>>14033084the net is just supposed to gracefully reduce the dimensionality down to the latent space, then back up
i can trim the dimensions but then I have to write the zeros back in
in the examples everyone scales the image between neg 1 and pos 1 or divide by 255 etc because its colour information,
i leave mine because its not colour and I don't want to transform it back and break the head shape
data isn't available but it looks like
I guess I will try and share the dataframe itself
Here is the link, its a csv, I use pandas to read it in to a dataframe
then run scikit learns MLP regressor with the NN built in, and layers explicitly passed in
https://file.io/ncrap2ZzmYHR>>14033135yes,
And I'm at the moment trying to create variations of head meshes using ML, my previous solution is various combinations of averages but there isnt enough real variation.
Eventually it could be generalised to all 3d mesh problems and create variations
points are in order xyzw xyzw etc
dataframe contains 63 examples in this set, I ripped from a nif>json conversion then extracted their vert morphs its easy enough to inject the morphs back into the json then go back to the game object nif
using this to view output in python
points = np.vstack((x, y, z)).transpose()
v = pptk.viewer(points)
