search

marmotlab / SCRIMP

offical code of paper 'SCRIMP: Scalable Communication for Reinforcement- and Imitation-Learning-Based Multi-Agent Pathfinding'
MIT License
37 stars 14 forks source link

cuda is not compatible #4

Open JDesperado opened 2 months ago

JDesperado commented 2 months ago

Hello, author. What version of cuda and Nvidia driver are you using? tensorflow1.14.0 officially supports cuda 10.0, torch1.11.0 officially supports cuda 10.2, but the 10.2 version does not support tesnsorflow1.14.0, and these issues also lead to driver incompatibility. May I ask how to solve this problem? I run this project under ubuntu18 system,I need your help.Looking forward to your reply,thank you!

Tviskaron commented 2 months ago

Hi, I’m not the author of this repo, but I’ll try to help.

First of all, the core network of SCRIMP is implemented in pytorch, so tensorflow is a redundant dependency. As a working example, I recommend the docker setup from the POGEMA benchmark. Here is the link: POGEMA Benchmark Dockerfile. If Docker isn’t an option, consider using cuda 11.6 with torch<2.0,>=1.9, which worked for our case.

JDesperado commented 2 months ago

Hi, I’m not the author of this repo, but I’ll try to help.

First of all, the core network of SCRIMP is implemented in pytorch, so tensorflow is a redundant dependency. As a working example, I recommend the docker setup from the POGEMA benchmark. Here is the link: POGEMA Benchmark Dockerfile. If Docker isn’t an option, consider using cuda 11.6 with torch<2.0,>=1.9, which worked for our case.

Thank you for your reply. I have taken your suggestion, but now I have encountered a new problem. As shown in the picture, I hope to get your help! 2024-09-03 10-22-17 的屏幕截图

Tviskaron commented 2 months ago

I haven't encountered such an error. It seems that there might be an issue with the initialization of self.PROB, which should default to OBSTACLE_PROB = (0.0, 0.5)