colorize.py 文件源码

def single_metrics(gt, pred, num_cl):
    t_px = np.zeros(num_cl)
    n_px = np.zeros(num_cl)
    n1_px = np.zeros(num_cl)
    px_class = np.unique(gt) 
    error = np.subtract(gt, pred)

    for i in px_class:
        t_px[i] += (np.where(gt == i)[0]).shape[0]
        n_px[i] += (np.where((gt == i) & (error == 0))[0]).shape[0]
        n1_px[i] += (np.where(pred == i)[0]).shape[0]

    return t_px, n_px, n1_px


# if __name__ == "__main__":
#   pic_start = int(sys.argv[1])
#   pic_end = int(sys.argv[2])
#   num_cl = 22
#   t_px = np.zeros(num_cl)
#   n_px = np.zeros(num_cl)
#   n1_px = np.zeros(num_cl)

#   for idx in range(pic_start, pic_end+1):
#       tmp_t, tmp_n, tmp_n1 = save_result(str(idx), num_cl, True)
#       t_px += tmp_t
#       n_px += tmp_n
#       n1_px += tmp_n1

#   t_sum = np.sum(t_px)
#   n_sum = np.sum(n_px)
#   px_acc = n_sum/t_sum
#   condition_1 = t_px != 0
#   c_n1 = np.extract(condition_1, n_px)
#   c_t1 = np.extract(condition_1, t_px)
#   condition_2 = (np.subtract(np.add(t_px, n1_px), n_px)) != 0
#   c_n2 = np.extract(condition_2, n_px)
#   c_d2 = np.extract(condition_2, (np.subtract(np.add(t_px, n1_px), n_px)))
#   mean_acc = np.sum(np.divide(c_n1, c_t1))/num_cl
#   mean_IU = np.sum(np.divide(c_n2, c_d2))/num_cl
#   fw_IU = np.sum(np.divide(np.extract(condition_2, np.multiply(t_px, n_px)), c_d2))/t_sum

#   print("========= metrics =========")
#   print("pixel accuracy: " + str(px_acc))
#   print("mean accuracy: " + str(mean_acc))
#   print("mean IU: " + str(mean_IU))
#   print("frequency weighted IU: " + str(fw_IU))
#   print("")

# if __name__ == "__main__":
#   num_cl = 22
#   pic_id = int(sys.argv[1])
#   save_compare_results(pic_id, num_cl)