Fix issues with GLN (#92)

The GLN model is the only model class that uses a bespoke environment
based on Python 3.7 instead of the shared environment that uses Python
3.9. Due to this, it is not tested in CI, and is generally prone to bugs
being introduced over time.

After trying to run the GLN model using the weights uploaded to
figshare, I found that the upload was incomplete, and some files were
missing from the checkpoint, preventing the model from being loaded. On
top of this, there were a few small issues in the wrapper that
accumulated over time, which also needed to be addressed to get the
model to run again. This PR changes the weights link to a complete
upload and also fixes the small issues in the model wrapper.

Author: kmaziarz

CHANGELOG.md

@@ -10,6 +10,7 @@ and the project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.
 ### Fixed
 
 - Fix incorrectly uploaded RetroKNN weights ([#91](https://github.com/microsoft/syntheseus/pull/91)) ([@kmaziarz])
+- Fix GLN weights and issues in its model wrapper ([#92](https://github.com/microsoft/syntheseus/pull/92)) ([@kmaziarz])
 
 ## [0.4.0] - 2024-04-10
 

docs/single_step.md

@@ -8,7 +8,7 @@ See table below for the links to the default checkpoints.
 | Model checkpoint link                                          | Source |
 |----------------------------------------------------------------|--------|
 | [Chemformer](https://figshare.com/ndownloader/files/42009888)  | finetuned by us starting from checkpoint released by authors |
-| [GLN](https://figshare.com/ndownloader/files/42012720)         | released by authors |
+| [GLN](https://figshare.com/ndownloader/files/45882867)         | released by authors |
 | [Graph2Edits](https://figshare.com/ndownloader/files/44194301) | released by authors |
 | [LocalRetro](https://figshare.com/ndownloader/files/42287319)  | trained by us |
 | [MEGAN](https://figshare.com/ndownloader/files/42012732)       | trained by us |

syntheseus/reaction_prediction/inference/default_checkpoint_links.yml

@@ -1,6 +1,6 @@
 backward:
   Chemformer: https://figshare.com/ndownloader/files/42009888
-  GLN: https://figshare.com/ndownloader/files/42012720
+  GLN: https://figshare.com/ndownloader/files/45882867
   Graph2Edits: https://figshare.com/ndownloader/files/44194301
   LocalRetro: https://figshare.com/ndownloader/files/42287319
   MEGAN: https://figshare.com/ndownloader/files/42012732

syntheseus/reaction_prediction/inference/gln.py

@@ -57,6 +57,10 @@ def __init__(self, *args, dataset_name: str = "schneider50k", **kwargs) -> None:
 
             self.model = RetroGLN(self.model_dir, chkpt_path)
 
+    @property
+    def name(self) -> str:
+        return "GLN"
+
     def get_parameters(self):
         return self.model.gln.parameters()
 
@@ -73,7 +77,9 @@ def _get_model_predictions(
             return process_raw_smiles_outputs_backwards(
                 input=input,
                 output_list=result["reactants"],
-                metadata_list=[{"probability": probability} for probability in result["scores"]],
+                metadata_list=[
+                    {"probability": probability.item()} for probability in result["scores"]
+                ],
             )
 
     def _get_reactions(