Normalizing geometric features is now supported by setting the flag --normalize_invariants. Geometric features are normalized by passing them through a BatchNorm1d as soon as they are computed. We use a different BatchNorm1d layer for each adjacency type, since the features are then fed to separate networks. Note that this is different from adding batch normalization to TEN. We are just using BatchNorm1d to standardize geometric features in the beginning of the forward pass: 31e7a88, c532725
A quality-of-life feature that avoids reading the entire dataset if --num_samples is set. This makes debugging take less time: a38b244
This small PR introduces two enhancements:
--normalize_invariants. Geometric features are normalized by passing them through aBatchNorm1das soon as they are computed. We use a differentBatchNorm1dlayer for each adjacency type, since the features are then fed to separate networks. Note that this is different from adding batch normalization to TEN. We are just usingBatchNorm1dto standardize geometric features in the beginning of the forward pass: 31e7a88, c532725--num_samplesis set. This makes debugging take less time: a38b244