search

CNCLgithub / GalileoEvents

Galileo + Events
MIT License
2 stars 0 forks source link

smashing inference bugs #9

Closed belledon closed 4 years ago

belledon commented 4 years ago

as of e30d9a4c8555b1f3b6ab7f7d754551c5dde718b4 there a few perplexing issues in the inference trace.

Below is a plot of the weighted particles followed by a visualization of the predictions from the unweighted samples as a function of time. THis is for a scene in which iron is very light.

In the plot there seems to be only 2 hypothesis which is rather weird... My first guess is that there is something with the rejuvinations proposal..

1_plot

Wrt to the visualization, I'm really not sure what's going on the with the continuity of motion.. its most likely not a bug in the vizualization code as it is mostly the same was it was in the older examples in the fall. My hunch is that it is a bug in the inference analysis/extraction pipeline as motion seems to start and end in reasonable places given the posterior.

1_viz

belledon commented 4 years ago

I ran an oracle (1 particle, has gt) and thankfully the markovian physics is equivalent to the full physics 0_viz

belledon commented 4 years ago

interestingly, adding more particles, even though they ahve the same initial state does lead to some variance in the trace

belledon commented 4 years ago

0_viz

this gets reflected in the log score of the inference trace 0_plot

belledon commented 4 years ago

@iyildirim this is what i posted on slack

10 particles on an incongruent task with material prior

0_viz

iyildirim commented 4 years ago

i see -- is this like the actual particle filter with proposals etc? It's not trying to simulate ground truth for debuggin purposes?