Open iancze opened 5 years ago
As a first-pass with Theano, let's implement a Fourier model, but without any sampling. This means that our dataset has the exact same (u,v) sampling as the FFT. Hopefully this means we wouldn't need to implement any custom operations.
Now that we've written up the math required to get from a generic sky model to visibilities, we should start down the path of setting up a model to compute visibilities.
Probably a good place to start is with a very simple analytic model in the sky plane (either a point source w/ some position offset, or a Gaussian, etc) and then try to run this through Theano and estimate visibilities and gradients.
This would probably be very helpful for the dust-modeling people, because we could even do this for a Tripathi-like Nuker profile, perhaps.
After that I think most of the hard work is done and we could even release something for the dust modeling folks.