Can not reproduce the performance in the paper

[PygmalionChen]

I have the same hardware environment and software version as the author. But I can't reproduce the results in the article by using the default parameters:

python tasks/R2R/main.py --name train_from_scratch
--plateau_sched
--lr 1e-4
--max_episode_len 30

Can you tell me about the specific training parameters ?