Takes 3min ! very slow in setting up data layer in new Pascal GPU

[ywu36]

Issue summary

Hi,
I bought a 1080Ti and installed it on my Ubuntu 14.04 machine.
When I try to load my trained model using the following code:

caffe.set_mode_gpu()
net = caffe.Net(model, weights, caffe.TEST)

It stuck nearly 3min in setting up data layer in my ConvNet:

I0709 11:55:58.938182 14391 layer_factory.hpp:77] Creating layer data
I0709 11:55:58.938237 14391 net.cpp:91] Creating Layer data
I0709 11:55:58.938251 14391 net.cpp:399] data -> data
3min later ....
I0709 11:58:43.847461 14391 net.cpp:141] Setting up data
I0709 11:58:43.847476 14391 net.cpp:148] Top shape: 1 3 160 160 (76800)
I0709 11:58:43.847477 14391 net.cpp:156] Memory required for data: 307200
....

It usally takes 0.001 second to set up the layer. I disable CUDNN usage, the problem remains.
The MNIST example does not have any problem.

Regards,
Rein

Steps to reproduce

caffe.set_mode_gpu()
net = caffe.Net(model, weights, caffe.TEST)

Your system configuration

Operating system:　ubuntu 14.04
Compiler: gcc
CUDA version (if applicable): 7.5
CUDNN version (if applicable): Disabled
BLAS:
Python or MATLAB version (for pycaffe and matcaffe respectively): Python 2.7.12

[ywu36]

It is because the CPU was loading the Pytorch models but the refreshing delays in the 'console' of Pycharm.