossentropy
self.correctprediction tf.equalself.prediction, self.y
self.accuracy tf.reducemeantf.castself.correctprediction, tf.float32
self.sess.runtf.globalvariablesinitializer
for epoch in rangeepochs:
for i, data,label in enumerateDataBatchtrainData, trainLabels, batchsize
, acc self.sess.runself.trainstep, self.accuracy, feeddictself if i10099:
print dd d fepoch, epochs, i, acc
print testing epoch:d accuracy: fepoch1, testtestData, testLabels,
def callself, x:
return self.sess.runself.prediction, feeddictself.x: x
helper function to get weight variable
def weightvariableshape:
initial tf.truncatednormalshape, stddev0.01
return tf.Variableinitial
helper function to get bias variable
def biasvariableshape:
6
initial tf.constant0.1, shapeshape
return tf.Variableinitial
example linear classifier
class LinearClassiferTFClassifier:
def initself, classes10:
self.sess tf.Session
self.x tf.placeholdertf.float32, shapeNone,32,32,3 input batch of image
self.y tf.placeholdertf.int64, shapeNone input labels
model variables
self.W weightvariable32323,classes
self.b biasvariableclasses
linear operation
self.y tf.matmultf.reshapeself.x,1,32323,self.W self.b
test the example linear classifier note you should get around 2030 accuracy
linearClassifer LinearClassifer
linearClassifer.traintrainData, trainLabels, epochs20
display confusion matrix
M confusiontestData, testLabels, linearClassifer
VisualizeConfussionM
1.8 Problem 6: Multi Layer Perceptron MLP 5 pts
Here you will implement an MLP. The MLP shoud consist of 3 linear layers matrix multiplcation and bias offset that map to the following feature dimensions:
32x32x3 hidden
hidden hidden
hidden classes
The first two linear layers should be followed with a ReLU nonlinearity. The final layer should
not have a nonlinearity applied as we desire the raw logits output see: the documentation for tf.nn.sparsesoftmaxcrossentropywithlogits used in the training
The final output of the computation graph should be stored in self.y as that will be used in the training.
Hint: see the example linear classifier Note: you should get around 50 accuracy
In : class MLPClassiferTFClassifier:
def initself, classes10, hidden100:
self.sess tf.Session
self.x tf.placeholdertf.float32, shapeNone,32,32,3 input batch of image
self.y tf.placeholdertf.int64, shapeNone input labels
7
your code here
test your MLP classifier note you should get around 50 accuracy
mlpClassifer MLPClassifer
mlpClassifer.traintrainData, trainLabels, epochs20
display confusion matrix
M confusiontestData, testLabels, mlpClassifer
VisualizeConfussionM
1.9 Problem 7: Convolutional Neural Netork CNN 7 pts
Here you will implement a CNN with the following architecture: ReLU Convkernelsize4x4 stride2, outputfeaturesn ReLU Convkernelsize4x4 stride2, outputfeaturesn2 ReLU Convkernelsize4x4 stride2, outputfeaturesn4 Linearoutputfeaturesclasses
In : def conv2dx, W, stride2:
return tf.nn.conv2dx, W, strides1, stride, stride, 1, paddingSAME
class CNNClassiferTFClassifier:
def initself, classes10, n16:
self.sess tf.Session
self.x tf.placeholdertf.float32, shapeNone,32,32,3 input batch of image
self.y tf.placeholdertf.int64, shapeNone input labels your code here
test your CNN classifier note you should get around 65 accuracy
cnnClassifer CNNClassifer
cnnClassifer.traintrainData, trainLabels, epochs20
display confusion matrix
M confusiontestData, testLabels, cnnClassifer
VisualizeConfussionM
1.10 Further reference
To see how state of the art deep networks do on this dataset see: https:github.comtensorflowmodelstreemasterresearchresnet
8