Add GradCAM analyzer

[JeanAnNess]

Add GradCAM analyzer

The following has been changed:

• Added struct GradCAM and function analyzer for the type in gradient.jl
• Added export GradCAM to ExplainableAI.jl
• Added Dictionary entry for GradCAM in heatmap.jl (Couldn't find the file in the repository anymore - Where is it?)

Papers / Blogs as reference

• XAI Methods - Guided GradCAM
• Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization

To-Do

• Potentially make the PR able to be automatically mergeable (?)
• Expand on visualisation function
  • Upscaling the 7x7 output and layering it over the input image (improve interpolation in example below)
  • Adding a function for it somewhere (TBD)
• Validate the choosen colorscheme

Example

# Imports
using ExplainableAI
using Flux
using Metalhead                         # pre-trained vision models
using HTTP, FileIO, ImageMagick         # load image from URL
using ImageInTerminal                   # show heatmap in terminal
using ImageTransformations: imresize    # resize of heatmap

# Load model
model = VGG(16, pretrain=true).layers
# Load input
url = HTTP.URI("https://raw.githubusercontent.com/Julia-XAI/ExplainableAI.jl/gh-pages/assets/heatmaps/castle.jpg")
img = load(url)
input = preprocess_imagenet(img)
input = reshape(input, 224, 224, 3, :)  # reshape to WHCN format

# Run XAI method
analyzer = GradCAM(model.layers[1], model.layers[2])    
expl = analyze(input, analyzer)         

# Show heatmap
output = heatmap(expl)

# Show overlayed heatmap with input
alpha = 0.6
hmap_scaled = imresize(output, size(img))
combined = img*(1-alpha) + hmap_scaled * alpha

Input	7x7 GradCAM output	Upscaled output	Upscaled and overlayed

[adrhill]

Awesome, thanks a lot!

I'm currently restructuring the package for a 1.0 release, moving the interface into XAIBase.jl, heat mapping into VisionHeatmaps.jl and LRP into RelevancePropagation.jl, so that's why src/heatmap.jl disappeared.

I'll take care of the rest!

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (75954c8) 95.89% compared to head (3613341) 96.29%.
Report is 1 commits behind head on master.

❗ Current head 3613341 differs from pull request most recent head 26ea235. Consider uploading reports for the commit 26ea235 to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #155      +/-   ##
==========================================
+ Coverage   95.89%   96.29%   +0.40%     
==========================================
  Files           4        5       +1     
  Lines          73       81       +8     
==========================================
+ Hits           70       78       +8     
  Misses          3        3              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[adrhill]

Awesome, looks great! 👍

I'm working on heatmap overlays in VisionHeatmaps, which should be merged soon (Julia-XAI/VisionHeatmaps.jl#4), so for now, calling heatmap on GradCAM will return the "7x7 GradCAM output" from your table above.