How does … work?¶
- User defines
tg = TrainGenerator(**configs),
get_data()is called, returning data & labels,
- fed to
- which are then printed, recorded.
- The loop repeats, or
validate() finishes, training may checkpoint, and
train() is called again. That’s the (simlpified) high-level overview.
Callbacks and other behavior can be configured for every stage of training.
dg = DataGenerator(**configs).
If not specified,
dginfers the number of batches, file extension, data loader, and other necessary info solely from
labels_path; this is “AutoData”.
- Only required is proper file naming; there’s to be a “common” off of which
set_nums, which is how it tracks all data internally.
- Exception to above is if the path is to a single file containing all data; see
x) and labels (
y) can be fetched with
DataGenerator.get(); by default it’ll validate the batch and reset necessary attributes in case data “runs out” - to prevent this, pass
To move on to next batch (which
.get()won’t do automatically), call