combine.py

"""Combine testing results of the three models to get final accuracy."""

import argparse
import numpy as np

def main():
    parser = argparse.ArgumentParser(description="combine predictions")
    parser.add_argument('--iframe', type=str, required=True,
                        help='iframe score file.')
    parser.add_argument('--mv', type=str, required=True,
                        help='motion vector score file.')
    parser.add_argument('--res', type=str, required=True,
                        help='residual score file.')

    parser.add_argument('--wi', type=float, default=2.0,
                        help='iframe weight.')
    parser.add_argument('--wm', type=float, default=1.0,
                        help='motion vector weight.')
    parser.add_argument('--wr', type=float, default=1.0,
                        help='residual weight.')

    args = parser.parse_args()

    with np.load(args.iframe) as iframe:
        with np.load(args.mv) as mv:
            with np.load(args.res) as residual:
                n = len(mv['names'])

                i_score = np.array([score[0][0] for score in iframe['scores']])
                mv_score = np.array([score[0][0] for score in mv['scores']])
                res_score = np.array([score[0][0] for score in residual['scores']])

                i_label = np.array([score[1] for score in iframe['scores']])
                mv_label = np.array([score[1] for score in mv['scores']])
                res_label = np.array([score[1] for score in residual['scores']])
                assert np.alltrue(i_label == mv_label) and np.alltrue(i_label == res_label)

                combined_score = i_score * args.wi + mv_score * args.wm + res_score * args.wr

                accuracy = float(sum(np.argmax(combined_score, axis=1) == i_label)) / n
                print('Accuracy: %f (%d).' % (accuracy, n))

if __name__ == '__main__':
    main()