get_smoothed_ccc.py

import torch
from models.utils import smooth_predictions, concordance_cc2_np
import numpy as np


if __name__ == '__main__':
    x = torch.load('predictions_val.pt', map_location='cpu')
    gt_v, gt_a, pred_v, pred_a = x['valence_gt'], x['arousal_gt'], x['valence_pred'], x['arousal_pred']
    names = list(gt_v.keys())
    all_v_gt, all_a_gt, all_v_pred, all_a_pred = [], [], [], []

    ccc_v, ccc_a = {}, {}

    for name in names:
        pv = smooth_predictions(pred_v[name], 35, mode='wiener')
        pa = smooth_predictions(pred_a[name], 35, mode='wiener')
        gv = gt_v[name].numpy()
        ga = gt_a[name].numpy()
        valid = (gv >= -1) & (ga >= -1)
        ccc_v[name] = concordance_cc2_np(pv[valid], gv[valid])
        ccc_a[name] = concordance_cc2_np(pa[valid], ga[valid])
        all_v_pred.append(pv[valid])
        all_a_pred.append(pa[valid])
        all_v_gt.append(gv[valid])
        all_a_gt.append(ga[valid])

    print (concordance_cc2_np(np.concatenate(all_v_pred), np.concatenate(all_v_gt)))
    print (concordance_cc2_np(np.concatenate(all_a_pred), np.concatenate(all_a_gt)))

    print ('Lowest ccc-v:')
    for name, v_ in sorted(ccc_v.items(), key=lambda x: x[1])[:10]:
        print (name, v_)
    print ('Highest ccc-v:')
    for name, v_ in sorted(ccc_v.items(), key=lambda x: -x[1])[:10]:
        print (name, v_)
    
    print ('Lowest ccc-a:')
    for name, a_ in sorted(ccc_a.items(), key=lambda x: x[1])[:10]:
        print (name, a_)
    print ('Highest ccc-a:')
    for name, a_ in sorted(ccc_a.items(), key=lambda x: -x[1])[:10]:
        print (name, a_)