Coverage for tests/test_models.py: 10%
294 statements
« prev ^ index » next coverage.py v7.14.1, created at 2026-05-30 08:25 +0000
1# This file is part of lsst.scarlet.lite.
2#
3# Developed for the LSST Data Management System.
4# This product includes software developed by the LSST Project
5# (https://www.lsst.org).
6# See the COPYRIGHT file at the top-level directory of this distribution
7# for details of code ownership.
8#
9# This program is free software: you can redistribute it and/or modify
10# it under the terms of the GNU General Public License as published by
11# the Free Software Foundation, either version 3 of the License, or
12# (at your option) any later version.
13#
14# This program is distributed in the hope that it will be useful,
15# but WITHOUT ANY WARRANTY; without even the implied warranty of
16# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17# GNU General Public License for more details.
18#
19# You should have received a copy of the GNU General Public License
20# along with this program. If not, see <https://www.gnu.org/licenses/>.
22import os
23from functools import partial
24from typing import cast
26import lsst.scarlet.lite.models as models
27import numpy as np
28from lsst.scarlet.lite import Blend, Box, FistaParameter, Image, Observation, Source
29from lsst.scarlet.lite.component import Component, FactorizedComponent, default_adaprox_parameterization
30from lsst.scarlet.lite.initialization import FactorizedInitialization
31from lsst.scarlet.lite.models import (
32 CartesianFrame,
33 EllipseFrame,
34 EllipticalParametricComponent,
35 FactorizedFreeFormComponent,
36 FittedPsfBlend,
37 FittedPsfObservation,
38 ParametricComponent,
39)
40from lsst.scarlet.lite.operators import Monotonicity
41from lsst.scarlet.lite.parameters import AdaproxParameter, parameter, relative_step
42from lsst.scarlet.lite.utils import integrated_circular_gaussian
43from numpy.testing import assert_array_equal
44from scipy.stats import gamma as gamma_dist
45from utils import ScarletTestCase
48def parameterize(component: Component):
49 assert isinstance(component, ParametricComponent)
50 component._spectrum = AdaproxParameter(
51 component.spectrum,
52 step=partial(relative_step, factor=1e-2, minimum=1e-16),
53 )
54 component._params = AdaproxParameter(
55 component._params.x,
56 step=1e-2,
57 )
60class TestFreeForm(ScarletTestCase):
61 def setUp(self) -> None:
62 yx0 = (1000, 2000)
63 filename = os.path.join(__file__, "..", "..", "data", "hsc_cosmos_35.npz")
64 filename = os.path.abspath(filename)
65 data = np.load(filename)
66 self.data = data
67 model_psf = integrated_circular_gaussian(sigma=0.8)
68 self.detect = np.sum(data["images"], axis=0)
69 self.centers = np.array([data["catalog"]["y"], data["catalog"]["x"]]).T + np.array(yx0)
70 bands = data["filters"]
71 self.observation = Observation(
72 Image(data["images"], bands=bands, yx0=yx0),
73 Image(data["variance"], bands=bands, yx0=yx0),
74 Image(1 / data["variance"], bands=bands, yx0=yx0),
75 data["psfs"],
76 model_psf[None],
77 bands=bands,
78 )
80 def tearDown(self):
81 del self.data
83 def test_free_form_component(self):
84 images = self.data["images"]
86 # Test with no thresholding (sparsity)
87 sources = []
88 for i in range(5):
89 component = FactorizedFreeFormComponent(
90 self.observation.bands,
91 np.ones(5),
92 images[i].copy(),
93 self.observation.bbox,
94 )
95 sources.append(Source([component]))
97 blend = Blend(sources, self.observation).fit_spectra()
98 blend.parameterize(default_adaprox_parameterization)
99 blend.fit(12, e_rel=1e-6)
