Scaling up the kernel size with D3D modules

[leeh43]

Hi, thank you for the great work and currently I have used the D3D packages to perform deformable convolution for volumetric segmentation. The performance is pretty promising with additional tricks in block design. However, when I scale up my kernel sizes (e.g., 5x5x5, 7x7x7), it always pops up the CUDA error as "RuntimeError: CUDA error: an illegal memory access was encountered". Is the c++ code optimized for 3x3x3 kernel size only? If yes, what do you recommend to do for next step about changing the kernel size?

Thank you so much!

[XinyiYing]

We scale up the kernel sizes to 5x5x5, 7x7x7, and can successfully run this code. Details are as follows:

[leeh43]

Thank you for your reply. I also can perform the same thing in test.py. However, when I try to use it for semantic segmentation, the code just blows up with the following error.

I have used the following block design with the kernel size of 5 or 7:

Do you mind to give some suggestions about it?

[XinyiYing]

Try to set the groups to 1.

[Pull2MAX]

Hi, thank you for the great work and currently I have used the D3D packages to perform deformable convolution for volumetric segmentation. The performance is pretty promising with additional tricks in block design. However, when I scale up my kernel sizes (e.g., 5x5x5, 7x7x7), it always pops up the CUDA error as "RuntimeError: CUDA error: an illegal memory access was encountered". Is the c++ code optimized for 3x3x3 kernel size only? If yes, what do you recommend to do for next step about changing the kernel size?你好，感谢你的出色工作。目前，我已经使用 D3D 软件包执行了用于体积分割的可变形卷积。通过在块设计中使用额外的技巧，性能相当不错。但是，当我扩大内核尺寸（例如 5x5x5、7x7x7）时，总是会弹出 CUDA 错误 "RuntimeError：CUDA error: an illegal memory access was encountered"。c++ 代码是否只针对 3x3x3 内核大小进行了优化？如果是，您建议下一步如何改变内核大小？

Thank you so much!非常感谢！

哥们 你怎么编译成功了 我编译好了 ，然后import不进来