search

XuyangBai / TransFusion

[PyTorch] Official implementation of CVPR2022 paper "TransFusion: Robust LiDAR-Camera Fusion for 3D Object Detection with Transformers". https://arxiv.org/abs/2203.11496
Apache License 2.0
619 stars 76 forks source link

fade strategy #45

Open wangyd-0312 opened 2 years ago

wangyd-0312 commented 2 years ago

hi, Mr. Bai. I got a very low performance while reproducing Transfusion-L. mAP:0.2484, NDS:0.2350. I only have a 8G NVIDIA 2080. And I trained on nuscenes-mini. I have searched related issues and found that I forgot the fade strategy. But I do not know how to achieve the fade strategy. Which parts of the code should I change? Thank you very much!! This is my log: 20220613_123356.log

XuyangBai commented 2 years ago

Hi @wangyd-0312 you should comment out the ObjectSampler in config file and resume the training from 15 epoch. https://github.com/XuyangBai/TransFusion/blob/399bda09a3b6449313ccc302df40651f77ec78bf/configs/transfusion_nusc_voxel_L.py#L30-L66

victor000000 commented 2 years ago

hi, Mr. Bai. I got a very low performance while reproducing Transfusion-L. mAP:0.2484, NDS:0.2350. I only have a 8G NVIDIA 2080. And I trained on nuscenes-mini. I have searched related issues and found that I forgot the fade strategy. But I do not know how to achieve the fade strategy. Which parts of the code should I change? Thank you very much!! This is my log: 20220613_123356.log

the learning rate is too large