Parameters of the paper's model

[gkrisp98]

Hi, I use the setup described in the script reproduce_paper_resuls.sh to reproduce paper's results, but the model I get has the following number of parameters:

| Name | Type | Params

0 | layers | ModuleList | 81.6 K
1 | output_layer | BRepNetFaceOutputLayer | 98.9 K
2 | classification_layer | Linear | 680

181 K Trainable params
0 Non-trainable params
181 K Total params
0.725 Total estimated model params size (MB)

And the following results:
DATALOADER:0 TEST RESULTS
{'test/Chamfer_iou': 0.8299525380134583,
'test/CutEnd_iou': 0.7150671482086182,
'test/CutSide_iou': 0.7787927389144897,
'test/ExtrudeEnd_iou': 0.8739719986915588,
'test/ExtrudeSide_iou': 0.9213229417800903,
'test/Fillet_iou': 0.978989839553833,
'test/RevolveEnd_iou': 0.5921052694320679,
'test/RevolveSide_iou': 0.7775700688362122,
'test/accuracy': 0.9383547306060791,
'test/mean_iou': 0.8084715604782104}

This is inconsistent with the results published in the paper, where the best performing model has 359k parameters and accuracy 92.52 ± 0.15 and IoU 77.10 ± 0.54. Can you tell me what I do wrong ?

[wll0307]

@gkrisp98
你好，我想要了解一下就是如何构建自己的数据还有就是如何用自己的数据来获取的seg文件的，这里面的好多都是要用到seg，我想构建我的数据就是不知道是如何获得的seg文件的