average_fusion.py

import dataloader

from utils.utils import *

if __name__ == '__main__':

    rgb_preds = 'record/spatial/spatial_video_preds.pickle'
    opf_preds = 'record/motion/motion_video_preds.pickle'

    with open(rgb_preds, 'rb') as f:
        rgb = pickle.load(f)
    f.close()
    with open(opf_preds, 'rb') as f:
        opf = pickle.load(f)
    f.close()

    dataloader = dataloader.spatial_dataloader(BATCH_SIZE=1, num_workers=1,
                                               path='/home/ubuntu/data/UCF101/spatial_no_sampled/',
                                               ucf_list='/home/ubuntu/cvlab/pytorch/ucf101_two_stream/github/UCF_data_references/',
                                               ucf_split='01')
    train_loader, val_loader, test_video = dataloader.run()

    video_level_preds = np.zeros((len(rgb.keys()), 101))
    video_level_labels = np.zeros(len(rgb.keys()))
    correct = 0
    ii = 0
    for name in sorted(rgb.keys()):
        r = rgb[name]
        o = opf[name]

        label = int(test_video[name]) - 1

        video_level_preds[ii, :] = (r + o)
        video_level_labels[ii] = label
        ii += 1
        if np.argmax(r + o) == (label):
            correct += 1

    video_level_labels = torch.from_numpy(video_level_labels).long()
    video_level_preds = torch.from_numpy(video_level_preds).float()

    top1, top5 = accuracy(video_level_preds, video_level_labels, topk=(1, 5))

    print top1, top5