search

bradyz / cross_view_transformers

Cross-view Transformers for real-time Map-view Semantic Segmentation (CVPR 2022 Oral)
MIT License
525 stars 80 forks source link

Questions concerning GPU memory #21

Open Solacex opened 2 years ago

Solacex commented 2 years ago

Hello,

Thanks for your inspiring work, and I am trying to reproduce the reported results with the provided code.

I run the code as you suggested : python3 scripts/train.py \ +experiment=cvt_nuscenes_vehicle data.dataset_dir=/media/datasets/nuscenes \ data.labels_dir=/media/datasets/cvt_labels_nuscenes

I made a slight change on the batch size, i.e., training on two A100 (CUDA11) with 8 samples on each card, but the results exhibit a large gap with the results in the paper. image

Are there any suggestions for debugging this?

Solacex commented 2 years ago
image

It seems better now? which is the reported metric in the paper, IoU@0.4 or IoU@0.5?

bradyz commented 2 years ago

Metrics reported in paper were iou@0.5, but we added the iou@0.4 metrics because we noticed that using the threshold of 0.5 was actually suboptimal.