search

cvlab-stonybrook / SAMPath

Repository for "SAM-Path: A Segment Anything Model for Semantic Segmentation in Digital Pathology" (MedAGI2023, MICCAI2023 workshop)
30 stars 6 forks source link

The difference between ignored_classes and ignored_classes_metric in the config file #16

Closed zhi-xuan-chen closed 2 months ago

zhi-xuan-chen commented 2 months ago

I noticed there two ignore class in the config, what is the difference of them?

jingweizhang-xyz commented 2 months ago
  1. We do not calculate loss nor metrics for unlabeled region. ("ignored_classes")
  2. We calculate loss for background region but do not include it when calculating metrics. ("ignored_classes_metric")
zhi-xuan-chen commented 2 months ago

OK, so if the background loss does not be counted into training, will it affect the performance?

jingweizhang-xyz commented 2 months ago

In that case the network will make a easy decision: predict very pixel to be gland because there will be no penalty on that.

zhi-xuan-chen commented 2 months ago

Right, thank you!