# Notes for meeting 28/11/22

Appologies that can't make meeting, here are my notes,
Kirsten Crane

Actions since last meeting:
- created venv locally
- installed packages required so far and updated requirements files
- created folder synthetic_data_metrics to have as package folder and added __init__.py - should we perhaps come up with a better package name at some point?
- split code in inception_score.py into inception_score.py, datasets.py, and utils.py in a first step toward some modularity
- added some docstrings
- created main.py just to load in data and run IS metric, can delete once development on driver class or similar is underway

Small problems encountered:
- anyone know of a way to auto format line length? none of the linting tools I looked at seemed to do that
- pushing to repo generated email 'flake8 Lint: All jobs have failed'
- as much as I understand what github actions and Projects are for, probably still require some demos to get the most out of this functionality. Added a to do item to our project but wasn't able to assign myself
- The debugger in vscode wasn't working for me. I think I have come across this problem before when using a virtual environment. One step is to make sure the venv is selected as the python interpreter, which I have done, but I vaguely remember another step around altering the launch.json perhaps. Does anyone know anything about this?

Bigger problems:
- when I run main.py I get memory allocation errors. Cifar10 (161 MB) loads fine and I think the Inception-v3 weights are loading fine (~100MB), I think it is calling predict on the model which is the issue.

Plans between this meeting and next for myself:
- Try and solve memory issue. First plan is to play with batch size. Next plan is to try and work out if GPU is required and if it is I'll have all the fun around installing the right versions of cuda etc.

Notes to anyone else:
- I had more of a look at the SyntGauge repo and their architecture and the use of an an Evaluator class which you create an instance of and then add metrics to is quite interesting. Everyone, and in particular Ian, should take a look and decide if we want to copy or come up with a different approach