one vs multiple models for a swap

[korrupt78]

I've heard some say that it helps to split the training set into subsets based on face orientation or expression, and then train a separate model for each subset.

Does that make any sense? Is it always a good idea, or only in some circumstances?

[DeliciousCaramels]

I am no expert but I have been thinking about this question a lot.

I assume the modelling is built to expect the face in different orientations, and as you know the face is a 3d object. I think if you do it with discrete orientations as you suggest the model makes assumptions about how the face will look at certain angles, despite it not being able to display or confirm, and, when you actually show it that data "for real" it is difficult for the model as it has to relearn things it has assumed.

I think a better strategy is to feed the model data in chunks. I like to sort the faces by average alignments and I split the data fed to the model based on these bins, eg, if I have 100k images extracted and 6 bins sorted I just divide each binning by 4, put them into different directories and just feed the different sets into the model as I go. Seems to work? Not sure if I am rubbing snake oil on training but it seems intuitively like a good idea.

[torzdf]

Theoretically splitting the data and training multiple models should lead to a better result (you are storing less information in the latent space).

However, I suspect that the time/work involved vs the perceived improvement would not make it worthwhile.

If you do try to experiment with this, I would be interested to see the results.