RISE returns different heatmap for same input

[bartmch]

Hey, after making your model torch>1.0 compatible I got the RISE "explainer" predictions for the covid-19 file "4-x-day1.jpg" twice. The predictions are the same:

but the heatmaps produced by RISE look totally different:

Does the same happen to your torch<1.0 model? If not, do you know what could be the issue?

[arpanmangal]

@SuryaKalia could you please look into this?

[SuryaKalia]

That behaviour is due to the randomly generated masks in the visualize function. Each call generates a new set of masks while the generate_new flag is true. Some solutions are:

1. Set generate_new = False at line 36, visualize.py - This will use the previously saved masks.npy
2. Set a random seed - That's redundant repetition of generating identical random masks over and over again.
3. Increase the number of masks generated. This is currently set as N=1000 at line 39, visualize.py. Having a large number of random masks would generate more consistent results due to the large sampling, though at the cost of compute time.

[bartmch]

Ok will do that now, quick side question: have you also looked into grad-cam visualization techniques?

[bartmch]

On the first solution ("Set generate_new = False at line 36, visualize.py - This will use the previously saved masks.npy"), I'm getting the following error on line 59 of "explanations.py"

[bartmch]

On the third solution ("Increase the number of masks generated"). I'm still getting totally different heatmaps. I changed from
explainer.generate_masks(N=1000, s=8, p1=0.1, savepath=maskspath)
to
explainer.generate_masks(N=10000, s=8, p1=0.1, savepath=maskspath)
and got the following heatmaps on the same data:

[bartmch]

Hi! Any updates on this?

[SuryaKalia]

Hi, thanks for pointing out. I've fixed that issue. The first solution should work now.

The 3rd solution - increasing the number of masks is still bound to give different visualizations. This is an inherent limitation of the RISE visualization method.

Grad-CAM is a good alternative. However, that had an issue with images which had very high predictive confidence for a given class, causing gradient saturation. To eliminate those issues we used RISE visualizations.