search

cxliu0 / PET

[ICCV 2023] Point-Query Quadtree for Crowd Counting, Localization, and More
MIT License
53 stars 5 forks source link

SHB result #9

Closed mru4913 closed 5 months ago

mru4913 commented 6 months ago

Hi，there:

I retrained SH-B with the code and the hyper-parameters given; However the best MAE was 7.6, which is far different than the result given by the paper. Is there any tricks I should conduct in order to replicate your result?

cxliu0 commented 6 months ago

Different environments may result in different optimal epochs. You may try to train the model for 2000 or 3000 epochs (optionally, add torch.backends.cudnn.deterministic = True before training for reproducibility). In general, SHB and SHA share the same training parameters.