def load_data(self):
# load train data
self.train_labels, train_images = dataset.read()
self.train_data_length = len(self.train_labels)
self.number_of_dimensions = train_images[0].flatten().shape[0]
#load test data
self.test_labels, test_images = dataset.read(dataset='testing')
self.test_data_length = len(self.test_labels)
"""Much Better accuracy without normalization"""
#train_images = train_images / 255
#test_images = test_images / 255
self.train_images_flattened = train_images.reshape(
train_images.shape[0], -1).T
self.test_images_flattened = test_images.reshape(
test_images.shape[0], -1).T
train_mean = np.nanmean(self.train_images_flattened)
self.train_images_flattened = self.train_images_flattened - train_mean
self.test_images_flattened = self.test_images_flattened - train_mean
self.weight_vector = np.random.random(
(self.number_of_dimensions, self.number_of_classes)) * 0.000008
self.one_hot_encoded_labels = self.create_one_hot_representation(
self.train_labels)
self.test_one_hot_encoded_labels = self.create_one_hot_representation(
self.test_labels)
def load_train_test(self):
Y_train, X_train = ld.read()
self.Y_test, self.X_test = ld.read('testing')
X_train = np.reshape(X_train,(self.training_size,28*28))/255
self.X_test = np.reshape(self.X_test,(self.testing_size,28*28))/255
Y_train = self.one_hot_representation(Y_train)
#Y_test = self.one_hot_representation(Y_test)
return (X_train, Y_train, self.X_test, self.Y_test)
def driver():
Y_train,X_train = read()
X_train = X_train.reshape((X_train.shape[0],784))
X_train = X_train/float(255)
#W = npy.zeros((784,10))
W = npy.zeros((784,10))
update_weights(X_train,Y_train,W)
def load_data(self):
self.train_labels, train_images = dataset.read()
self.train_data_length = len(self.train_labels)
self.test_labels, test_images = dataset.read(dataset='testing')
self.test_data_length = len(self.test_labels)
"""normalize data set"""
train_images = train_images / 255
test_images = test_images / 255
self.train_images_flattened = train_images.reshape(
train_images.shape[0], -1)
self.test_images_flattened = test_images.reshape(
test_images.shape[0], -1)
#train_std = np.nanstd(self.train_images_flattened, axis=0)
train_mean = np.nanmean(self.train_images_flattened)
self.train_images_flattened = self.train_images_flattened - train_mean
self.test_images_flattened = self.test_images_flattened - train_mean
self.train_images_squared = np.einsum('ij,ij->i',
self.train_images_flattened,
self.train_images_flattened)
def main():
train_labels, train_images = load_dataset.read(
"training",
"/Users/samarth/Desktop/Fall 2018/CSE 575/Assignment 2/MNIST")
train_images = np.reshape(train_images, (60000, 784)) / 255.0
# train_images = getPCAData(train_images,50)
# train_images = np.reshape(train_images, (60000, 784))
print(len(train_images))
test_labels, test_images = load_dataset.read(
"testing",
"/Users/samarth/Desktop/Fall 2018/CSE 575/Assignment 2/MNIST")
test_images = np.reshape(test_images, (10000, 784)) / 255.0
k_list = [1, 3, 5, 10, 30, 50, 70, 80, 90, 100]
i = 0
final = []
for test_image in test_images:
dist = compute_distance_matrix(100, train_images, train_labels,
test_image)
# print(dist)
# print(pred)
final.append(dist)
print('Calculating the matrix for k = 100 for test image[' + str(i) +
']')
i += 1
with open('sorted_list_latest', 'wb') as fp:
pickle.dump(final, fp)
accuracy, accuracy_list = test_accuracy(k_list, test_images, test_labels)
plt.plot(k_list, accuracy_list, color='g')
plt.xlabel("K")
plt.ylabel("Accuracy")
plt.title("KNN Graph")
plt.show()
def driver():
Y_train,X_train = read()
X_train = X_train.reshape((60000,784))
X_train = X_train/float(255)
Y_test,X_test = read("testing")
X_test = X_test.reshape((10000,784))
X_test = X_test/float(255)
Y_pred_k_based = dict()
accuracies = dict()
for k in [1, 3, 5, 10, 30, 50, 70, 80, 90, 100]:
Y_pred_k_based[k] = list()
labels = compute_euclidean_distance(X_train,Y_train,X_test)
for i in range(len(X_test)):
all_l = labels[i][1]
for k in [1, 3, 5, 10, 30, 50, 70, 80, 90, 100]:
label = predict(all_l[0:k])
lst = Y_pred_k_based[k]
lst.append(label)
for k in [1, 3, 5, 10, 30, 50, 70, 80, 90, 100]:
lst_labels = Y_pred_k_based[k]
count = 0
for i in range(0,len(X_test)):
