Speedy improvements

[mrariden]

When using the --save_png flag, the io.save_masks() method attempts to make a PNG of the original image/predicted outlines/predicted masks/cell pose. This takes a lot of time, especially for large images. I've added a default flag to disable the block in io.save_masks() that makes the matplotlib figure. I've exposed this flag to the CLI and API.

Alternatively, the image data could be reshaped manually, rather than letting matplotlib do the work. (@carsen-stringer)

Example CLI data after setting default behavior to skip matplotlib:

• 10 Mb image, 2600x1500
  • Base runtime: 98 seconds
  • With --save_png: 105 seconds
  • With --save_png and --save_mpl: 121 seconds
• 100 Mb image, 5700x6500
  • Base runtime:
  • With --save_png: 668 seconds
  • With --save_png and --save_mpl: 1562 seconds

This also is related to the OOM issues #735, since the memory spikes at the end of matplotlib figure creation.

[codecov]

Codecov Report

Merging #793 (167f273) into main (5fb74fd) will increase coverage by 0.04%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main     #793      +/-   ##
==========================================
+ Coverage   63.36%   63.40%   +0.04%     
==========================================
  Files          14       14              
  Lines        2964     2965       +1     
==========================================
+ Hits         1878     1880       +2     
+ Misses       1086     1085       -1     

Files	Coverage Δ
cellpose/__main__.py	66.44% <ø> (ø)
cellpose/cli.py	100.00% <100.00%> (ø)
cellpose/io.py	63.25% <100.00%> (ø)

... and 1 file with indirect coverage changes

[mrariden]

Tests initially not passing due to persistent GH actions weirdness. (Environment wasn't able to be created). Passing now. @carsen-stringer