Structure:

• data/: contains all data related files(datasets,utils for data,...)
• logging/: contains code for parsing, logging and experiment tracking
• models/: contains model implementations and maybe common modules
• notebooks/: contains various notebooks for experimentations/visualizations
• tests/ contains tests for various modules
• config.py: global data and data paths
• main.py: entry point
• run_tests.py: entry point for tests

Practices:

• Inside modules/packages the imports are done using full paths(from Data.utils import ..) even if utils is a twin module
• packages/modules are lower snake case while classes are Capital camel case
• Models should provide a train_step, val_step, eval_step, get_criterion & get optimizer
• Models receive hyperparameters and training details from a config.yml file
• main.py entry file will make sure that all components are on the same device and will combine modules(data,model,logger) to create and run an experiment
• Before each experiment the current state of models & code should be saved by a commit, the commit hash will be included in the logging folder