'feature' not in batch.keys() in loss_superpixels_average

sbysbysbys / UOV

Apache License 2.0

15 stars 2 forks source link

'feature' not in batch.keys() in loss_superpixels_average #1

Closed synsin0 closed 3 months ago

synsin0 commented 3 months ago

Hello, I preprocess with SAN and find that the script do not save xxx_lidarseg.npz, features=[] actually, and the collate function does not include 'feature' key to batch dict. How to generate features for superpoints?

sbysbysbys commented 3 months ago

Sorry, I may have accidentally deleted pretrain.dataloader_nuscenes_ov, I've added it back now. Please let me know if you encounter any issues.

synsin0 commented 3 months ago

Thanks! I have another question with task settings, does the model learn open-vocabulary semantics on pretraining stage or annotation-free stage? Is pretrain is only for backbone warmup?

sbysbysbys commented 3 months ago

It learns semantics in both stages, with the first stage focusing on scene understanding and the second stage on scene perception. So pre-training not only warmup parameters but also endows our models with scene understanding and transfer learning capabilities.

synsin0 commented 3 months ago

Great Work! Now I can start the pretrain stage and will reach you if meet further issues. Thanks a lot.