evaluation.py

import numpy as np


def _fast_hist(label_pred, label_true, num_classes):
    mask = (label_true >= 0) & (label_true < num_classes)
    hist = np.bincount(
        num_classes * label_true[mask].astype(int) +
        label_pred[mask], minlength=num_classes ** 2).reshape(num_classes, num_classes)
    return hist


def confusion_matrix(predictions, gts, num_classes):
    hist = np.zeros((num_classes, num_classes))
    for lp, lt in zip(predictions, gts):
        hist += _fast_hist(lp.flatten(), lt.flatten(), num_classes)
        
    return hist


def kappa_score(confusion):
    n_classes = confusion.shape[0]
    sum0 = np.sum(confusion, axis=0)
    sum1 = np.sum(confusion, axis=1)
    expected = np.outer(sum0, sum1) / np.sum(sum0)

    w_mat = np.ones([n_classes, n_classes], dtype=np.int)
    w_mat.flat[:: n_classes + 1] = 0
    
    k = np.sum(w_mat * confusion) / np.sum(w_mat * expected)
    return 1 - k


def evaluate(predictions, gts, num_classes):
    hist = np.zeros((num_classes, num_classes))
    for lp, lt in zip(predictions, gts):
        hist += _fast_hist(lp.flatten(), lt.flatten(), num_classes)
        
    #print(hist)
    # axis 0: gt, axis 1: prediction
       
    acc = np.diag(hist).sum() / hist.sum()
    acc_cls = np.diag(hist) / hist.sum(axis=1)
    acc_cls = np.nanmean(acc_cls)
    iu = np.diag(hist) / (hist.sum(axis=1) + hist.sum(axis=0) - np.diag(hist))
    mean_iu = np.nanmean(iu)
    freq = hist.sum(axis=1) / hist.sum()
    fwavacc = (freq[freq > 0] * iu[freq > 0]).sum()
    kappa = kappa_score(hist)
    return acc, acc_cls, mean_iu, iu, fwavacc, kappa