image classifier using KNN algorithm and cifar 10 dataset

image_classifier_using_knn.py

# image classifier using KNN and cifar 10 dataset

import numpy as np
import pickle

# sort dictionary by key
def sort_key(dictionary):
    keys = list(dictionary.keys())
    keys.sort()

    sorted_dictionary = {}
    for key in keys:
        sorted_dictionary[key] = dictionary[key]

    return sorted_dictionary

# get batches from files
def get_batches(files, prefix=''):
    # multiple files
    if type(files) is list:
        result = {}

        for file in files:
            data = {}

            # unpickle batches
            with open(prefix + file, 'rb') as fo:
                data = pickle.load(fo, encoding='bytes')

            for i in range(len(data[b'labels'])):
                label = data[b'labels'][i]

                if label not in result:
                    result[label] = []

                result[label].append(data[b'data'][i])

        return sort_key(result)

    # a file
    elif type(files) is str:
        result = {}
        data = {}

        with open(prefix + files, 'rb') as fo:
            data = pickle.load(fo, encoding='bytes')

        for i in range(len(data[b'labels'])):
            label = data[b'labels'][i]

            if label not in result:
                result[label] = []

            result[label].append(data[b'data'][i])

        return sort_key(result)

# Manhattan distance
def mdistance(x, y):
    return np.abs(np.sum(x - y))

# Euclidean distance
def edistance(x, y):
    return np.sqrt(np.abs(np.sum((x-y) ** 2)))

# function that returns majority of a list
# Example : [1,2,1,1] -> 1
def get_majority(data):
    max_item = None
    max_value = None
    for item in set(data):
        tmp = 0
        for datum in data:
            if item is datum:
                tmp += 1

        if max_value is None or max_value < tmp:
            max_value = tmp
            max_item = item

    return max_item

# predict function
def predict(data, test, distance=mdistance):
    min_label = []
    min_label_value = []

    for index in data:
        for batch in data[index]:
            d = distance(batch, test)

            # always append till length of min_label does not equal K.
            if len(min_label) < K:
                min_label.append(index)
                min_label_value.append(d)

            elif max(min_label_value) > d:
                i = np.argmax(min_label_value)
                min_label[i] = index
                min_label_value[i] = d

    return get_majority(min_label)

# Hyperparameters
K = 5
D = edistance 

if __name__ == "__main__":
    # get batches == train
    files = ['1', '2', '3', '4', '5']
    images = get_batches(files, prefix='cifar-10-batches-py/data_batch_')

    # get test data
    test_images = get_batches('test_batch', prefix='cifar-10-batches-py/')

    result = []
    for test_image_index in test_images:
        cnt = 0

        for batch in test_images[test_image_index]:
            label = predict(images, batch, distance=D)
            result.append(label is test_image_index)
            print("predict : %d, answer : %d"%(label, test_image_index))

            cnt += 1
            
            # check 100 images
            # 10 (0 ~ 9) * 10

            # edit this value to adjust the number of test images
            if cnt == 10:
                break

    # print result
    result_np = np.array(result, dtype='float32')
    print("Average : %f"%np.mean(result_np))

# result 
# the number of test images : 100
# 
# K = 3
# using Manhattan distance function
# ---
# Average : 0.250000
# ---
# 
# using Euclidean distance function
# ---
# Average : 0.220000
# ---
# 
# K = 5
# using Manhattan distance function
# ---
# Average : 0.250000
# ---
# 
# using Euclidean distance function
# ---
# Average : 0.200000
# ---