luohongyin / ford_vqa

ford_vqa

Training

Run ./setup_training.sh and it will download and pretrained word embeddings.

Run python singleton.py EXP_NAME for training. EXP_NAME should be defined in experiments.conf.

Example: python singleton.py best.

Evaluating

Run ./copy_exp.sh EXP_NAME ITER_NUM.

For example: ./copy_exp.sh best 10000.

Run python test_single.py EXP_NAME. This generates scores.pkg file.

For example: python test_single.py best.

To get final scores, run python eval.py 0 (or 1). 0 for evaluating on video clips and 1 for entire video.

Format of scores.pkg

Test scores can be loaded by _, _, test_scores = pickle.load(open('scores.pkg')).

test_score = [['q_id', 'v_id', score], ....]