search

VITA-Group / FasterSeg

[ICLR 2020] "FasterSeg: Searching for Faster Real-time Semantic Segmentation" by Wuyang Chen, Xinyu Gong, Xianming Liu, Qian Zhang, Yuan Li, Zhangyang Wang
MIT License
524 stars 107 forks source link

RuntimeError:CUDA error:API call is not supported in the installed CUDA driver #82

Open LikeLidoA opened 2 years ago

LikeLidoA commented 2 years ago

hello @chenwydj 555~ I stuck in step one.....
My environment : Hardware:A100-40G Nvidia-Driver:450.xxx(forgot the details) CUDA11.0 python3.6.9 cuDNN8.0.4 TensorRT7.2.5.1 Before I run "train_search.py",I have checked all the requirements. The samples of TensorRT run well. After I run "train_search.py", the "train" is fine,I get a folder named like "search-pretrain-256x512_F12.L16_batch3-20220608-xxxxxx",there are something in it. Also I can see the terminal show all the 20 epoch have finished But when it comes to "validation",something wrong. the terminal show "use TensorRT for latency test" then comes"RuntimeError:CUDA error:API call is not supported in the installed CUDA driver"

Is update the driver the only way I can solve the problem、、、？

ZhouZhengda commented 1 year ago

Hello,l also encountered this problem. when l loaded my data in dataloader "for i, samples_batch in enumerate(data_loader): ", there will be an error shows"RuntimeError:CUDA error:API call is not supported in the installed CUDA driver". l think it maybe caused by the use of Multiple Processes in Python, because when l delete the order "torch.multiprocessing.set_start_method('spawn')", and put all the data in cuda when it was in cpu，they no longer report errors. Hope my answer can help you.