BlueA_infer

Blue Whale A inference optimization project using Jetson Nano Authors: Daniel De Leon (Cal Poly) and Danelle Cline (MBARI)

Creating working evnironment

Create a python3 virtual environment with access to global libraries on the Nano
- virtualenv --python=python3 --system-site-packages <virtual_env_name>
Activate virtual environment
- in bash: source <virtual_env_name>/bin/activate
- in fish: . <virtual_env_name>/bin/activate.fish
Install juypter notebook in the virtual environment
- pip3 install jupyter
Install custom jupyter notebook kernel
- ipython kernel install --name "<kernel_name>" --user
Run jupyter notebook and start a new notebook with your custom "kernel-name"

python3 /usr/lib/python3.6/dist-packages/uff/bin/convert_to_uff.py <model_name.pb>

Run uff2engine.ipynb:
- Make sure "Automatically deduced input nodes" and "Automatically deduced out nodes" names are placed in the parser.register_input and parser.register_output function - not the names from graph in TF before making the pb file
- The dimensions of the input, however, do need to match those from TF
- NOTE: builder.build_engine(network,config) function takes about 1.5 minutes - Nano became completely occuppied with this process. Saw significant slow down
- TODO: Find best max_batch_size and max_workspace_size parameter values for engine

Run inferEngine.ipynb
- uncomment #!{sys.executable} -m pip install pycuda to install pycuda in the correct path