create fixnoise code

[andrevitorelli]

Creates fixnoise code. I am not sure if this is the correct way, so I summon @EiffL and/or @aguinot to review. This is really just a short few lines of code. I get the mean and std of the images, create noise images based on them, shear just the same, and then I rotate them by 90 degrees before adding to the main image.

[aguinot]

I am not sure about your implementation. Here are a few things that are important:

• Your implementation does not include psf deconvolution/reconvolution
• From my understanding, you are measuring the noise on sheared images which is bad for 2 reasons:
  • Your input noise image will be different for all sheared version and that is a big problem!
  • You are not using the true noise but the deconv/reconv image that is also an issue because the noise on this image is already correlated.
• Since you are using randoms here, one of your input should be a seed or a random generator for reproducibility and doing tests consistently. (I prefer personally prefer random generator. It easier to handle mass processing).

I would encourage you to look at the implementation in ngmix which is actually even simpler than yours. Basically you leave your method generate_mcal_image as it is (without the fixnoise addition). then you do the following:

def get_fixnoise_mcal(img, psf, rng):
    """
     Get mcal image using fixnoise correction.
     param:
         img: galaxy stamp
         psf: corresponding psf
         rng: either seed or random generator
     output:
         mcal_images
     """
     mcal_img = generate_mcal_image(img, psf)

     # handle the noise
     noise_img = get_noise(img, rng)
     input_noise = rot90(noise_img)
     mcal_noise = generate_mcal_image(input_noise, psf)
     # rotate back the noise
     final_noise = rot90(mcal_noise, k=-1)

     # correction
     final_result = mcal_img + final_noise

     return final_result

You can have a look at their implementation here

Also, given the autodiff framework I am not sure if what I proposed above work as it is or if it require some changes. I guess generating randoms within the autodiff part is not ideal so you will probably need to that outside the function (generating the noise image). Which make sense since in a more realistic context the noise will be provided by the step before the metacal.

[aguinot]

Apart for the comment I left, the rest looks good to me

[andrevitorelli]

Closed for reorganization of development.