if lst_labels[i] == Y_test[i]:
count = count+1
accuracies[k] = count/float(len(X_test))
plot_graph(accuracies)
dum_accuracies_file(accuracies)
def loadData(data):
#
#Read data
Label, Img1 = read(data)
#
#normalise data
Img1 = Img1 / 255
#
#flattendata
Img = np.asarray([img.flatten() for img in Img1])
#create a frame
Frame = pd.DataFrame(Img)
Frame['label'] = Label
return Frame
def predict(W):
Y_test,X_test = read("testing")
X_test = X_test.reshape((X_test.shape[0],784))
X_test = X_test/float(255)
Y_pred = []
prod = npy.dot(X_test,W)
values = softmax(prod)
Y_pred = npy.argmax(values,axis=1)
count = 0
for i in range(0,X_test.shape[0]):
if(Y_test[i] == Y_pred[i]):
count=count+1
return count/float(X_test.shape[0])
# coding: utf-8
# In[3]:
import numpy as np
import load_dataset
import timeit
import collections
import timeit
from bisect import bisect
import scipy.spatial.distance as sd
# In[4]:
y_train , x_train= load_dataset.read("training","MNIST")
y_test, x_test= load_dataset.read("testing","MNIST")
x_train = x_train.reshape([60000,28*28])
x_test = x_test.reshape([10000,28*28])
# In[ ]:
distanceMatrix = sd.cdist(x_test,x_train)
# In[ ]:
k = [1,3,5,10,30,50,70,80,90,100]
import numpy as np
#from scipy.special import expit
from load_dataset import read, show
[train_lab, train_img] = read()
[test_lab, test_img] = read("testing")
#reshape to nx784 matrix
train_img = train_img.reshape(
train_img.shape[0], (train_img.shape[1] * train_img.shape[2])).astype(int)
test_img = test_img.reshape(
test_img.shape[0], (test_img.shape[1] * test_img.shape[2])).astype(int)
#for test
#train_tl=train_lab[:10000]
#train_ti=train_img[:10000]
#test_tl=test_lab[:5000]
#test_ti=test_img[:5000]
#print(train_tl.shape,train_ti.shape,test_tl.shape,test_ti.shape)
#in logistic regression, all we need is find this w (ignore b)
w = np.zeros((10, 784), dtype=float)
#final predict function
def sigmoid(w, X):
return (1 / (1 + np.exp(-np.matmul(X, w)))).astype(float)
#find w
#first step: relabel to 0,1. such that for 5, y[4][i] is a 0-1 label array.
def __init__(self, image, label):
self.image = image
self.label = label
class BallTreeNode:
def __init__(self, image, radius, leafs):
self.imageData = image
self.radius = radius
self.left = None
self.right = None
self.leafs = leafs
path = "../assignment2_data/"
trainingData = load_dataset.read("training", path)
testData = load_dataset.read("testing", path)
size = 200
trainLbls = trainingData[0][:size * 6]
trainImgs = trainingData[1][:size * 6]
testLbls = testData[0][:size]
testImgs = testData[1][:size]
ks = [1, 3, 5, 10, 30, 50, 70, 80, 90, 100]
trainingData = []
testData = []
for i in range(len(trainLbls)):
image = [[0] * 28 for _ in range(28)]
from __future__ import division
import scipy
import numpy as np
import load_dataset as ld
import matplotlib.pyplot as plt
train_data = ld.read()
alpha = 0.001
y_train_data = train_data[0]
x_train_data = train_data[1].reshape([60000, 28 * 28]) / 255
test_data = ld.read('testing')
y_test_data = test_data[0]
x_test_data = test_data[1].reshape([10000, 28 * 28]) / 255
weights = np.tile(np.zeros(784), (9, 1))
def post_prob(x, y):
if y == 9:
return 1 / sum_exps(x)
else:
return np.exp(np.dot(weights[y], x)) / sum_exps(x)
def sum_exps(x):
exp_array = []
for weight in weights:
exp_array.append(np.exp(np.dot(weight, x)))
return (1 + np.sum(exp_array))
import load_dataset
import time
import numpy as np
startTime = time.time()
path = "../assignment2_data/"
trainings = load_dataset.read("training", path)
tests = load_dataset.read("testing", path)
class ImageData:
def __init__(self, label, image):
self.label = label
self.image = image
self.image = image
testSize = 10000
trainSize = 60000
learningRate = 0.00002 / trainSize
trainingData = [
ImageData(label,
image.flatten().astype(float))
for label, image in zip(trainings[0][:trainSize], trainings[1][:trainSize])
]
testData = [
ImageData(label,
image.flatten().astype(float))
for label, image in zip(tests[0][:testSize], tests[1][:testSize])
def load_train_test(self):
Y_train, X_train = ld.read('training')
Y_test, X_test = ld.read('testing')
X_train = np.reshape(X_train, (training_size, 28 * 28))
X_test = np.reshape(X_test, (testing_size, 28 * 28))
return (X_train, Y_train, X_test, Y_test)
