The changes here are aimed at splitting some of the functionality in ec33.py out into a few different places:
ec3.config.ModelConfig - holds a dataclass for configuring stuff, this could potentially be souped up with hydra later but for now is just a way of setting defaults + making namespace more explicit
ec3.data_exchange - module for the socket client stuff and memmapping functionality
ec3.recognizer.model.Recognizer - probably didn't need to make this a subpackage yet but it feels like the architecture stuff (e.g. things in clip_model) could maybe all be lumped into one area
ec3.ec34 - pared down version of ec33.py
To-do:
get flow for command line args -> config -> script ironed out
continue to pull the data exchange logic out from training script and maybe switch it to a context manager
move loss function set up logic out of ec34 and see how to configure that
This is still a work in progress.
The changes here are aimed at splitting some of the functionality in
ec33.py
out into a few different places:ec3.config.ModelConfig
- holds a dataclass for configuring stuff, this could potentially be souped up with hydra later but for now is just a way of setting defaults + making namespace more explicitec3.data_exchange
- module for the socket client stuff and memmapping functionalityec3.recognizer.model.Recognizer
- probably didn't need to make this a subpackage yet but it feels like the architecture stuff (e.g. things in clip_model) could maybe all be lumped into one areaec3.ec34
- pared down version of ec33.pyTo-do: