tickets/DM-32436: implement scarlet lite #47
Conversation
`modelToHeavy` was modified to remove an incorrect thresholding algorithm used to truncate convolved scarlet models and replace it with an algorithm that grows the deconvolved footprint with the rectangular PSF.
Deblended children near the edge of an image can have their footprints dilated outside of the exposure by the PSF width. This commit ensures that the dilated model is always contained fully in the exposure. This is valid, as later in pipe_tasks the science pipelines flag edge sources so it isn't expected that all of their flux is contained in the image.
fred3m
force-pushed
the
tickets/DM-32436
branch
from
fc41bce
to
b2d34a5
Compare
| @@ -771,17 +964,15 @@ def deblend(self, mExposure, catalog): | |||
| try: | |||
| t0 = time.time() | |||
There was a problem hiding this comment.
The internet suggests time.monotonic or time.perf_counter, in case of system time changes (doubt that comes up much, but anyway).
| @@ -807,7 +998,10 @@ def deblend(self, mExposure, catalog): | |||
| continue | |||
|
|
|||
| # Update the parent record with the deblending results | |||
| logL = blend.loss[-1]-blend.observations[0].log_norm | |||
| if self.config.version == "scarlet": | |||
| logL = -blend.loss[-1]+blend.observations[0].log_norm | |||
There was a problem hiding this comment.
There should be spaces around the plus: https://developer.lsst.io/python/style.html#binary-operators-should-be-surrounded-by-a-single-space-except-for
(that goes for other lines in this file too)
| """Convert a scarlet model to a `MultibandFootprint`. | ||
| Parameters | ||
| ---------- | ||
| source : `scarlet.Component` |
| @@ -146,3 +146,78 @@ def modelToHeavy(source, mExposure, blend, xy0=Point2I(), dtype=np.float32): | |||
| model = MultibandImage(mExposure.filters, model, valid.getBBox()) | |||
| mHeavy = MultibandFootprint.fromImages(mExposure.filters, model, footprint=foot) | |||
| return mHeavy | |||
|
|
|||
|
|
|||
| def liteModelToHeavy(source, mExposure, blend, xy0=Point2I(), dtype=np.float32, useFlux=False): | |||
There was a problem hiding this comment.
Even if you don't mutate xy0 currently, this should still be xy0=None and conditionally initialized to Point2I() below (https://developer.lsst.io/python/style.html#a-mutable-object-must-not-be-used-as-a-keyword-argument-default). Also, if you don't actually need a Point2I, you could just pass x0 and y0 floats.
There was a problem hiding this comment.
Isn't Point2I immutable?
There was a problem hiding this comment.
Oh, apparently not. Yes, I'll fix this.
| origin = source_box.origin | ||
| else: | ||
| source_box = source.bbox | ||
| shape = (source_box.shape[0], source_box.shape[1] + py, source_box.shape[2] + px) |
There was a problem hiding this comment.
Don't you only need to grow the box by 2*dh, 2*dw instead of py, px (if the psf shape is odd)?
There was a problem hiding this comment.
Good catch, I think that you're right.
fred3m
force-pushed
the
tickets/DM-32436
branch
2 times, most recently
from
4ef56e3
to
eeb47e8
Compare
|
Ok, I think that I addressed everything now. Once approved I'll run Jenkins before merging. |
fred3m
force-pushed
the
tickets/DM-32436
branch
from
eeb47e8
to
cbb4b98
Compare
No description provided.