Results of pre-trained model

[zyzyyy123]

Hello, I am learning to use your repository. I used the model in models/pre-trained/combined-large-step=1000000.ckpt in the model download link you provided to directly predict without fine-tuning. Could you please tell me what does the sample_smiles obtained in this case mean? Should it theoretically be equivalent to the input smile sequence?

The script I use:

The data I use:

The result I get:

[EBjerrum]

No, not the input SMILES sequence, but the input SMILES molecule. That's the whole idea of the data augementation, we pretrain it to learn about molecules, not SMILES sequences.

[zyzyyy123]

Thank you very much for your reply! I checked and they are indeed different smiles sequences representing the same molecule, but the decoding effect of using the pre-train model on one of my own molecular sets is not that good, and I want to continue training on this molecular set. Should I use the pre-train method or the fine-tune method based on the current pre-trained model?

[EBjerrum]

Hard to say whats the issue and reason for your not so good performance without understanding more about what you are trying to do and what the source of your SMILES strings are. As example, it works best if the SMILES strings are similar to the ones used during training which is RDKit unkekulized (which standard comes out of Chem.MolToSmiles). But apart from that, fine-tuning is the way to go, if you have a specific task in mind.