python train_nce_distill_model.py \
  --region_feat_path=region_features.hdf5 \
  --phrase_feat_path=phrase_features.hdf5 \
  --glove_path=glove.hdf5

python train_nce_distill_model.py \
  --region_feat_path=region_features.hdf5 \
  --phrase_feat_path=phrase_features.hdf5 \
  --glove_path=glove.hdf5 \
  --phrase_to_label_json=phrase_to_label.json

python eval_model.py \
  --region_feat_path=region_features_test.hdf5 \
  --phrase_feat_path=phrase_features_test.hdf5 \
  --glove_path=glove.hdf5 \
  --restore_path=checkpoint.meta

@InProceedings{Wang_2021_CVPR,
    author    = {Wang, Liwei and Huang, Jing and Li, Yin and Xu, Kun and Yang, Zhengyuan and Yu, Dong},
    title     = {Improving Weakly Supervised Visual Grounding by Contrastive Knowledge Distillation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {14090-14100}
}