I am trying to install version 2.05.00.2 on Ubuntu 16.04.4 LTS
I did a "make install" which completes successfully. I can successfully run hello_ncs_py and hello_ncs_cpp
However when I try to run mvNCCompile -h it core dumps (Illegal instruction). If I run as sudo then it doesn't core dump but complains that there is no module named mvnc. If I just start python3 and try to import mvnc it works fine.
Note that I am currently running in a virtualenv. Without it I cannot find caffe module and again I can find it if I directly run python3.
@sbv Did you try installing the ncsdk and enabling the virtualenv option in the ncsdk/ncsdk.conf file? By default it is set to install without virtualenv support, but you can enable virtualenv support by editing the ncsdk/ncsdk.conf file.
Yes I was setting virtualenv support from the ncsdk.conf file. Any other suggestions? I'm thinking I should go back and try NCSDK1 since I'm having so much trouble with NCSDK2.
@sbv https://movidius.github.io/ncsdk/virtualenv.html has more information about virtualenv and NCSDK 2 that may be helpful. Also you have to set the virtual env option in the conf file before installing the NCSDK. I'm sorry that my previous message wasn't clear. I don't know if you were using virtualenv with NCSDK1, but NCSDK 1 doesn't officially have virtualenv support.
Thanks @Tome_at_Intel. Yes I had modified the conf file before I did the install and yes I was using NCSDK2. I was just suggesting that I give up on NCSDK2 since I'm having so much trouble and I assume NCSDK1 is better supported. I don't need to use a virtualenv, in fact I originally tried the installation without it and still had issues.
The issue seems to be with TensorFlow - that is why it crashes - it seems like it is expecting avx instructions which aren't supported on the machine I am using. See https://github.com/tensorflow/tensorflow/issues/17411 Using sudo just delayed the illegal instruction since running as sudo messes up my paths. I will try on another machine and see if that solves the problem.
@sbv Ah okay. I have seen that on one of my machines also. Have you downgraded your version of TensorFlow to version 1.4 or 1.5? It seems that some users have had success with doing so.