search

theFoxofSky / ddfnet

The official implementation of the CVPR2021 paper: Decoupled Dynamic Filter Networks
MIT License
214 stars 34 forks source link

Also questions about the kernel size #33

Open ydhongHIT opened 2 years ago

ydhongHIT commented 2 years ago

The kernel size can not be larger than 13x13? I set the kernel size to 15x15 and encounter the error "CUDA error: an illegal memory access was encountered". But everything is ok when the kernel size is smaller than 13x13 or equals to it. Besides, I note that the speed is obviously lower as the kernel size goes up, more obvious than common depthwise convolutions. Is that because the implementation is not enough efficient?

theFoxofSky commented 2 years ago

Yes, the Cuda code only speeds up small kernels due to the limited shared memory. You can modify the Cuda code to support large kernels.