Dmitriy Gorovoy

Seems like I finally solved the issue. However, now I got another issue The problem is with setting an attribute in retrosynthesis.py: graphs.node_feature=node_feature because in the Molecule.py property was used:...

@FanmengWang @bhadreshpsavani After applying except: pass I got an error different from yours: ``` Traceback (most recent call last): File "tests.py", line 133, in predictions = synthon_task.predict_reactant(batch, num_beam=10, max_prediction=5) File...

@FanmengWang Are you sure we can skip this part? because it is beam search and, for instance, num_synthon_beam is used even in __init__ part of Retrosynthesis in retrosynthesis.py. And in...

@FanmengWang Hi! Concerning your first note, yeah I agree, I mixed something and beam search is used in a different part. So it should be fine. And yeah, I have...

Where did you find that code? I see the one here https://torchdrug.ai/docs/tutorials/retrosynthesis.html#prepare-the-dataset but it is different and has no argument lazy at all. Moreover, the code from the link worked...

seems like it has already been solved https://github.com/DeepGraphLearning/torchdrug/pull/24

I changed the dataset feature from node_feature to atom_feature as you proposed in one of the previous issues. Except for this, everything is right as in the tutorial. Initially, I...

@KiddoZhu I with the help of Chence Shi, I got close to the reported results. If you need I can send you my eventual hyperparameters. And in general, seems like...

@delpqhiz Hi! Sure. here are the results (they are a bit lower than the reported ones but at least they are close): top-1 accuracy: 0.422131 top-10 accuracy: 0.797131 top-3 accuracy:...

@delpqhiz thank you for letting me know! It's weird that top-k accuracies for k>2 decreased after your modification but overall I believe it works. Concerning your second problem, I will...