search

tusen-ai / SST

Code for a series of work in LiDAR perception, including SST (CVPR 22), FSD (NeurIPS 22), FSD++ (TPAMI 23), FSDv2, and CTRL (ICCV 23, oral).
Apache License 2.0
765 stars 96 forks source link

about the training loss? #155

Closed rockywind closed 6 months ago

rockywind commented 10 months ago

Hi, thank you for sharing the great work. When I load the pretrained model, the training loss will get bigger and bigger. The config is the same as the fsdv2_argo_2x.py. Can you give me some advices? image

Abyssaledge commented 10 months ago

I do not really understand your setting. Please be more specific. Why you need the pretrained model for training?

rockywind commented 6 months ago

Hi, the pre-trained model has been developed using one-fifth of the avaliable datasets.