QM9 Dataset Proportions Sum Issue

[feiji110]

The proportions listed for the QM9 dataset in examples/test_CGCNN.sh do not sum to 1 ([0.4110, 0.3026, 0.0602, 0.0735, 0.0015]). Could you please clarify how these values were calculated?
examples/test_CGCNN.sh

[ftherrien]

Hi @feiji110,

Thanks for the issue. The reason it is not adding up to 1 is because the rest is "empty atoms" (15%). When I originally calculated this statistic I was using a subset of QM9 of less than 20 atoms. In that subset the proportions are: [0.4110, 0.3026, 0.0602, 0.0735, 0.0015, 0.1513] where 0.1513 is the number of empty atoms.

I originally used proportions that did not add up to 1 so one could use the sum to control the graph size. However, in the current version of the code the proportion is actually meant to add up to one.

I will update the code to automatically normalize the config proportions. FYI the actual proportions should be: [0.51145674, 0.35286766, 0.05638106, 0.07804472, 0.00124983]

In practice, proportions can be set very loosely and I did not set a convergence criteria for it, so the actual numbers do not matter much.