Image gain restoration

[ysBach]

I remember previously astroscrappy.detect_cosmics returned the cleaned array multiplied by gain factor, but now it does not. (I could not find CHANGELOG.. )

astroscrappy.detect_cosmics(np.ones((100,100)), gain=3)[1]

array([[1., 1., 1., ..., 1., 1., 1.],
       [1., 1., 1., ..., 1., 1., 1.],
       [1., 1., 1., ..., 1., 1., 1.],
       ...,
       [1., 1., 1., ..., 1., 1., 1.],
       [1., 1., 1., ..., 1., 1., 1.],
       [1., 1., 1., ..., 1., 1., 1.]], dtype=float32)

It used to be

array([[3., 3., 3., ..., 3., 3., 3.],
       [3., 3., 3., ..., 3., 3., 3.],
       [3., 3., 3., ..., 3., 3., 3.],
       ...,
       [3., 3., 3., ..., 3., 3., 3.],
       [3., 3., 3., ..., 3., 3., 3.],
       [3., 3., 3., ..., 3., 3., 3.]], dtype=float32)

if my memory is correct, and this is why I used my custom function to do nccd.data = cleanarr / gain.

Now I find my processed images have extremely small counts & SNR (as I have gain ~ 10).

How can I write a code which incorporates both versions of astroscrappy within a single function?
(e.g., for inbkg, bkg, pssl stuff, I could use try-except clause)

[ysBach]

I found a simple code below may work:

divfactor = detect_cosmics(np.ones((3, 3)), gain=1., niter=0)[1][0, 0]
if divfactor != 1.:
    _ccd.data = cleanarr / divfactor

but any better idea?

A digression:
Interestingly, it takes only ~ 40 us on MBP 15", but this becomes ~30x slower (1.2 ms) on MBP 16".
(tested using %timeit detect_cosmics(np.ones((3, 3)), gain=1., niter=0)[1][0, 0] == 1., timings does not change much w.r.t. gain, shape, etc.).
Maybe because I am running python on Rosetta in the latter machine..? Still factor 30 is strange.

MBP 15": [2018, macOS 11.6, i7-8850H (2.6 GHz; 6-core), RAM 16 GB (2400MHz DDR4), Radeon Pro 560X (4GB)]
MBP 16": [2021, macOS 12.0.1, M1Pro, 8P+2E core, GPU 16-core, RAM 16GB]

[ysBach]

I am still using the above code.

BTW, trying on mini-forge, I found this:

1. the above test was done on default anaconda (intel, running on Rosetta2, ~ 1200 µs)
2. the code is still slow in default mini-forge installation (apple arm64):

%timeit divfactor = detect_cosmics(np.ones((3, 3)), gain=1., niter=0)[1][0, 0]
643 µs ± 29 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

3. but from mini-forge (apple arm64) with veclib (see here; install numpy as conda install numpy "blas=*=*accelerate*"), it is much faster (x40+)

In [4]: %timeit divfactor = detect_cosmics(np.ones((3, 3)), gain=1., niter=0)[1][0, 0]
14.5 µs ± 59 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

I think this speedup is not due to np.ones or [1](0, 0) (they reduced the %timeit by only < 1 µs and < 100 µs order, respectively).