Open kevin-li opened 8 years ago
Hi @kevin-li, Did you solve this?
I have the same issue, what is the best way to run the same net on multiple incoming images? ps. I know about the batch option, but it needs the images to be already present
hi @shelhamer, any suggestion on this topic? thanks
I don't remember exactly, but I was using the Windows version. Try smaller batches, it worked for me for cudaSuccess(4 vs. 0)
. I think it was my CUDA problem and I had to reinstall it the hard way...
hi @kevin-w-li thanks for the comment. I use batch_size 1, I still get the error message Check failed: status == CUDNN_STATUS_SUCCESS (4 vs. 0) CUDNN_STATUS_INTERNAL_ERROR for some networks.
For example for fcn8s, in this case I should reload the caffe.Net(deploy_file,caffemodel, caffe.TEST) for each image to get it working.
for other nets such as alexnet it works nicely just by loading the net once and using it for multiple images!
if anyone has an idea, please comment
I am using ubuntu 14, and cuda 7.5+cudnn 5
As an update, I installed cuda 8.0 and libcudnn5-dev_5.1.10
I still get the same error ...
You said it works for AlexNet. Could you paste your net definition here? Or try reduce the net size...? Sorry I don't really know the problem.
Hi,
I double checked, this problem only exist for the latest version of NV-caffe and not for the BVLC/caffe
so I will report a bug on nv-caffe.
cheers!
problem solved for nv-caffe in here https://github.com/NVIDIA/caffe/issues/299
maybe this issue can be closed.
I have a forloop in which I load different data into the
Input
layer and runs forward. It crashes my Nvidia driver. I found that this net can only do a forward once, and on the secondnet.forward()
it crashes, givingerror == cudaSuccess (4 vs. 0)
unspecified launch failure. But if I load the net again in the forloop, it works fine.I did not have this problem before, not sure why this shows up now. I tried different versions of Nvidia driver and it's the same.