search

wengong-jin / icml18-jtnn

Junction Tree Variational Autoencoder for Molecular Graph Generation (ICML 2018)
MIT License
509 stars 190 forks source link

is '#Predict stop' in jtnn_dec.py topological prediction? #44

Open anny0316 opened 5 years ago

anny0316 commented 5 years ago

Hello WenGong： is '#Predict stop' in jtnn_dec.py topological prediction of this article? Meanwhile, In mpn.py: ATOM_FDIM = len(ELEM_LIST) + 6 + 5 + 4 + 1 BOND_FDIM = 5 + 6 MAX_NB = 6 #maximum degree of junction tree but , in jtmpn.py: ATOM_FDIM = len(ELEM_LIST) + 6 + 5 + 1 BOND_FDIM = 5 MAX_NB = 15 #maximum degree of the molecular graph Why is there such a difference? Please give me your advices. Thank you very much.

anny0316 commented 5 years ago

@wengong-jin, hello wengong, Don't you need to predict tree roots in the latest code? By default, in the code, the first node of each tree is the root, is it right?

anny0316 commented 5 years ago

@wengong-jin, hello wengong, how do the following values come into being? by the way, how to realize 'Constrained Optimization' which is mentioned in the paper. thank you very much.

valid = 0.9991 unique@1000 = 1.0 unique@10000 = 0.9997 FCD/Test = 0.9770302413177916 SNN/Test = 0.522326049871644 Frag/Test = 0.9950979926332992 Scaf/Test = 0.8655089872053796 FCD/TestSF = 1.5980327517965094 SNN/TestSF = 0.4996388119246172 Frag/TestSF = 0.9926974330760409 Scaf/TestSF = 0.1174452677242035 IntDiv = 0.8562054073435843 IntDiv2 = 0.8503170074513857 Filters = 0.9743769392453208 logP = 0.02464815889709815 SA = 0.15781023266502325 QED = 2.1869624593648385e-05 NP = 0.0962078166269753 weight = 8.657725423864576