Closed mjvakili closed 5 years ago
Hi,
I was currently trying it with two parameters (Omega_c and sigma8), but going down to 1 first is a better start. For the two parameter case I had delta_theta = [0.02,0.02], but the differences between the upper and lower derivatives are very low (note the y-axes labels):
Which results in loss curves that just go randomly up and down:
So I will have to think about how to normalize/standardize the Cl's.
Are these the noisy Cl's? Can you try a larger delta_theta and one parameter (omega_cdm) perhaps?
Yes this is from the noisy Cl's.
I don't really have a feeling for how large delta_theta should be for omega_cdm, do you mean an order of magnitude larger?
yeah perhaps something of order 10^-1! You can take a look at fig.12 of this paper: https://arxiv.org/pdf/1607.01761.pdf You might get a feeling of how large delta_theta should be by looking at Delta Cl/Cl given a logarithmic change in omega_m and the errorbars
I implemented the 1 parameter version now, with delta_theta = 0.11 the difference between the upper/lower simulations is still very small:
Taking the logarithm of the Cl's doesn't seem to help the network learn either. It is interesting that the loss curves are almost the same whether I take the log or not. I think it will the answer will lie in the standardization of the data/derivatives.
Can you plot (delta Cl)/Cl (only the theoretical Cl with no noise) as a function of l? In the same plot, can you show (sigma Cl /Cl) as a function of l? delta Cl = Cl(0.28) - Cl(0.26) ; sigma Cl = (var Cl)**0.5
The issue with the upper/lower derivatives is now resolved. It was a problem with not correctly rescaling the data and not with the numerical derivative calculation. Closing this issue here.
Hi Erik,
Can you plot Cl for different values of omega_cdm? This will help you get a sense of how Cl's are sensitive to varying omega_cdm. Also as a start, can you just estimate one parameter (omega_cdm) with IMNN? You can keep the rest of the parameters fixed!