search

metadriverse / cat

[CoRL'23] Adversarial Training for Safe End-to-End Driving
https://metadriverse.github.io/cat/
MIT License
49 stars 3 forks source link

How much GPU Ram is needed? and does the code work with the full data from validation/testing_interactive? #5

Open CesarLiu opened 8 months ago

CesarLiu commented 8 months ago

Dear Authors,

thanks in advance! As mentioned in the title, I wonder how much GPU RAM does it need to train? I have a GPU with 11GB RAM, but it always stop with CUDA out of memory?

I used raw_scenes_500 provided, it works fine. But when generating my own data from waymo scenarios validation/testing_interactive, there is always KeyError: 'DRIVEWAY' or 'polyline'. Do you have any idea

BenStoler commented 6 months ago

I also ran into issues with GPU VRAM, which was solved by just reducing the batch size from the default 256 to something like 128. Were you able to resolve the KeyError issue? I'm also interested in potentially adding in new scenarios.

CesarLiu commented 6 months ago

I used a 3090, which solved the memory issue. For the key error, I haven't tried until now.