To load Python modules on Great Lakes, run source setup.sh.
Do not execute the shell script with bash or ./ as this will not modify the current environment.
The setup script should run pip install -r requirements.txt, but run if needed.
On Great Lakes, you may see the message Defaulting to user installation because normal site-packages is not writeable; this is expected.
To run Tensorboard, first make sure it is installed pip install tensorboard.
Then, you can run at the command line: tensorboard --logdir runs/.
On Great Lakes, you will need to use python3 -m tensorboard.main --logdir runs/
This will start a webserver (e.g. at http://localhost:6016/) with information about your training.
If the Python extension is installed, VSCode will offer some built-in Tensorboard support. The displayed in-line quick-start button may not work with Great Lakes, since the Python module must first be loaded (see Setup).
However, even without the plugin you can view the board in a VSCode tab (since VSCode is a web browser). When you start the board, just click "Preview in editor" on the pop-up that comes up.
Go to (Great Lakes)[https://greatlakes.arc-ts.umich.edu/]. Go to "My Interactive Sessions" and Request a standard desktop instance with GPU, maximum RAM and cores, and one GPU. It may take a few minutes to provision.
Start the VNC viewer (recommended) or SSH (not supported).
Make sure you have activated the environment first with source setup.sh.
To train the model, run python3 src/train_model.py.
Start a Tensorboard server to view results (see above). Model will periodically save checkpoints in the models/ directory and run on the test set.