Caffe

From UMIACS
Jump to navigation Jump to search

Cuda + cuDNN

On machines running RHEL7 and the nvidia driver installed you can build Caffe with Cuda and cuDNN. This has been tested as of Nov 12th 2015 commit 19028e7975fa464ce6835d513a25111e6aedda47.

$ nvidia-smi -L
GPU 0: GeForce GTX TITAN X (UUID: GPU-dd20c890-3d0f-942f-6e85-7b43c6bd03c5)
GPU 1: GeForce GTX TITAN X (UUID: GPU-b1d93efa-3d89-bb58-d4cb-b517ead36b27)
GPU 2: NVS 310 (UUID: GPU-3c199b93-5899-ba43-bbe3-870a98f76457)

To build caffe follow these instructions.

module add cuda/7.5.18 cudnn/v3
git clone https://github.com/BVLC/caffe.git 
cd caffe
cp Makefile.config.example Makefile.config
# set the following options in a Makefile.config
#   !!! Make sure there are no trailing whitespace after these options
# CUDNN := 1
# BLAS := open
# CUDA_DIR := /opt/common/cuda/cuda-7.5.18/
make 
make test 
make runtest

Matlab

If you need to use then please use Matlab 2015a and make sure you do the following commands in addition to what is above. Currently Matlab 2015a only supports GCC 4.7.x for compiling mex files.

module add matlab/2015a gcc/4.7.4
make matcaffe

CPU Only

To build caffe follow these instructions.

git clone https://github.com/BVLC/caffe.git 
cd caffe
cp Makefile.config.example Makefile.config
# set the following options in a Makefile.config 
# CPUONLY := 1 
# BLAS := open 
make 
make test 
make runtest