BC03 3DoF, 1 Sphere, PN++ SVD Pointwise Flow
WITHOUT the consistency loss. NOTE: cannot look at the evaluation flow visualizations or MSEs because that doesn't have the correct model.eval() but the action selection should be the same. Unfortunately doesn't look that good?
Created on May 13|Last edited on May 16
Comment
Edit: I think there has to be some bug somewhere.
The train/MSE_loss goes down to below 0.004. If you compare with the naive method:
This train/MSE_loss here (PN++ SVD) is better (i.e., lower) but this is per-point on flow, and that includes when we zero-out the tool mask, so those points end up getting more 0s and thus that lowers the train/MSE_loss I think. Nonetheless, the train/MSE_loss plots do look good (they are decreasing, and have expected behavior?) so I'm not sure what's up with such poor evaluation...
We cannot look at the eval/MSE_loss or eval flow visualizations (these are pre-May 14).\
Targets are in (-1,1) for the flow vectors, as expected. It does a good job with predicting up / down movement, but not the sideways movement. I wonder if a consistency loss would help in that case ...
Section 1
Run set
5
Add a comment