Spconv 2.2.x

[L-Reichardt]

This pull request is just so this fork shows up linked to original repository, and does not need to be merged.

I have retained with Spconv 2.1.21, with slightly lower results. The purpose of this pull request is just to offer weights/code with the updated Spconv library.

[xizaoqu]

Counld you provide the setting or log file of the training results of mIOU 63.5? Thanks!

[L-Reichardt]

@xizaoqu I've updated the ReadMe in my fork.

[xizaoqu]

@xizaoqu I've updated the ReadMe in my fork.

The LR is 10 times larger than the official setting. Do you use multiple GPUs for training?
Besides, do you suffer from overfitting? I use 4GPUs with the batch size of 4 and get the best mIOU of ~63 at about the 10th epoch. But afterward, the performance will jitter dramatically, sometimes even below 59.

[L-Reichardt]

@xizaoqu Thank you, it was a typo. I've used the settings in the config file (LR of 0.001). I used a single Nvidia 3060 for training.

I had the same "issue". The model reached peak performance at around 10-15 epochs. Most likely the model can be further improved with an ablation study on the training hyperparameters and heavier augmentation. I wont train it anymore for some time though as I am working on another project. Maybe in the future, I will retrain the Knowledge-Distilled version to the new SPConv Version

[xizaoqu]

@xizaoqu Thank you, it was a typo. I've used the settings in the config file (LR of 0.001). I used a single Nvidia 3060 for training.

I had the same "issue". The model reached peak performance at around 10-15 epochs. Most likely the model can be further improved with an ablation study on the training hyperparameters and heavier augmentation. I wont train it anymore for some time though as I am working on another project. Maybe in the future, I will retrain the Knowledge-Distilled version to the new SPConv Version

I see. Thanks!