ornet.analysis.flow.region_flow: Rebuilding of GMM from serialized params

[magsol]

In region_flow, the GMM for each frame of the video is rebuilt from the serialized means and variances (precisions). From what I can surmise, this is done so that predict_proba can be called on each pixel to derive a weighting scheme for the subsequent coloring.

However, the entire GMM fitting process has to play out first. At best, this introduces a lot of overhead; why not either serialize the pixel probabilities when the model is first run alongside the model parameters, or simply evaluate the probability of each pixel under each component (which is what the scikit GMM predict_proba does anyway!), for which you already have fully-specified parameters at the start of the function.

At worst, I fear there could be unobserved consequences--that the gmm.fit function performs some randomized or stochastic steps that can't be reproduced, resulting in a visualization that is entirely dependent on the specific run. The fact that the parameters are initialized is good, and in theory the resulting model training should be entirely deterministic, but without looking at the source there's no way to know.