Open gihanpanapitiya opened 3 months ago
Hi @gihanpanapitiya,
I noticed that the losses are different, as in the notebook:
clip loss: -0.92124 L2 loss: 0.33059
This seems to be caused by the corrupted checkpoints. Can you help double-check them?
How can I double check?
These are are checkpoint related parameters I use,
########## for foundation ##########
parser.add_argument("--MoleculeSTM_model_dir", type=str, default="../data/demo/demo_checkpoints_SMILES")
parser.add_argument("--MoleculeSTM_molecule_type", type=str, default="SMILES", choices=["SMILES", "Graph"])
parser.add_argument("--vocab_path", type=str, default="../MoleculeSTM/bart_vocab.txt")
########## for generation ##########
parser.add_argument("--MegaMolBART_generation_model_dir", type=str, default="../data/pretrained_MegaMolBART/checkpoints")
########## for foundation and generation projection ##########
parser.add_argument("--language_edit_model_dir", type=str, default="../data/demo/demo_checkpoints_SMILES")
For zero-shot molecule editing, I am not getting the expected results as shown in the repository's notebook. Do you have a suggestion to improve this result.
This is what I get.