High mapping score with big KL reg

[Fclef]

Hi

I tried to map 3000 cells to 1000 voxels. I see the KL reg stayed in large compare to tutorial case. Waht could be the possible reason for this?

Tim

INFO:root:Allocate tensors for mapping.
INFO:root:Begin training with 271 genes and rna_count_based density_prior in cells mode...
INFO:root:Printing scores every 100 epochs.
Score: -0.002, KL reg: 38.268
Score: 0.704, KL reg: 26.072
Score: 0.712, KL reg: 26.075
Score: 0.719, KL reg: 26.078
Score: 0.723, KL reg: 26.081
Score: 0.730, KL reg: 26.085
Score: 0.733, KL reg: 26.088
Score: 0.735, KL reg: 26.089
Score: 0.738, KL reg: 26.090
Score: 0.740, KL reg: 26.090
INFO:root:Saving results..

[Hejin0701]

Hi @Fclef . Thanks for your question. The values of KL divergence is dependent on the number of voxels in the spatial samples. Normally the larger number of voxels in the samples, the larger the KL divergence can be. As long as it stabilizes at a values and does not fluctuate too much it should not be a big concern.