Closed nperraud closed 5 years ago
hmm... not sure what do you mean by "right insights"?
I meant (I wrote this to @nperraud): how would you be able to tell if the model really understood the data and did not merely exploit some artifacts?
I think the healpix maps we make should be fairly artefact free.. they contain only projected particle count from N-body simulation. Other applications may need to investigate this.
Let me try to rephrase the question. What kind of visualization would be useful for physicists when they work with scnn? We are trying to see what would make the package more useful.
I think the 1D filter profile would be very nice! (as you emailed). Hopefully we will see different size and non-Gaussian profiles.
Here are the 5 filters learned in the first layer of the whole_sphere experiment. The first plot is the filters in the spectral domain, the second are the 1D filter profiles.
The learned filters look quite similar across experiments, you can find a lot of them in the results branch. There is one folder per experiment, and for each layer you find:
layer?_coefficients.png
: the raw coefficients learned by the neural network, that is the coefficients of the Chebyshev polynomials.layer?_spectral.png
: the filters represented in the spectral domain, that is the evaluated Chebyshev polynomicalslayer?_section.png
: the 1D filter profile, that is the filters localized at the equator.@tomaszkacprzak can you interpret anything out of those?
@mdeff: I think we can close this issue.
Tomek, what do the physicists would like to see to be able to see in order to tell if the model picked up the "right insights"?
This might not only improve the paper, but may make the package used by more people.