Port issue / Pod crash, and other issues still present on Runpod

[qudabear]

The issue where the notebook "can't find a port in range ___" still exists. It doesn't matter if you add that port or an infinite number of ports to the pod, it still crashes, invariably but randomly, every session, multiple times. It has cost me quite literally hundreds of dollars in delays.

The suggested settings also do not work, except in the case that you are training ONLY face pictures, and nothing else. Dreambooth is capable of styles, but with this notebook, it is nearly impossible to get a style right. 2e-6 does not work, nor does 1e-6, for anything above around 30 photos.

The problem lies, most likely, with the text encoder, and balancing it with other variables, but, even though you've been asked multiple times on multiple forums for specific values, you've never given a clear answer to this question: how can we calculate the text encoder value, based on our other values?

Is the text encoder supposed to be set to 15-20% of the images? The total steps? Is it relative to the learning rate? How can we objectively (not guess) calculate what to input for the text encoder, so that we can train something beyond "10 images, 50 steps an image..." ?

Please give objective values to try for larger datasets, styles, multiple people, etc., and which learning rates and text encoder settings work for what type of dataset - and WHY - and please fix the port issue. -_- Please also, if you're going to defend this method, offer image evidence of models showing variation of poses and lighting of full body, half body, and face, not just celebrity head shots. Not trying to be mean, and I do appreciate what you do, but very frustrated with all the money and time I've lost using this "fast" method, because of lack of clarification and errors.

Current dataset is around 70 photos. What learning rate? How many steps? What text encoder? How did you arrive at the numbers you suggest?

OSError: Cannot find empty port in range: 3000-3000 (adding port 3000 does not resolve the issue).