训练过程问题 about bp-net

[kakaxi314]

We can see from your command
torchrun --nproc_per_node=4 --master_port 4321 train.py gpus=[0] ...
that you run with 4 nodes, but only provide 1 GPU.
You should give 4 GPU ids (e.g. gpus=[0,1,2,3]) for 4 nodes.
If you only have 1 GPU, you can directly use python train.py gpus=[0] ... to run the code.

[stw2021]

Thank you very much for your reply! I have solved this problem.
But when I test with the checkpoints provided, the result is different from the paper:

[kakaxi314]

Please show me your command.

[stw2021]

command is HYDRA_FULL_ERROR=1 python test.py gpus=[0] name=BP_KITTI ++chpt=BP_KITTI net=PMP num_workers=4 data=KITTI data.testset.mode=test data.testset.height=352 test_batch_size=1 metric=RMSE ++save=true

[kakaxi314]

You evaluated the test set.
As we couldn't get the ground truth of the test set of KITTI dataset, I simply used the input sparse depth map as the ground truth.

Thus, the displayed RMSE result for test set is meaningless. You need to submit the saved results to KITTI server for fair evaluation.

[stw2021]

Got it! Thank you very much for your patient reply.