cluster_evaluation.py

import numpy as np
import cv2
from sklearn import metrics
import candidate_selection as cand_selec


# def pairwise_distance(f_name, img_feats, query_feats):
#     if f_name == "BOVW":
#         diq = np.sqrt(np.sum((img_feats - query_feats) ** 2))
#     elif f_name == "MPEG7":
#         diq = np.sqrt(np.sum((img_feats[0] - query_feats[0]) ** 2)) + np.sqrt(
#             np.sum((img_feats[1] - query_feats[1]) ** 2)) + np.sqrt(
#             np.sum((img_feats[2] - query_feats[2]) ** 2))
#     elif f_name == "SIFT":
#         bf = cv2.BFMatcher(cv2.NORM_L2, crossCheck=True)
#         matches = bf.match(img_feats, query_feats)
#         matches = sorted(matches, key=lambda x: x.distance)
#         matches = len(matches)
#         diq = matches
#
#     return diq


def my_v_measure_score(labels_true, labels_pred):
    return metrics.v_measure_score(labels_true, labels_pred)


def my_calinski_harabasz_score(fname, dataset, labels):
    if (fname == "MPEG7" or fname == "SIFT"):
        dataset = np.array(dataset)
        dataset = dataset.reshape(
            dataset.shape[0], (dataset.shape[1]*dataset.shape[2]))
        return metrics.calinski_harabasz_score(dataset, labels)
    else:
        return metrics.calinski_harabasz_score(dataset, labels)


# def my_silhouette_score(fname, dataset, labels):
#     if (fname == "SIFT"):
#         dataset = np.array(dataset)
#         dataset = dataset.reshape(
#             dataset.shape[0], (dataset.shape[1] * dataset.shape[2]))
#         return metrics.silhouette_score(dataset, labels, metric='euclidean')
#     else:
#         return metrics.silhouette_score(dataset, labels, metric='euclidean')


def silhouette_score(f_name, data_points, labels, k, dist_matrix):
    num = len(labels)
    silh_points = []

    for i in range(num):
        clu = labels[i]
        same = 0.0
        diff = 0.0
        for j in range(num):
            if labels[j] == clu:
                same += dist_matrix[i][j]
            else:
                diff += dist_matrix[i][j]

        same_count = labels.count(labels[i])
        a = same / same_count
        b = diff / (num - same_count)
        silh_points.append((b - a) / max(b, a))

    final_score = 0.0
    for i in range(k):
        temp = 0.0
        for j in range(num):
            if i == labels[j]:
                temp += silh_points[j]

        if labels.count(i) == 0:
            temp = 0
        else:
            temp /= labels.count(i)
        final_score += temp

    return final_score / k