Look at your needs, if you want to run a little larger neural network (e.g. AlexNet), preferably with the GTX 770 or better, Titan, K40 and other GPUs. If only mnist on the run to play the normal card can.
There is not much CPU requirement, the memory of video card is more than 3g
To use CUDNN, the GPU must be capable of operating at 3.0. It's possible without a GPU, but it's very slow.
There is no requirement for the GPU, the only requirement is that the graphics card support Cuda (A-card tear-Ben. )。 If your video card does not support CUDA, it does not matter, you can run with the CPU. In the last line of your network profile Solver.prototxt file, set it to CPU mode.
https://www.zhihu.com/question/27647998
Caffe compile time, depending on the Python version, and the version you are currently using inconsistent, resulting in Python, but the version is different.
Workaround: Try to keep only one version of Python in your computer. Mac system comes with a good
Caffe GPU version Configuration under Windows * * *
Sometimes it appears that a header file in the NumPy is not found
Cause: The NumPy path in Makefile.config in Caffe is not correct
Resolution: Enter the Caffe Makefile.config file, and carefully review the path portion of the python_include,python_lib, and modify the comment for that comment out. and then recompile
About the experience of running someone else's demo
In the case of Caffe with no problem, is mainly used in the Solver training network or testing the data source path, must be carefully changed. When using the Python interface, you often write a data provider py file that works with the network in Train.prototxt (or val.prototxt) to indicate which provider is used to provide data for that network. Usually it is to write the path of the picture into the text, then according to the text to read the picture, this text is usually train.txt or test.txt
caffe-windows Environment Configuration (detailed on GitHub's official Bvlc/caffe recommended configuration method) * * * * *