Reproduction on the R2R

[Fanzy27]

Hello @MrZihan.
Thank you for providing your work.
I am trying to follow your process. However, I noticed that there are some differences between the hyperparameters specified in the run_r2r.sh script under map_nav_src/scripts and those mentioned in the "Training Details" section of your paper.
After using the hyperparameters from your code, I obtained results that are slightly different from those reported in your paper. I am wondering if I should use the hyperparameters described in your paper instead, or if there are any additional considerations I should be aware of to achieve better alignment with your results.

Difference:
run_r2r.sh:
ngpus: 4 batchsize: 4 pretrain iters: 65k finetune iters: 20k
your paper:
ngpus: 3 batchsize: 4 pretrain iters: 100k finetune iters: 50k

Or my command for finetune is wrong?
CUDA_VISIBLE_DEVICES='0,1,2,3' python3 -m torch.distributed.launch --master_port 29504 --nproc_per_node=${ngpus} main_nav.py $flag
--tokenizer bert
--bert_ckpt_file ../datasets/R2R/exprs_map/pretrain/cmt-vitbase-mlm.mrc.sap-init.lxmert-aug.speaker-new/ckpts/model_step_65000.pt
--eval_first