101 # Test with thresholding (sparsity)
102 sources = []
103 for i in range(5):
104 component = FactorizedFreeFormComponent(
105 self.observation.bands,
106 np.ones(5),
107 images[i].copy(),
108 self.observation.bbox,
109 bg_rms=self.observation.noise_rms,
110 bg_thresh=0.25,
111 min_area=4,
112 )
113 sources.append(Source([component]))
115 blend = Blend(sources, self.observation).fit_spectra()
116 blend.parameterize(default_adaprox_parameterization)
117 blend.fit(12, e_rel=1e-6)
119 # Test with peak centers specified
120 sources = []
121 peaks = list(np.array([self.data["catalog"]["y"], self.data["catalog"]["x"]]).T.astype(int))
122 for i in range(5):
123 component = FactorizedFreeFormComponent(
124 self.observation.bands,
125 np.ones(5),
126 images[i].copy(),
127 self.observation.bbox,
128 peaks=peaks,
129 )
130 sources.append(Source([component]))
132 blend = Blend(sources, self.observation).fit_spectra()
133 blend.parameterize(default_adaprox_parameterization)
134 blend.fit(12, e_rel=1e-6)
136 # Tests for code blocks that are difficult to reach,
137 # to complete test coverage
138 component = blend.sources[-1].components[0]
139 self.assertFalse(component.resize(self.observation.bbox))
140 component.morph[:] = 0
141 component.prox_morph(component.morph)
144class TestParametric(ScarletTestCase):
145 def setUp(self) -> None:
146 filename = os.path.join(__file__, "..", "..", "data", "hsc_cosmos_35.npz")
147 filename = os.path.abspath(filename)
148 data = np.load(filename)
149 self.data = data
150 self.model_psf = integrated_circular_gaussian(sigma=0.8)
151 self.detect = np.sum(data["images"], axis=0)
152 self.centers = np.array([data["catalog"]["y"], data["catalog"]["x"]]).T
153 bands = data["filters"]
154 self.observation = Observation(
155 Image(data["images"], bands=bands),
156 Image(data["variance"], bands=bands),
157 Image(1 / data["variance"], bands=bands),
158 data["psfs"],
159 self.model_psf[None],
160 bands=bands,
161 )
163 def tearDown(self):
164 del self.data
166 def test_cartesian_frame(self):
167 bbox = Box((31, 60), (1, 2))
168 frame = CartesianFrame(bbox)
169 y = np.linspace(1, 31, 31)
170 x = np.linspace(2, 61, 60)
171 x, y = np.meshgrid(x, y)
173 self.assertBoxEqual(frame.bbox, bbox)
174 self.assertTupleEqual(frame.shape, bbox.shape)
175 assert_array_equal(frame.x_grid, x)
176 assert_array_equal(frame.y_grid, y)
177 self.assertIsNone(frame._r2)
178 self.assertIsNone(frame._r)
180 def test_ellipse_frame(self):
181 y0 = 23
182 x0 = 36
183 major = 3
184 minor = 2
185 theta = 2 * np.pi / 3
186 bbox = Box((50, 39), (2, 7))
187 r_min = 1e-20
188 frame = EllipseFrame(
189 y0,
190 x0,
191 major,
192 minor,
193 theta,
194 bbox,
195 )
197 self.assertEqual(frame.x0, x0)
198 self.assertEqual(frame.y0, y0)
199 self.assertEqual(frame.major, major)
200 self.assertEqual(frame.minor, minor)
201 self.assertEqual(frame.theta, theta)
202 self.assertEqual(frame.bbox, bbox)
204 y = np.linspace(2, 51, 50)
205 x = np.linspace(7, 45, 39)
206 x, y = np.meshgrid(x, y)
207 r2 = frame._xa**2 + frame._yb**2 + r_min**2
208 r = np.sqrt(r2)
210 self.assertEqual(frame._sin, np.sin(theta))
211 self.assertEqual(frame._cos, np.cos(theta))
212 self.assertBoxEqual(frame.bbox, bbox)
213 self.assertTupleEqual(frame.shape, bbox.shape)
214 assert_array_equal(frame.x_grid, x)
215 assert_array_equal(frame.y_grid, y)
216 assert_array_equal(frame.r2_grid, r2)
217 assert_array_equal(frame.r_grid, r)
219 # There doesn't seem to be much utility to testing the
220 # gradients exactly, however we test that they all run
221 # properly.
