Noise handling + New outputs

[aguinot]

Note 1: Should be merged after #67 Note 2: This should fix issue #54

@matroxel if you have some time, I would be very happy to have your opinion on how I implemented the noise using custom features of galsim.config.

Noise handling

I "fixed" the handling the handling of the noise in _euclidlikeimsim. Originally the noise was added through these lines:

poisson_noise = galsim.noise.PoissonNoise(rng)
if self.draw_method == "phot":
    logger.debug("Adding poisson noise to sky photons")
    sky_image1 = sky_image.copy()
    sky_image1.addNoise(poisson_noise)
    image.quantize()  # In case any profiles used InterpolatedImage, in which case
    # the image won't necessarily be integers.
    image += sky_image1
else:
    logger.debug("Adding poisson noise")
    image += sky_image
    image.addNoise(poisson_noise)

The default drawing method is "auto" so it will do the else statement which is a problem because the majority of our objects are drawn using the phot method so the poisson noise will be counted twice (this is the case on all the images I have done so far).
With the new implementation, the noise is handled independently of the drawing method. That raise a new issue, the bright objects are drawn with the fft method so there will be no poisson noise on them. If we believe this is important I have some ideas how to fix that.

New outputs

With the new implementation of the noise I was able to make new ouptuts:

• Noise image: This will output the exact noise on the image including the sky background (except if the sky_subtract==True in that case it is also subtracted on the noise)
• Sky image: This will output the sky background. I want to emphasize on background, it is NOT a realization of the sky background including noise. I insist because it is different than the output of _romanimsim. This is just a personal preference. Let me know if you want me to change it.
• Weight image: This is an inverse variance map. The variance is measured from the noise image with numpy.var() because it might not possible to get it analytically. (Is it possible to get it from GalSim?)

The new outputs are done with custom galsim.config.ExtraOutputBuilder. I can easily add more or modify the existing one. I did not thought useful to output a noiseless image since you can get it from the noise image. If you think it could be useful to have one, I can add it.

Tests

Here is a list of tests I have done:

• The noise is only computed once and reused where ever it is needed
• Requesting noise_image, sky_image, weight_image when the noise is not requested:
  • Raise an error
  • The error is raised as early as possible, before the full image is computed
• Requesting sky_image when the background is subtracted:
  • Raise an error
  • This on is debatable if it should be an error or not. If the background is subtracted the sky_image is by definition 0
• The new computation of the noise give the same results as before (except for the boject poisson noise issue mention above)

All feedback is very welcome 🙂.

[matroxel]

Noise is already added to the fft-drawn objects at this line: https://github.com/GalSim-developers/GalSim-Euclid-Like/blob/cce5f516a1926ef520fde13bf89a0606d76a6b3e/euclidlike_imsim/stamp.py#L285

So you should be able to always treat the image like the self.draw_method == "phot" block quoted in your post.

[matroxel]

I think your changes will work also.

[aguinot]

Noise is already added to the fft-drawn objects at this line:

https://github.com/GalSim-developers/GalSim-Euclid-Like/blob/cce5f516a1926ef520fde13bf89a0606d76a6b3e/euclidlike_imsim/stamp.py#L285

So you should be able to always treat the image like the self.draw_method == "phot" block quoted in your post.

Thank you for pointing that, I did not look carefully enough to the code. Thank you for having a look at the changes!

[aguinot]

@rmandelb the conflicts have been solved. It is ready for review.