BC03 3DoF, 1 Sphere, PN++ SVD Pointwise Tool Flow (after bugfix).
Has 15 trials with and without the consistency loss (and, w/out consistency but with OLDER v02 3DoF data). After the 05/16 bug fix where we properly scale `data.pos`.
Created on May 16|Last edited on May 27
Comment
Whew, I think we might have fixed the earlier PN++ with SVD stuff. Here I'm doing the two versions with and without consistency loss together. (Since this is after 05/14, evaluation statistics can be fully trusted.) Observations:
UPDATE 05/26/2022: ignore these two runs (though I have them here just in case):
- New v08 data, without consistency loss.
- Learning seems slow but trending up after 200 (!) epochs.
- New v08 data, with consistency loss.
- Learning seems slow but trending up after 200 (!) epochs.
LOOK AT THESE RUNS:
- Old v02 data, without consistency loss.
- Older wandb report HERE. This was trending up after ~50 epochs. However, I don't think we can look at the eval/MSE_loss here but we can possibly see if the eval/info_done_final or train/MSE_loss has a similar structure?
- Update: wow, this is doing VERY well. The results over 5 trials seem a little better than they were earlier (but here we don't zero-pad, etc.) so I would hope this is better. :) Turns out that's the case!
- Old v02 data, with consistency loss.
- I ran these 05/20/2022 and 05/21/2022. (Edit: and 05/22 as the 05/21 runs crashed)
- These results are now updated, whew. TL;DR with consistency seems slightly worse even at the end (unlike the ee2flow results with v04 data) but the similar trend is that the rate of improvement at the end is higher with consistency. Maybe it just requires more training?
Look at the end for more GIFs, and maybe some more targeted analysis of the results, these are a bit crammed with 20 curves total.
Section 1
5
5
v02 data, no consistency
5
v02 data, with consistency
5
GIFS
Code
After these commits:
Add a comment