Low accuracy on mnist with gpu
zhangyuanW opened this issue · comments
Hi everyone, I am new to caffe and I have already compiled this windows version with VS2013 community, Matlab 2015a, CUDA 7.5 and cuDNN v4. My PC is Win10 with GTX 765M. The compilation is successful without error.
The problem is that when I try to run the mnist example, I can get reasonable accuracy with CPU training (0.97 accuracy after 500 iterations). However, when I turn to GPU mode, it only gives 0.11 accuracy even after 10000 iterations. This is the same whether I use bat and run caffe.exe or run the corresponding matlab wrapper.
The solver and network structure is just the lenet in /example/mnist/. The log file for gpu training is attched.
If anyone has encountered similar problem, please help me out. Thank you!
It is working now, after I adapted my compute_30,sm_30 in CommonSettings.props.