Production inference quality extremely low

[robrohan]

Hello - I have to say this project is amazing. You've done some incredible work.

I've been playing around with the model and had some great results when running directly in python in "training mode" (aka just loading the checkpoints and doing inference). However, when I try playing with the model in script mode (which, of course, doesn't work because of types, so using trace()), and serving it from a production inference server, the results are not even close to what it would create directly from python.

I've tried other people's onnx "builds" as well, with the same results - max image size 518, and the depth map is unusable. The depth map looks just like a splatter of black and white - almost random looking.

Is there a tutorial or somewhere I can look for help on trying to run this model in a production environment? Or is it a know issue that you have to run this model "unfrozen" directly in python via the checkpoints?

Thank you for your help, and congratulations on such a great piece of work.