Skip to content

Latest commit

 

History

History
119 lines (72 loc) · 3.54 KB

File metadata and controls

119 lines (72 loc) · 3.54 KB

v0.0.21 (2026-08-18)

Feat

  • datamodules: wire multi-conformer ETKDG+MMFF pipeline into Graph3DDataModule._calculate_coords
  • datamodules: add multi-conformer ETKDG+MMFF helpers for 3D coords

Refactor

  • datamodules: generalize embed timeout helper and bump default budget

Perf

  • datamodules: reduce conformer defaults to numConfs=5, maxIters=500

v0.0.20 (2026-08-18)

Feat

  • datamodules: thread optional coords through Graph3DDataModule to skip ETKDG re-embedding
  • datamodules: add shared coords_utils helpers for 3D atomic coordinates

Refactor

  • datamodules: delegate 3D pretraining coord helpers to coords_utils

v0.0.19 (2026-08-17)

Fix

  • torch: resolve leaf encoder when finetuning nested Finetuner artifacts

v0.0.18 (2026-08-16)

Fix

  • cli: sample dense-mode validation compounds globally so val size tracks sampling_rate

v0.0.17 (2026-08-16)

Fix

  • losses: honor reduction in DropoutLoss so it returns per-element output when wrapped by MultiLoss/MultitaskLoss

v0.0.16 (2026-08-14)

Fix

  • losses: drop caller-supplied reduction kwarg in DropoutLoss so it can be wrapped by MultiLoss/MultitaskLoss
  • finetuner: advance global_step in full manual-opt path so MultiLoss weight curriculum interpolates

v0.0.15 (2026-08-14)

Fix

  • finetuner: unpack MultiLoss tuple in training_step

v0.0.14 (2026-08-14)

Fix

  • finetuner: route wrapper MLM path through forward_tokens

v0.0.13 (2026-08-14)

Fix

  • losses: revert MultiLoss to always-tuple return + align callers (stage 1/1)

v0.0.12 (2026-08-13)

Feat

  • pretraining: add GPS3DPretraining and GT3DPretraining (stage 1/1)
  • pretraining: sparsity-agnostic multitask loader + collate (stage 4/5)
  • cli: wire dense branch into prepare_dataset (stage 3/5)
  • cli: dense-mode prep helpers (stage 2/5)
  • cli: rename prepare command + sparse schema toggle (stage 1/5)
  • losses: register 8 dropout-* concrete aliases (stage 2/3)
  • losses: add DropoutLoss wrapper for per-label random masking (stage 1/3)
  • cli: auto-discover multitask coords + docs update (stage 4/4)
  • cli: wire graph3d branch into pretrain_encoder (stage 3/4)
  • pretraining: thread coords through on-the-fly wrappers (stage 2/4)
  • cli: schema + shared coords loader for graph3d pretraining (stage 1/4)
  • pretraining: add E3GNNPretraining model + schema (stage 3/4)
  • pretraining: add Graph3DPretrainingDataModule (stage 2/4)

Refactor

  • encoders: unify 3D encoders on graph.pos contract (stage 1/4)
  • pretraining: unify canonical + MLM RoFormer (stage 4/5)
  • pretraining: delete PretrainingEncoder duplicates for graph models (stage 3/5)
  • encoders: hoist forward to base for gatedgcn/gps/gt/attentivefp (stage 2/5)
  • encoders: unify canonical + pretraining GIN (stage 1/5)

v0.0.11 (2026-08-12)

Feat

  • predictors: add BatchEnsembleLinear primitive for SNN (stage 1/2)

Fix

  • predictors: rewire SNN with BatchEnsembleLinear (stage 2/2)
  • encoders: reconcile E3GNN with reference implementation (stage 1)
  • layers: reconcile SpatialEncoder / SpatialEncoder3d internals (stage 3/3)
  • encoders: reconcile GPS3D and GT3D encoders (stage 2/3)
  • encoders: reconcile GPS and GT 2D transformer encoders (stage 1/3)
  • encoders: reconcile GIN, AttentiveFP, GatedGCN with reference implementations

Refactor

  • encoders: drop PyG private-API dependency in E3GNN and land cleanups (stage 2)

v0.0.10 (2026-08-03)

Feat

  • first commit