#2 [Marimo Interface] Enhancement toward project completion

[HugoHakem]

Marimo interface

To come up to biological interpretation we need to be able to efficiently visualize each of the method
Why? There is a large amount of data.

App description:

1. UMAP plot of real image and counterfactual image.
   i. embedding would be generated using the classifier.
   ii. option to add 2 UMAP plots to visualize on the same raw different colorization of dots (for instance real vs fake AND class)

2. When selecting in the UMAP plot:
   i. plot real image with their corresponding counterfactual.
   ii. option to choose which channel to plot.
   iii. option to choose which counterfactual to plot: (for class 0, either plot class 1 2 or 3 etc.)
   vi. option to visualise different counterfactual (since starganv2 generate different output, why not visualizing different outputs.

3. For each pair of real and counterfactual plot:
   i. show the corresponding attribution map that leads to the smallest minimal mask.
   ii. show the minimal mask with a sliding bar allowing to either increase from that minimal mask or either reduce its size.

4. Other colorization of UMAP:
   i. coloration in term of size of the smallest minimal mask -> identify failure mode
   ii. coloration in term of attribution techniques that leads to the smallest minimal mask.

Things to be aware:
When building the app, it is way faster to have already every data stored somewhere.
The problem is that if we want to plot masking scenario with a sliding bar allowing you to see the mask for different mask size:

• Either we have to recompute every time (takes time)
• Either we have to store the data somewhere (takes space).
  Which option should be preferred?

[HugoHakem]

Masking scenario: store every mask from -5% before the minimum mask and 5% after the minimal mask. (limit to a certain number of fix mask like 10 for exemple)