The indicators of training results are abnormal

[erasan23]

When I trained directly with the processed Structured3D dataset, the final training result seemed to be wrong. I don't know what I did wrong that led to this.

When training, I used the command: python main.py --dataset_name=stru3d --dataset_root=dataru3d --num_queries=800 --num_polys=20 --semantic_classes=-1 --job_name=train_stru3d --num_workers=0

The relevant indicators obtained after the training are as follows: Averaged stats: room_prec: 0.0000 (0.1604) room_rec: 0.0000 (0.1415) corner_prec: 0.0000 (0.0777) corner_rec: 0.0000 (0.0607) angles_prec: 0.0000 (0.0626) angles_rec: 0.0000 (0.0500) loss: 0.8760 (0.8640) loss_ce: 0.1437 (0.1420) loss_coords: 0.3569 (0.3627) loss_raster: 0.3638 (0.3593) loss_ce_unscaled: 0.0719 (0.0710) loss_coords_unscaled: 0.0714 (0.0725) loss_raster_unscaled: 0.3638 (0.3593) cardinality_error_unscaled: 9.7000 (9.8400)

The charts in wandb are as follows:

My dataset file directory is shown below:

[ywyue]

Hi, could you please try setting --num_workers=2? btw, what's your pytorch version?

[erasan23]

When num_workers>0, this error occurs: EOFError:Ran out of input. In addition, my pytorch version is 1.9.0+cu111.

[erasan23]

Thanks. The problem is solved by setting --num_worker=2. But for Windows, the code in main.py needs to be modified as follows:

First, add the following code:

Then, set the input parameter 'mutiprocessing_context' in the Data_loader:

[ywyue]

It is great the problem was solved! Thanks.