PyTorch1.2 and CUDA10.0 about rotated_iou HOT 6 CLOSED

from rotated_iou.

hi, my code doesn't support Pytorch 1.2, because some C++ APIs of Pytorch have been changed somewhere between 1.4 and 1.5 if I remember it correctly. The CUDA 10.0 is probably OK, because I only use very basic CUDA features.

I think it's possible to make it work with Pytorch 1.2. At least two places should be modified:

1. all tensor.data_ptr<T>() should be replaced with tensor.data<T>().
2. all TORCH_CHECK macro in utils.h should be replaced with AT_CHECK.

Other modifications can be done by trial and error.

from rotated_iou.

hi, my code doesn't support Pytorch 1.2, because some C++ APIs of Pytorch have been changed somewhere between 1.4 and 1.5 if I remember it correctly. The CUDA 10.0 is probably OK, because I only use very basic CUDA features.

I think it's possible to make it work with Pytorch 1.2. At least two places should be modified:

1. all tensor.data_ptr<T>() should be replaced with tensor.data<T>().
2. all TORCH_CHECK macro in utils.h should be replaced with AT_CHECK.

Other modifications can be done by trial and error.

Many thanks.

1. CUDA10 encountered error information like: 'CUDA kernel failed, invalid device function';

Any helps would be appreciated sincerely!

from rotated_iou.

hi, I mean you should modify the C++ source code.
First, change line 23-24 of cuda_op/sort_vert.cpp from

sort_vertices_wrapper(b, n, m, vertices.data_ptr<float>(), mask.data_ptr<bool>(),
                         num_valid.data_ptr<int>(), idx.data_ptr<int>());

to

sort_vertices_wrapper(b, n, m, vertices.data<float>(), mask.data<bool>(),
                         num_valid.data<int>(), idx.data<int>());

Since tensor.data_ptr<T> is a new API and not supported in old Pytorch versions.

Second, cuda_op/utils.h uses the macro TORCH_CHECK, which is probably not supported in old Pytorch versions. You should replace all TORCH_CHECK in the file to AT_CHECK.

I don't have CUDA-toolkit 10.0 on my machine so I cannot test these changes. I'm not sure if they would work. If you meet further issues, I recommend you to focus on the C++ APIs of Pytorch. Other parts should be OK.

from rotated_iou.

Many thanks! It is successful to compile on the PyTorch1.2 and CuDA10.0 by converting 'tensor.data_ptr<>' and 'TORCH_CHECK' that you mentioned, and another data type in 'box_intersection_2d.py' and 'min_enclosing_box.py'.

Thanks again for your help.

from rotated_iou.

By the way, if I want to use it for KITTI loss calculation, how should I do?

from rotated_iou.