222 gradients = (
223 frame.grad_x0,
224 frame.grad_y0,
225 frame.grad_major,
226 frame.grad_minor,
227 frame.grad_theta,
228 )
230 input_grad = np.ones(bbox.shape, dtype=float)
231 original_grad = input_grad.copy()
232 for grad in gradients:
233 grad(input_grad, False)
234 grad(input_grad, True)
236 # Make sure that none of the methods changed the input gradient
237 assert_array_equal(input_grad, original_grad)
239 def test_grad_sersic_n_index(self):
240 """Finite-difference check of the Sersic gradient w.r.t. n.
242 Protects against regressions in the analytical gradient (e.g. the
243 np.log10 vs np.log mistake fixed for audit finding C-1). The
244 analytical gradient holds bn(n) fixed -- differentiating the
245 inverse incomplete gamma function is intentionally omitted -- so
246 the finite-difference comparison does the same.
247 """
248 bbox = Box((33, 33), origin=(0, 0))
249 y0, x0 = 16.0, 16.0
250 sigma_y, sigma_x = 5.0, 4.0
251 theta = 0.3
252 spectrum = np.array([1.0])
254 def sersic_fixed_bn(n_value, bn_value, ellipse):
255 r = ellipse.r_grid
256 return np.exp(-bn_value * (r ** (1 / n_value) - 1))
258 for n in (0.5, 1.5, 2.0, 3.0, 4.0):
259 ellipse = EllipseFrame(y0, x0, sigma_y, sigma_x, theta, bbox)
260 bn = gamma_dist.ppf(0.5, 2 * n)
261 morph = sersic_fixed_bn(n, bn, ellipse)
262 input_grad = np.ones((1,) + morph.shape)
263 params = np.array([y0, x0, sigma_y, sigma_x, theta, n])
265 d_n_analytical = models.grad_sersic(input_grad, params, morph, spectrum, ellipse)[5]
267 # Centered finite difference: df/dn ~ (f(n+e) - f(n-e)) / (2e)
268 # where f(n) = sum(sersic_fixed_bn(n)). bn is held fixed in both
269 # evaluations to match the analytical formulation, which omits
270 # the dbn/dn contribution.
271 eps = 1e-5 * max(1.0, abs(n))
272 loss_plus = np.sum(sersic_fixed_bn(n + eps, bn, ellipse))
273 loss_minus = np.sum(sersic_fixed_bn(n - eps, bn, ellipse))
274 d_n_fd = (loss_plus - loss_minus) / (2 * eps)
276 np.testing.assert_allclose(
277 d_n_analytical,
278 d_n_fd,
279 rtol=1e-4,
280 err_msg=f"grad_sersic d/dn mismatch at n={n}",
281 )
283 def test_grad_circular_gaussian(self):
284 """Finite-difference check of the circular Gaussian gradient.
286 Protects against regressions in grad_circular_gaussian (audit
287 finding C-2). The forward model is morph = exp(-r2) where
288 r2 = ((x-x0)/(2*sigma))**2 + ((y-y0)/(2*sigma))**2, so the gradient
289 must scale with 1/(2*sigma**2). A previous version of the code used
290 a fixed factor of 2, ignoring sigma entirely.
291 """
292 bbox = Box((33, 33), origin=(0, 0))
293 y0, x0 = 16.3, 16.7
294 spectrum = np.array([1.0])
295 frame = CartesianFrame(bbox)
297 for sigma in (0.6, 0.8, 1.5, 3.0):
298 morph = models.circular_gaussian((y0, x0), frame, sigma=sigma)
299 input_grad = np.ones((1,) + morph.shape)
300 params = np.array([y0, x0])
302 d_analytical = models.grad_circular_gaussian(
303 input_grad, params, morph, spectrum, frame, sigma=sigma
304 )
306 # Centered finite difference on f(y0, x0) = sum(morph(y0, x0)).
