Wansit Hepburn

好的 十分感谢

> 没有人复现出论文中说的63%的MIOU吗？ 老兄是用啥配置跑的 现在有复现成功吗？求教

> weight‘s shape should be (19,),not (1,19) in train_SemantciKITTI.py，line 94 > > ``` > class_weights = torch.from_numpy(train_dataset.get_class_weight()).float().cuda() > class_weights=class_weights.squeeze(0) > self.criterion = nn.CrossEntropyLoss(weight=class_weights, reduction='none') > ``` class_weights = torch.from_numpy(train_dataset.get_class_weight()).float().view(19).cuda() #...

> 这是我跑的整体流程，你可以参考一下：https://blog.csdn.net/m0_62648611/article/details/135607266 请问复现结果和文中差距大吗？

> Had the same issue, solved it by changing all the occurences of #include to #include . The reason is that the namespace THC is deprecated and functions were moved...

> Sorry for late reply. I suggest you customize a new `Evaluator` instead of `SegMetric`, in which you can define your own operation, such as remapping the predicted label to...

And when I update this files, I meet some problems File "/opt/data/private/whu_wsy/seg/FRNet-master-init/frnet/models/segmentors/frnet.py", line 83, in extract_feat voxel_dict = self.voxel_encoder(voxel_dict) File "/opt/conda/envs/frnet/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl result = self.forward(*input, **kwargs) File...

> And when I update this files, I meet some problems > > File "/opt/data/private/whu_wsy/seg/FRNet-master-init/frnet/models/segmentors/frnet.py", line 83, in extract_feat voxel_dict = self.voxel_encoder(voxel_dict) File "/opt/conda/envs/frnet/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl result =...

Thank you for your prompt response. After using the new config, I noticed a significant decrease in training speed, almost 1.5 times slower. Even after disabling mixed precision through onecycle-50k.py,...

> Thanks for your continuous attention to our work. In the updated version, we tried to use `iter-based` training instead of `epoch-based` training and use `mixed precision` (successful in `SemanticKITTI`...