Use datashader if too many points in train_board / densityplot

[BernhardAhrens]

If we just color by the marginal density of observations (one-off), this should not add any extra time. A 2D density estimate maybe do expensive for each epoch

[Meng2077]

Not sure if this fast enough to do for all epochs. I used histogram2d from Plots to do such plotting.

[mreichMPI-BGC]

Not sure if still active, but FWIW, I'd tend to 1) sample to say max 1000-2000 points, 2) use some alpha=0.05 or so to show density

[BernhardAhrens]

What you describe was my quick fix ;-) We sample max 10000 points (maybe we decrease?) and have some alpha. Maybe we just need some scaling for the alpha if there are fewer points to 10000 points