BC03 3DoF, 1 Sphere, PN++ Pointwise Flow no consistency (debugging, prior data)
Using older data (ladle_algorithmic_v02) to debug.
Created on May 16|Last edited on May 16
Comment
Yeah, something is really not right here.
These are 05/14 runs so we can look at eval results.
Why are the results this bad? Is there a bug or should we get more sideways movement? Eval MSEs seem OK.
By the way this is with the SAME data (`ladle_algorithmic_v02`) as these results:
The main change I can think of is with the scaling. Let me check this again. Not sure if the MSEs can be directly comparable to the earlier results... flow vectors should be in the range of (-1,1) here. Point cloud should have been originally scaled as well by 250x (but then divided by 250 when passing into the network). But why is PN++ doing well with just that averaging layer?
Section 1
Run set
5
Add a comment