Hi Ian, Ryan, and Jeff,
This is a follow-up to our meeting today. We discussed the idea of adding parametric models of solar system bodies to operate in series with the rest of the non-parametric neural net. The goal (in my opinion) is to use the model as prior knowledge about the structure of the source, as well as to retrieve the (physically meaningful) parameters of the parametric model from the data.
I'm copy-pasting a very simple code to adjust the free parameters of a limb-darkened disk model, along with the code I used to make this into a PyTorch tensor at the same resolution as the data. You'll also need the file "uranus_mu_2021-09-30.npy", which is a 2-D array of emission angles on the Uranian disk at super-resolution and so is rather large (650 MB) - I have staged this for you on Google Drive for now. Is the dependence of the model on this array, or one like it, an issue? If so, we can discuss how to best implement this - for example, I could back up a step and send you the code I use to produce the mu values on a given pixel grid.
I am very happy to provide any additional information you might need, e.g., reasonable parameter ranges/priors, more complex models, or a restructure of this model into a different format.
I look forward to continuing to collaborate on this!
def adjust_lddisk(mu, a, flux, shift=None):
'''
limb-darkened disk model to update each step (?)
parameters
------
mu: 2-D numpy array with emission angle values of planet.
a: limb darkening parameter (exponential model)
flux: planet flux at mu=1.0
shift: tuple (xshift, yshift) to change planet center, default None
shift is (vertical, horizontal) for matplotlib with origin = 'lower'
'''
model = lddisk_model(mu, a, flux, law = 'exp')
if shift is not None:
model = ndimage.shift(model, shift)
model[model < flux/1e4] = 0.0 # just make very small vals zero
return model
import the model
modelfname = 'data/uranus_mu_2021-09-30.npy' #has pixel scale 1 mas
theta0 = [0.21, 63.7, 0.0, 0.0] # [ld, flux, xshift, yshift] from 1-D optimizer of limb darkened model
mu = np.load(modelfname)
zoom the model to desired resolution - note that zoom hates NaNs
zfactor = 0.001 / coords.cell_size
mu[np.isnan(mu)] = 0.0
mu = ndimage.zoom(mu, zfactor)
mu[mu < 1e-3] = np.nan #account for zoom hating NaNs by just setting very small mu back to NaN
ldmodel = adjust_lddisk(mu, theta0[0], theta0[1], shift=(theta0[2], theta0[3]))
pad_width = int((npix - ldmodel.shape[0])/2)
ldmodel = np.pad(ldmodel, pad_width)
put ldmodel into starting model
ldmodel = np.fft.fftshift(ldmodel) # need this because base_cube gets automatically fftshifted
ldmodel = np.expand_dims(ldmodel, 0) # basecube needs to be (nchans, nx, ny)
startmodel = torch.tensor(ldmodel, dtype=torch.float64, device=device)
Hi Ian, Ryan, and Jeff,
This is a follow-up to our meeting today. We discussed the idea of adding parametric models of solar system bodies to operate in series with the rest of the non-parametric neural net. The goal (in my opinion) is to use the model as prior knowledge about the structure of the source, as well as to retrieve the (physically meaningful) parameters of the parametric model from the data.
I'm copy-pasting a very simple code to adjust the free parameters of a limb-darkened disk model, along with the code I used to make this into a PyTorch tensor at the same resolution as the data. You'll also need the file "uranus_mu_2021-09-30.npy", which is a 2-D array of emission angles on the Uranian disk at super-resolution and so is rather large (650 MB) - I have staged this for you on Google Drive for now. Is the dependence of the model on this array, or one like it, an issue? If so, we can discuss how to best implement this - for example, I could back up a step and send you the code I use to produce the mu values on a given pixel grid.
I am very happy to provide any additional information you might need, e.g., reasonable parameter ranges/priors, more complex models, or a restructure of this model into a different format.
I look forward to continuing to collaborate on this!
def adjust_lddisk(mu, a, flux, shift=None):
'''
limb-darkened disk model to update each step (?)
import the model
modelfname = 'data/uranus_mu_2021-09-30.npy' #has pixel scale 1 mas
theta0 = [0.21, 63.7, 0.0, 0.0] # [ld, flux, xshift, yshift] from 1-D optimizer of limb darkened model
mu = np.load(modelfname)
zoom the model to desired resolution - note that zoom hates NaNs
zfactor = 0.001 / coords.cell_size
mu[np.isnan(mu)] = 0.0
mu = ndimage.zoom(mu, zfactor)
mu[mu < 1e-3] = np.nan #account for zoom hating NaNs by just setting very small mu back to NaN
ldmodel = adjust_lddisk(mu, theta0[0], theta0[1], shift=(theta0[2], theta0[3]))
pad_width = int((npix - ldmodel.shape[0])/2)
ldmodel = np.pad(ldmodel, pad_width)
put ldmodel into starting model
ldmodel = np.fft.fftshift(ldmodel) # need this because base_cube gets automatically fftshifted
ldmodel = np.expand_dims(ldmodel, 0) # basecube needs to be (nchans, nx, ny)
startmodel = torch.tensor(ldmodel, dtype=torch.float64, device=device)