Closed najwalb closed 8 months ago
Hi, thanks for interested in this project. I finished this project quite long time ago, but I think I simply used the default hyperparameters. I trained the Transformer by USPTO-50K dataset instead of USPTO-MIT dataset.
I will close this issue if there is no more discussion on this.
Hi, thanks a lot for your reply, it's quite helpful :)
The paper reports both ground-truth and round-trip accuracies. The repo does not provide any information about the round-trip accuracy however. Could you share the parameters you used in MolecularTransformer to get the numbers you reported in the paper? For example, what was your beam_size, n_best, etc?