Hi MXNET community
I am trying to cross compile mxnet from sourec for jetson nano and TX2. I have followed the steps listed on http://mxnet.incubator.apache.org/versions/master/install/install-jetson.html But i had to make some changes to teh steps in order to compile. The changes are as follows:
- Cloning the v1.4.x branch instead of origin as the steps indicate installing numpy 1.15.2 which later contradicts with version required during installing python bindings
- I think thers’s a typo in the step that suggests changing make file, “-gencode arch=compute-63, code=sm_62”. I kept getting errors regarding compute-63 arch as its not supported. I was able to get rid of this error by changing “-gencode arch=compute-63, code=sm_62” to " -gencode arch=compute_53,code=sm_53 -gencode arch=compute_62,code=sm_62"
I was able to cross compile but couldn’t get python3 script involving mxnet to run. It gives an error related to not able to find libmxnet.so (OSError: path/to/libmxnet.so No such file or directory)
I did clone the same repo, install all the dependencies, updated PATH and LD_LIBRARY variables on jetson nano.
Please anyone who experienced similar issues or have cross compiled please leave suggestion.
PS: Just to confirm distributed training is not supported in pre-built python wheels ???