Doodleverse/segmentation_gym

Bug in `seg_images_in_folder` when models are ensembled but config files do not all contain same implementation parameters

dbuscombe-usgs opened this issue · 1 comments

dbuscombe-usgs commented

When more than one model is used for prediction, right now, seg_images_in_folder will read in each model config file and concatenate models into a list

However, if the config files do not contain the same implementation parameters, the script will pass only the config parameters from the last config file selected

Config parameters this refers to:

  1. "TESTTIMEAUG"

  2. "WRITE_MODELMETADATA"

  3. "OTSU_THRESHOLD"

  4. "DO_CRF"

This is not a problem when only one of the config files lists the above parameters, but will catch someone out one day unknowingly, so should be fixed. leaving it for now since it is not a priority. Will revisit later

dbuscombe-usgs commented

I don't have any smart ideas here for an easy fix. A proper json parser would be required I think; one that read in each config file and cross-checked them.

In lieu of that, I have amended the seg_images script to print out the inference parameters to screen, so at least the user can see what ones are being used by do_seg