DECODE convergence

[alinkaKuliesh]

Dear DECODE developers,

I was running "Training a DECODE model" on Google Colab without any changes to the original code and parameters.  I observed that values of Jaccard Index are converging to 0.6 after 100 epochs, which is much lower than the values reported in your paper. The same applies to the RMSE. My values didn’t get lower than 50 nm. Unfortunately, I couldn’t resolve this issue on my own and hope that you could help me. 

Sincerely,Alina

[Haydnspass]

Hey,

The absolute numbers on Jaccard and RMSE depend pretty heavily on your simulation parameters (i.e. density, SNR etc.).
The curve that you show does not look like final convergence to me. Have you tried out DECODE on your own data? :)

[alinkaKuliesh]

Thanks for an answer! Indeed, after I played around with the SNR and density of emitters I observed different behaviour during training in terms of metric performance.

[Ano-Li]

Dear Dr Haydnspass，

I also get this situation with my training process, which the JI was about 59%(Precision was pretty good for 96% but bad for recall at 60%) and the rmse for lateral was about 45 nm when convergence(while it was about 28 nm at epoch 14).

I think we have the extremely same problem, so can I get the parameters used in paper to obtain a better model?