您好作者！我在复现monster时出现了一些问题：
1、按照config文件train_sceneflow.yaml训练了20w step后，在sceneflow测试集上测试，并没有达到论文里的0.37或开源权重的0.377的结果，我是否忽略了某些步骤？比如batchsize应设置为8而不是12
wandb: Run summary:
wandb: train/1px 0.9339
wandb: train/3px 0.96797
wandb: train/5px 0.97763
wandb: train/epe 0.45731
wandb: train/learning_rate 0.0
wandb: train/loss 7.28777
wandb: val/d1 4.54764
wandb: val/epe 0.39265
wandb:
wandb: 🚀 View run dry-vortex-34 at: https://wandb.ai/13339231179wxb-nanjing-university-of-science-and-technology/sceneflow/runs/bys9qbip
wandb: ⭐️ View project at: https://wandb.ai/13339231179wxb-nanjing-university-of-science-and-technology/sceneflow
wandb: Synced 5 W&B file(s), 28500 media file(s), 0 artifact file(s) and 0 other file(s)
wandb: Find logs at: ./wandb/run-20250426_181132-bys9qbip/logs
2、复现的monster的zero-shot性能与Table5中差异非常大
Validation ETH3D: EPE 8.819202, D1 9.717720
Validation MiddleburyF: EPE 1.2212312916914623, D2 8.390393275767565
Validation KITTI2012: EPE 0.9127147690844291, D3 4.514730703692086, 1.60-FPS (0.624s)
Validation KITTI2015: EPE 1.47173886269331, D3 6.332149741057671, 1.60-FPS (0.624s)
望解答！
您好作者！我在复现monster时出现了一些问题：
1、按照config文件train_sceneflow.yaml训练了20w step后，在sceneflow测试集上测试，并没有达到论文里的0.37或开源权重的0.377的结果，我是否忽略了某些步骤？比如batchsize应设置为8而不是12
2、复现的monster的zero-shot性能与Table5中差异非常大
Validation ETH3D: EPE 8.819202, D1 9.717720
Validation MiddleburyF: EPE 1.2212312916914623, D2 8.390393275767565
Validation KITTI2012: EPE 0.9127147690844291, D3 4.514730703692086, 1.60-FPS (0.624s)
Validation KITTI2015: EPE 1.47173886269331, D3 6.332149741057671, 1.60-FPS (0.624s)
望解答！