search

apax-hub / apax

A flexible and performant framework for training machine learning potentials.
MIT License
12 stars 1 forks source link

Per Batch Padded Dataset #281

Closed M-R-Schaefer closed 1 month ago

M-R-Schaefer commented 1 month ago

Adds a dataset which does not use tf.data and pads samples per batch instead of everything to the largest size. Very advantageous for training datasets containing very differently sized samples.

M-R-Schaefer commented 1 month ago

Needs a one line change but is otherwise ready for review. Note the change in the training config.

M-R-Schaefer commented 1 month ago

pre-commit.ci autofix