Fit training - adding data, tweaking settings while training & loss graph

[Fubared]

So have read that updating / adding new faces while training is a good idea, and using Fit / Final training is useful. I have done both of these while using Realface.

At around 200k iterations of fit training I added some new faces to the existing set by merging in a new aligment.
At 300k iterations I switched to new A & B data (same people) for final training.
At 360k iterations I added more new B to the final training.
While the training previews looked pretty good, the Tools preview was much worse, with very blurry swap, especially around the eyes.
So at 380k or so, I then changed the training configuration, increasing eye and mouth multiplier in Global & increasing input & output size in Realface settings.

Graph below shows the points clearly.

1) Will changing Global settings mid training messed up the training to date?
2) Will changing the Realface input & output settings mid training messed up the training to date?
3) Does the graph look right based on the process ive gone through (adding data, final training & changing settings)?

Thanks

[bryanlyon]

Don't look at your loss numbers. They're really only there to see trends. They don't provide any useful information about the actual swap and will actually only confuse the issue. Instead look at the previews, which will show the actual model results and can be used to judge the quality.

1 and 2 will generally only affect new models. Existing models will continue training using their original settings.

[Fubared]

Ok thanks for reply.

Very useful to know that changing the settings mid training has no impact, as these are set at the start.
And while I understand the loss graph is just an indication of trend and previews are what we need to pay attention to, why did the graph line jumped so clearly after settings were changed, if they don't affect an existing model?

[bryanlyon]

Simply restarting a model will cause jumps like that since optimizer weights are not saved or reloaded.