Implementation_on_visdial_bert

1. It is really hard to eliminate the direct path from history to answer, so in the current stage, we suggest to remove history.
You can change the dataloader files to do this.
2. For implementation of P2, we recommend the R4 Loss which is found after the paper publication. Meanwhile, we suggest enlarge the learning rate.
The experience is 3~5 epochs for stage 2. 200k iterations for stage 1.
3. We also suggest to decrease the loss of object prediction and LM.
4. We found that the ensemble of vd-bert result and our results can bring a further improvement though the single performance of vd-bert is about 75.5% and the ensemble of ours is lower than 75%
5. About the fusion of NDCG and MRR, only use the first place of MRR result and others use the NDCG result is a good strategy, which we omitted in the challenge, of course other fusion strategy using first several result of MRR can make some trade-off. This fusion strategy will not hurt NDCG a lot. In the leaderboard, when you find the MRR and R@10 is different from the experience, the author may apply this strategy.