I want to take existing inference code (training was already done) in TensorFlow and deploy it on the TX2 to get a demo fast.
My initial thought was to setup TensorFlow on the TX2, thus saving the conversion of the code to TX2-specific code at the cost of possible poor performance, until I will learn the other alternatives, like using tensorRT.
So far, I was able to build TensorFlow 1.0.1 following Jetsonhacks’ tutorial on the TX2.
The code requires TensorFlow 1.0.1, python 2.7 and modules like ‘scipy’, ‘matplotlib’ & ‘PIL’
However, I failed to install the modules following https://scipy.org/install.html,
e.g. “pip install --user scipy matplotlib”.
The modules depend on other modules that are not installed, e.g. mkl_rt (which I guess I don’t need on the Jetson, right?).
So my questions are:
- How to deploy these modules on the TX2?
Is there some distribution like Anaconda that can be easily deployed on the TX2 and simplify the process?
- Is it inevitable to use tensorRT or equivalent tool? how can I do it and migrate the code to tensorRT?
Any help is highly appreciated,