Currently, inference in done using inference.py works completely independent from any dataset class and methods defined in dataset.py. However, many functionality is duplicated, not only making things inefficient and unreadable, but also increasing the risk of a mismatch between what happens during presto training and what happens during inference in production. We should adapt inference.py to make use as much as possible of the functionality defined in dataset.py.
kvantricht
commented
2 months ago
Currently, inference in done using
inference.pyworks completely independent from any dataset class and methods defined indataset.py. However, many functionality is duplicated, not only making things inefficient and unreadable, but also increasing the risk of a mismatch between what happens during presto training and what happens during inference in production. We should adaptinference.pyto make use as much as possible of the functionality defined indataset.py.