Skip to content

Latest commit

 

History

History
40 lines (31 loc) · 1021 Bytes

README.md

File metadata and controls

40 lines (31 loc) · 1021 Bytes

visualization

  • GradCam++

Usage: just set cfg.VISUALIZED.LAYER_NAMES and follow how to visualize.

cfg.VISUALIZED.LAYER_NAMES = [
  "backbone.bottom_up.res3.3.conv3",
  "backbone.bottom_up.res4.5.conv3",
  "backbone.bottom_up.res5.2.conv3",
  "backbone.fpn_output3",
  "backbone.fpn_output4",
  "backbone.fpn_output5",
  "backbone.top_block.p6",
  "backbone.top_block.p7",
  "head.cls_subnet.0",
  "head.cls_subnet.2",
  "head.cls_subnet.4",
  "head.cls_subnet.6",
  "head.bbox_subnet.0",
  "head.bbox_subnet.2",
  "head.bbox_subnet.4",
  "head.bbox_subnet.6"
] # example for retinanet

# how to get cfg.VISUALIZED.LAYER_NAMES
# model = build_model(cfg)
# for n, module in model.named_modules():
#    print(n)

Note:

This module is experimental!

It is modified and different from https://github.com/yizt/Grad-CAM.pytorch

It generates heatmaps with loss but Grad-Cam.pytorch generates heatmaps with score.