Uni-Mol
Uni-Mol copied to clipboard
Integrating Uni-Mol2 with unimol_tools
Description:
I’m working to integrate Uni-Mol2 into the unimol_tools
repository and encountered some issues with configuration compatibility. I tried modifying the architecture and model configuration, as well as adjusting mol.dict
to mol.dict_new.txt
, but I’m not achieving the expected performance.
Here’s a summary of the integration steps I took:
-
Modified Architecture in
unimol.py
- Substituted
molecule_architecture
parameters to match Uni-Mol2 specifications. - Adjustments include reducing
encoder_layers
to 12,encoder_embed_dim
to 768, and setting other parameters as per Uni-Mol2-84M specifications.
- Substituted
-
Model Configuration
- Updated
MODEL_CONFIG
to point to the Uni-Mol2 checkpoint (checkpoint-84M.pt
). - Adjusted path in
weights
and confirmed the correct.pt
file location.
- Updated
-
Dictionary Adjustment
- Created a new dictionary file,
mol.dict_new.txt
, with 128 rows instead of the original 31, to match Uni-Mol2’s expected128x128
dimension.
- Created a new dictionary file,
Observed Issue:
The model is not performing as anticipated, and the integration does not seem fully compatible with Uni-Mol2’s configuration. Specifically:
- The expected behavior and accuracy are not achieved despite matching configurations to Uni-Mol2 specs.
- Potential misalignment in the dictionary file or model configuration could be causing the issue.
Questions:
- Could you identify any potential missteps or overlooked configuration requirements to ensure seamless integration?
- Are there plans to add official support for Uni-Mol2 within the
unimol_tools
repository? This would be highly beneficial to ensure compatibility and streamline the setup process.
Steps to Reproduce:
- Modify
unimol.py
as per the steps above. - Adjust
MODEL_CONFIG
to point to thecheckpoint-84M.pt
. - Replace
mol.dict.txt
withmol.dict_new.txt
containing 128 rows. - Run the model and observe deviations from expected performance.
Thank you for your assistance and for considering future Uni-Mol2 support in unimol_tools
!