search

tsinghua-fib-lab / DRL-urban-planning

A deep reinforcement learning (DRL) based approach for spatial layout of land use and roads in urban communities. (Nature Computational Science)
https://rdcu.be/dlRPZ
MIT License
159 stars 33 forks source link

Unable to utilize cuda during training #8

Closed WilliamTianYe closed 4 months ago

WilliamTianYe commented 5 months ago

Hello,

Your work is excellent. I encountered an issue while trying to reproduce your work, as I am unable to run it on CUDA. I can confirm that CUDA is functioning properly in my PyTorch setup.

When I attempted to print self.actor_critic_net.actor_net.shared_net.node_encoder.weight.device in the script, it displayed device(type='cuda', index=0). However, I found that the data returned by self.batch_data in the forward function of SGNNStateEncoder is on the CPU. I tried to move this data to the same device as self.node_encoder.weight.device, but found that they were still on the CPU. When I printed out self.node_encoder.weight.device again, I discovered that it was now on the CPU, which contradicts the information I printed earlier.

Could you please check if there might be some part of the model training process that moves the model back from CUDA to the CPU?

Thank you for your assistance.

DavyMorgan commented 4 months ago

Hi, I checked the codes, the model training is performed on CUDA and there is no operations that moves the model back to CPU.