The evaluation results for 2D+3D

Hello, thanks for your amazing work. I'm currently attempting to replicate the results, however, despite numerous attempts and parameter adjustments, I haven't been able to achieve the results shown in the paper. The command I've been using like: python scripts/ScanRefer_train.py --use_multiview --use_normal --batch_size 8 --epoch 200 --lr 0.002 --coslr --tag 3dvg-trans+ The results: For the environent, I use the pytorch1.8.1 + cu111. Could you provide some guidance on the best practices or specific parameter configurations that might help me get closer to the results described in the paper?

Thank you in advance for your insights!