Document the weight sets, including ap2_tf_paper
Browse files
README.md
CHANGED
|
@@ -1,3 +1,43 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: mit
|
| 3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
---
|
| 4 |
+
|
| 5 |
+
# QCMLForge pretrained models
|
| 6 |
+
|
| 7 |
+
Pretrained checkpoints for [QCMLForge](https://github.com/Awallace3/qcmlforge)
|
| 8 |
+
(`apnet_pt`). Files are resolved on demand by
|
| 9 |
+
`apnet_pt.hf_pretrained`; set `QCMLFORGE_AUTO_DOWNLOAD_PRETRAINED=1` to allow
|
| 10 |
+
non-interactive downloads.
|
| 11 |
+
|
| 12 |
+
## APNet2 weight sets
|
| 13 |
+
|
| 14 |
+
| `weights=` | files | description |
|
| 15 |
+
|---|---|---|
|
| 16 |
+
| `qcmlforge` (default) | `am_ensemble/am_{0..4}.pt`, `ap2_ensemble/ap2_{0..4}.pt` | APNet2 ensemble trained by QCMLForge. |
|
| 17 |
+
| `ap2_tf_paper` | `ap2_tf_paper/atom_models/atom{0..4}.pt`, `ap2_tf_paper/pair_models/pair{0..4}.pt` | The AP-Net2 ensemble published with the paper ([`zachglick/apnet`](https://github.com/zachglick/apnet)), converted from TensorFlow. |
|
| 18 |
+
|
| 19 |
+
```python
|
| 20 |
+
from apnet_pt.pretrained_models import apnet2_model_predict
|
| 21 |
+
|
| 22 |
+
pred = apnet2_model_predict(dimers, weights="ap2_tf_paper")
|
| 23 |
+
```
|
| 24 |
+
|
| 25 |
+
### About `ap2_tf_paper`
|
| 26 |
+
|
| 27 |
+
These are the original TensorFlow SavedModel weights converted to PyTorch
|
| 28 |
+
checkpoints, not a retrained approximation: every atom-model tensor transfers
|
| 29 |
+
bit-exactly, and predictions match recorded TensorFlow output to ~1e-6 on
|
| 30 |
+
multipoles and ~1e-4 kcal/mol on interaction energies (float32 reduction order).
|
| 31 |
+
|
| 32 |
+
Two things they require, both handled by `weights="ap2_tf_paper"`:
|
| 33 |
+
|
| 34 |
+
- each `pair{i}` must be loaded with its own `atom{i}` — the checkpoints carry
|
| 35 |
+
no embedded atom submodel;
|
| 36 |
+
- `quadrupole_scale` must be `1.5` (it is stored in the checkpoint config, not
|
| 37 |
+
the state dict; dropping it shifts electrostatics by ~0.5 kcal/mol with no
|
| 38 |
+
error).
|
| 39 |
+
|
| 40 |
+
They also cannot be loaded by the fused `APNet2_AM_MPNN` model.
|
| 41 |
+
|
| 42 |
+
Details, the parity table, and the conversion procedure:
|
| 43 |
+
[docs/apnet2-tensorflow-weights.md](https://github.com/Awallace3/qcmlforge/blob/main/docs/apnet2-tensorflow-weights.md).
|