search

Julie-tang00 / Point-BERT

[CVPR 2022] Pre-Training 3D Point Cloud Transformers with Masked Point Modeling
MIT License
541 stars 65 forks source link

Some question about few-shot learning in fine-tune stage. #60

Open luoyoucode opened 1 year ago

luoyoucode commented 1 year ago

Hi, thank you for your code first. I got some questions when I code my few-shot learning stage:

In every ModelNet Few-shot Dataset fold, for example, 5way_10shot, there are 50 items = 5 Class * 10 shot in training data

` 2023-03-14 16:53:13,496 - ModelNetFewShot - INFO - Load processed data from data/ModelNetFewshot/5way10shot/0.pkl...

2023-03-14 16:53:13,516 - ModelNetFewShot - INFO - The size of train data is 50

2023-03-14 16:53:13,516 - ModelNetFewShot - INFO - Load processed data from data/ModelNetFewshot/5way10shot/0.pkl...

2023-03-14 16:53:13,528 - ModelNetFewShot - INFO - The size of test data is 100 `

and it's totally training data. I want to know how can I split the query sets and support sets into the 50 items and hold the 5way_10shot task setting. When I sample an item of the query set in training data, I can't find 10 shots data in the corresponding class, because there are 9 shots in training data.

yuxumin commented 1 year ago

Hey, i think it may be a different definition for point cloud few-shot learning task, check here https://proceedings.neurips.cc/paper/2020/file/50c1f44e426560f3f2cdcb3e19e39903-Paper.pdf， see section 3.1