BC #02 MMMultiSphere PNet2_svd_pointwise_flow
See if flow can be used with SVD for the multi-sphere case.
Created on May 2|Last edited on May 3
Comment
Update: interesting, this is actually getting performance as good as the RGB baseline at the very end, though it seems to take much longer to learn? Hmm ... this is both based on epochs, and the data size is the same (minibatch size is smaller for PN++ but again that shouldn't matter too much).
FYI SVD pointwise flow is MUCH better than naive PN++ regressing to the 3DOF EE, so that's still good to see.
Binary Success Rate, Train/Eval GIFs.
SVD Pointwise Flow
3
Naive RGB
3
Example GIFs
See here for the RGB policy examples. https://wandb.ai/mooey5775/mixed_media/reports/BC-02-MMMultiSphere-Naive-Methods-Regress-Straight-to-3DoF-use-PN-and-RGB-tests---VmlldzoxOTM2MDkz
Here we show the SVD pointwise GIFs. There's a LOT of confusion with red and yellow balls. If that alone were corrected, SVD pointwise would do way better.
Code
Using these commits modulo changes in launch BC script:
Add a comment