Platform | Download link |
---|---|
Linux(cpu, openblas)1 | |
Linux(gpu, cuda9.0)13 | |
Linux(gpu, cuda9.2)13 | |
Windows x64(cpu, openblas) | |
Windows x64(gpu, cuda9.2)3 | |
Mac OS(cpu, appleblas)2 | x |
1: use LD_LIBRARY_PATH=$LD_LIBRARY_PATH:. gluoncv-detect
in case you are using the supplied shared libraries under the same directory. You can add export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:~/path/to/prebuilt-package
to ~/.bashrc
to permanently add library into search path.
2: use DYLD_FALLBACK_LIBRARY_PATH=. gluoncv-detect
in case you are using the supplied shared libraries under the same directory. You can add export DYLD_FALLBACK_LIBRARY_PATH=$DYLD_FALLBACK_LIBRARY_PATH:~/path/to/prebuilt-package
to ~/.bash_profile
to permanently add library into search path.
3: You may want to disable CUDNN autotune feature which is useful in large batch training, but not desirable during single image inference by export MXNET_CUDNN_AUTOTUNE_DEFAULT=0
in Linux/MacOS or set MXNET_CUDNN_AUTOTUNE_DEFAULT=0
in Windows.
-
You can download prebuilt libmxnet binaries from PyPI wheels. For example, you can download mxnet 1.3.0 wheels from PyPI, extract libmxnet.{so|dll} by opening the wheel as zip file(you may change the suffix of *.whl to *.zip). For Linux and Mac, the targeting shared library is libmxnet.so, for windows, the library is libmxnet.dll.
-
You can then replace the libmxnet.* in previously downloaded binary package with the extracted libmxnet from PyPI wheel.
By doing this, You may switch between mxnet cpu/gpu or blas versions without build from source.