make inference faster, and add more specificity to do_seg wrt inputs and outputs

[dbuscombe-usgs]

1. Inference is slow. How do we speed it up?
2. add functionality to control which outputs are created

[ebgoldstein]

Has anybody profiled the code to see what step is the longest in the inference code? (i.e., via cPython or the even the tensorflow profiler?)
I have not done it, but i presume its just the actual model.predict call?
And do we have an understanding of model.predict speed on GPU vs CPU?

[dbuscombe-usgs]

In many of my applications, a single model.predict is a small fraction of the total time. I use several other function calls to prepare inputs and outputs. I also tend to call the model several times for the same image. Writing outputs to file often takes longer than creating them. So, parallelizing the model inference makes sense anyway

According to the docs, model(x) should be quicker than model.predict(x) on a single input

It would be really nice to not have to rely on GPU, which will be out of reach for a lot of users and hinder cloud deployments

[dbuscombe-usgs]

I added a small change to enable the option to use model in parallel using joblib using gym, zoo, etc.

doodleverse_utils/doodleverse_utils/prediction_imports.py

Line 133 in 0999c40

Mc = []

0999c40

[dbuscombe-usgs]

Superseded by #31