error in deformable_col2im: invalid argument

[chengdazhi]

error in deformable_col2im: invalid argument

[chengdazhi]

Possible cause: channel dimension too large.

You can try to change the src/deform_conv_cuda_kernel.cu at line No.68 to

const int CUDA_NUM_THREADS = 1024;
const int kMaxGridNum = 65535;
inline int GET_BLOCKS(const int N) {
  return std::min(kMaxGridNum, (N + CUDA_NUM_THREADS - 1) / CUDA_NUM_THREADS);
}

and see if this can solve the problem.
 

[gasvn]

I met a similar error. When the input image is larger than 1024x1024 (1, 512, 128, 128 for that layer), the same invalid argument error occurs. I think the reason for this error is that the pre-allocated memory space is not enough. But I not familiar with the Cuda program in pytorch, so I can't find a solution yet.

[happycoding1996]

Same problem. Pytorch 0.4.0, python 3.5, cuda 8.0.

[chengdazhi]

Hi, could you provide a short code snippet to reproduce the bug? Thanks a lot!

[chengdazhi]

FYI, my environment is PyTorch 0.4.1, python 3.6/3.7, cuda 9.0.

[shaoqb]

Similar error，how to solve it? Pytorch 0.4.1, python 3.7, cuda 9.0

[jwwangchn]

@chengdazhi It solved my problem. Thanks.

[chengdazhi]

@jwwangchn Thanks for your feedback. I have updated the .cu file in the commit 50a282 .

[abhijay9]

Hi,

I have set up the deform conv on a machine with a single 1060 and it works but when I tried using the same on a machine with (multi-gpu) GTX TITAN X it gives me an error:

error in modulated_deformable_im2col_cuda: invalid device function

This is already present in the file deform_conv_cuda_kernel.cu
#13 (comment)

What do you think might be the problem?

[MurrayC7]

Hi,

I have set up the deform conv on a machine with a single 1060 and it works but when I tried using the same on a machine with (multi-gpu) GTX TITAN X it gives me an error:

error in modulated_deformable_im2col_cuda: invalid device function

This is already present in the file deform_conv_cuda_kernel.cu
#13 (comment)

What do you think might be the problem?

@chengdazhi @abhijay-g Did you figure it out? I also met this issue.