This convolution is not supported by cudnn, MXNET convolution is applied

[robotzheng]

when I run the trainning of coco data, it reports "This convolution is not supported by cudnn, MXNET convolution is applied." but, when I run the demo, it does not reports this. it is very slow. Can you help me?

[javierjsa]

Hi,

I have the same problem with demo.py (haven't tried training yet). Which cuda/cudnn versions are you using? I have cudnn 5.1.10 and cuda 8.0.61.

Regards

[lc8631058]

@javierjsa @robotzheng have you solved this problem??

[javierjsa]

@lc8631058 I'm afraid I haven't, but I think it's just a matter of using the right cudnn/cuda versions

[realwecan]

I can confirm that I had this issue before, and it could be resolved by installing the appropriate version of cudnn.

[lc8631058]

@javierjsa so it's just something like Warning and has no influence to the result?

[javierjsa]

@lc8631058 The result might be the same, but I'd say it's probably slower. If I understand the error/warning message, it means it's using the CPU to perform the convolution instead of the GPU. However, I'm just a newbie.

[lc8631058]

@javierjsa thanks a lot

[dajiangxiaoyan]

I have met this problem. Anyone fix it ?

[betterhalfwzm]

I have met this problem. Anyone fix it ? @javierjsa @robotzheng @javierjsa @dajiangxiaoyan @realwecan

[javierjsa]

@betterhalfwzm Sorry, I didn't. However, if you are into semantic segmentation, you should take a look at Mask R-CNN (https://github.com/matterport/Mask_RCNN). They claim to outperform both FCIS and MNC.

[DuinoDu]

I met this problem when using mxnet and I fix it by switch cuda and cudnn version from cuda-9.1 to cuda-9.0. BTW, I install mxnet from src.

[light201212]

because dilation convolution is not support in cudnn5,using cuda8+cudnn7,install mxnet from src will be ok