ReDet

ReDet copied to clipboard

Hi,

I tried your code with the 30xx GPU series which only works for Cuda toolkit >=11.0 (using the mvcc library adapted with Cuda 11.0). Unfortunately, your code can't work with Cuda 11.0. I think the reason is from the "compile.sh" file. Can you have some guides to deal with the Cuda 11.0.

Thanks.

I guess the reason is: cuda11.0 requires higher version pytorch (>1.3), while some ops in our code are designed for pytorch<1.5.

If so, to fix this, you need to replace all AT_CHECK with TORCH_CHECK in the source code (.cpp and .cu). See https://github.com/pytorch/pytorch/issues/36581

Thanks for your suggestion. I will try it and inform the result.

Thanks for your suggestion. I will try it and inform the result. hello, have you worked it out? I still meet with the problems in 30** GPU

Have you tested successfully on CUDA11.0 ?@hphuongdhsp

Hi,

I tried your code with the 30xx GPU series which only works for Cuda toolkit >=11.0 (using the mvcc library adapted with Cuda 11.0). Unfortunately, your code can't work with Cuda 11.0. I think the reason is from the "compile.sh" file. Can you have some guides to deal with the Cuda 11.0.

Thanks.

Have you tested successfully on CUDA11.0 ?@hphuongdhsp

Hi, I am also trying it on CUDA11.0. I want to know if you success? Because 11.0 need higher torch version that causes many problems. I have tried for almost 1 week.

find . -type f -exec sed -i 's/AT_CHECK/TORCH_CHECK/g' {} +

I guess the reason is: cuda11.0 requires higher version pytorch (>1.3), while some ops in our code are designed for pytorch<1.5.

If so, to fix this, you need to replace all AT_CHECK with TORCH_CHECK in the source code (.cpp and .cu). See pytorch/pytorch#36581

Thanks, it works with cuda 10.2 and pytorch1.7 !

I guess the reason is: cuda11.0 requires higher version pytorch (>1.3), while some ops in our code are designed for pytorch<1.5. If so, to fix this, you need to replace all AT_CHECK with TORCH_CHECK in the source code (.cpp and .cu). See pytorch/pytorch#36581

Thanks, it works with cuda 10.2 and pytorch1.7 !

@chongkuiqi is the ReDet code compatible with cuda 10.2 and pytorch1.7?

I guess the reason is: cuda11.0 requires higher version pytorch (>1.3), while some ops in our code are designed for pytorch<1.5. If so, to fix this, you need to replace all AT_CHECK with TORCH_CHECK in the source code (.cpp and .cu). See pytorch/pytorch#36581

Thanks, it works with cuda 10.2 and pytorch1.7 !

@chongkuiqi is the ReDet code compatible with cuda 10.2 and pytorch1.7?

Oh! I'm sorry that i had this issue when using the S2ANet code, and this solution works with cuda 10.2 and pytorch1.7, and i'm not sure if the ReDet code is compatible with cuda 10.2 and pytorch1.7.

I guess the reason is: cuda11.0 requires higher version pytorch (>1.3), while some ops in our code are designed for pytorch<1.5.

If so, to fix this, you need to replace all AT_CHECK with TORCH_CHECK in the source code (.cpp and .cu). See pytorch/pytorch#36581

能请问一下这个cuda ops在哪吗 工程里面没有这个文件呀

@colorfully-ybf See mmdet/ops

@colorfully-ybf See mmdet/ops

ok，I will try it.

@colorfully-ybf See mmdet/ops

@csuhan @colorfully-ybf I tried changing it in the ops folder, still, many errors exists. But please share your results @colorfully-ybf maybe in your implementation, it will work. Thanks!