GitLab

New binning methods. They are not really necessary as I gave up on learning with the binned pointcloud loss.
There also might be an error as the binning method is too big for 10/15k data points

Option of binning the pointcloud before calculating mse.

Behavior of linear layers is changed in a strange way. So look at this parameter closely before chosing it anything other than 2 or 3

I hope the whole autoencoder is working.

I hope files within ShellnetJakob are fine again.

Ready to tackle the two main problems.
Memory and computation time when working with the big data.

Tried to find where zeros come from. Understood, that it is form relu cutting negative values. But this should be correct.

Work I did:
Shell_sample works always. Normalization should be good. Check if always between 0 and 1.
Plotting fixed. There is only a few data for test_loader. So it will often be the same.

save_result should be activated
In the log 2624 there is a shaping error 4095 and 4096 particles
Normalization is not yet working
In the plot seem to be the same original image several times. Random?

Two Problems: 1. plot data has some symmetry issues. Axes are changed?
2. My autoencoder gets many zeroes. I tracked them down a little already. Seems, that it starts from the first shellconv layer on and gets more every time!

Idea provided by Nico.
Starting a few runs to see whats happening.

Plotting for results working.
Startet a shell_sample in order to reduce points as preprocessing step.
Determined that the knn calculation in the last shellup Layer is the memory consuming step.
(Added memory control functions)
Changed to MSE-loss.