search

MKorablyov / LambdaZero

4 stars 0 forks source link

TPNN #130

Closed L-sky closed 3 years ago

L-sky commented 3 years ago
  1. Update trainer class (use recommended functions, instead of deprecated ones)
  2. Added trainer subclasses that allow use of scheduler with step on epoch / on batch.
  3. TPNN_v0 - wrapper that accepts arbitrary equivariant model.
  4. Equivariant models: TPNN_ResNet, TPNN_Unet.
  5. Modification of train function (with closure) - supports usage of LBFGS optimizer.
  6. Modified mpnn training to match the structure.
  7. Added degree based normalization (keep unit variance for messages sum).