IgnacioHeredia
commented
2 years ago Update
Training on a cleaner/smaller dataset with a MobileNet does not seem to produce patterns. So it probably (?) had to do with the noisiness of the dataset. Still no intuition on why a noisy dataset would be more likely to produce these patterns.
Feel free to close the issue, though still interested to hear if someone has ever experienced something similar.
Hi all,
I had a question. When I use
Fullgrad(or its smooth version) I experience very strong checkerboard patterns. I'm aware this is a fairly well know issue with gradient methods, but still curious because:GradCamimplementation (from this very same repo) does not show themtorchvision.models. No fancy stuff.I'm skeptical of opening an issue because the patterns only appear in the model after training. The model with weights pretrained on ImageNet doesn't show these patterns. Therefore it might very well be unrelated to Fullgrad, so feel free to close the issue.
But I am still curious if someone else has had this problem, in case it could be a side effect of the interpolation or whatnot.
Model after training
FullGrad
One can start to sense the patterns.
Smooth FullGrad
Averaging over the default 100 samples really magnifies the patterns.
Of course, this behavior repeats for all images. Here's another example:
GradCam
No checkerboard, although one sees these reticular-like explanations. They also seem to appear in other implementations thus are probably a distinctive feature of GradCam.
Model only with ImageNet pretraining
Smooth FullGrad
No patterns.
GradCam
No patterns, still kind of reticular.
So, just sharing in case anyone had any thoughts. Pointers are welcome, thanks! :)