Command line argument to limit posterior generation memory usage

[sjfleming]

Create a new input argument to control the batch size used by the posterior inference step.

Unclear if the problematic line is here

CellBender/cellbender/remove_background/infer.py

Line 413 in 0feb5e0

batch_size=20,

or here

CellBender/cellbender/remove_background/infer.py

Line 383 in 0feb5e0

n_cells = min(100, cell_inds.size)

but figure this out.

Also explore memory usage during posterior generation. Why is it spiking? Is this avoidable?

[sjfleming]

Totally clearing out CUDA memory after training but before posterior generation seems to help out a lot. Not sure if it will totally solve this for everybody, but this will be in the 0.3.0 release, and should fix the issue for most people

[hoondy]

Changing batch_size=20, and n_cells = min(10, cell_inds.size) from CellBender/cellbender/remove_background/infer.py worked for me.

[sjfleming]

Closed by #238