experiment
aitoolbox.experiment
defines the experiment tracking and performance evaluation components. Because all
implemented components are completely independent from the TrainLoop engine they can be used either on their own
in a more manual mode or as part of the TrainLoop functionality available in aitoolbox.torchtrain
. Due to the
independence of the components, certain elements, for performance evaluation can even be utilized for evaluation of
non-PyTorch models.
In general, aitoolbox.experiment
helps the user with the following:
- Structured and reusable performance evaluation logic definition
- Tracked training performance history primitive
- High level experiment tracking API
- Low level experiment tracking primitives for model saving and performance results saving
- Saved model re-loading low level primitives