307 eps = 1e-4
308 f_yp = np.sum(models.circular_gaussian((y0 + eps, x0), frame, sigma=sigma))
309 f_ym = np.sum(models.circular_gaussian((y0 - eps, x0), frame, sigma=sigma))
310 f_xp = np.sum(models.circular_gaussian((y0, x0 + eps), frame, sigma=sigma))
311 f_xm = np.sum(models.circular_gaussian((y0, x0 - eps), frame, sigma=sigma))
312 d_fd = np.array([(f_yp - f_ym) / (2 * eps), (f_xp - f_xm) / (2 * eps)])
314 np.testing.assert_allclose(
315 d_analytical,
316 d_fd,
317 rtol=1e-4,
318 atol=1e-8,
319 err_msg=f"grad_circular_gaussian mismatch at sigma={sigma}",
320 )
322 def test_grad_gaussian2(self):
323 """Finite-difference check of the elliptical 2D Gaussian
324 gradient.
326 ``grad_gaussian2`` returns d/d{y0, x0, sigma_y, sigma_x,
327 theta} of ``sum(spectrum * input_grad * morph)``. With
328 ``spectrum = [1]`` and ``input_grad = ones``, the loss is
329 just ``sum(morph)``.
331 Also exercises ``EllipseFrame.grad_major`` and
332 ``grad_minor``: the ``-2/major * _xa**2`` base in those
333 methods is already ``d(r**2)/d(major)``, while the
334 ``grad_x0/y0/theta`` base is ``(1/2) * d(r**2)/d(...)``.
335 Multiplying by 2 (``use_r2=True``) or by ``1/r``
336 (``use_r2=False``) was therefore producing 2x the correct
337 gradient for the size parameters.
338 """
339 bbox = Box((33, 33), origin=(0, 0))
340 spectrum = np.array([1.0])
341 # y0, x0, sigma_y, sigma_x, theta
342 y0, x0, sigma_y, sigma_x, theta = 16.3, 16.7, 5.0, 4.0, 0.3
343 empty_params = np.array([])
345 ellipse = EllipseFrame(y0, x0, sigma_y, sigma_x, theta, bbox)
346 morph = models.gaussian2d(empty_params, ellipse)
347 input_grad = np.ones((1,) + morph.shape)
348 params = np.array([y0, x0, sigma_y, sigma_x, theta])
349 d_analytical = models.grad_gaussian2(input_grad, params, morph, spectrum, ellipse)
351 # Smaller eps for theta because the morph is more sensitive
352 # near theta=0.3 (radians), and a too-large eps flattens the
353 # finite-difference signal.
354 eps_list = [1e-4, 1e-4, 1e-4, 1e-4, 1e-5]
355 d_fd = np.zeros(5)
356 for i, eps in enumerate(eps_list):
357 perturb = np.zeros(5)
358 perturb[i] = eps
359 yp, xp, syp, sxp, tp = params + perturb
360 ym, xm, sym, sxm, tm = params - perturb
361 f_plus = np.sum(models.gaussian2d(empty_params, EllipseFrame(yp, xp, syp, sxp, tp, bbox)))
362 f_minus = np.sum(models.gaussian2d(empty_params, EllipseFrame(ym, xm, sym, sxm, tm, bbox)))
363 d_fd[i] = (f_plus - f_minus) / (2 * eps)
365 np.testing.assert_allclose(d_analytical, d_fd, rtol=1e-3, atol=1e-7)
367 def test_grad_integrated_gaussian(self):
368 """Finite-difference check of the integrated Gaussian
369 gradient. ``grad_integrated_gaussian`` returns
370 d/d{y0, x0, sigma}.
371 """
372 bbox = Box((33, 33), origin=(0, 0))
373 spectrum = np.array([1.0])
374 frame = CartesianFrame(bbox)
375 params = np.array([16.3, 16.7, 1.5])
376 morph = models.integrated_gaussian(params, frame)
377 input_grad = np.ones((1,) + morph.shape)
379 d_analytical = models.grad_integrated_gaussian(input_grad, params, morph, spectrum, frame)
381 eps = 1e-4
382 d_fd = np.zeros(3)
383 for i in range(3):
384 perturb = np.zeros(3)
385 perturb[i] = eps
386 f_plus = np.sum(models.integrated_gaussian(params + perturb, frame))
387 f_minus = np.sum(models.integrated_gaussian(params - perturb, frame))
388 d_fd[i] = (f_plus - f_minus) / (2 * eps)
390 np.testing.assert_allclose(d_analytical, d_fd, rtol=1e-3, atol=1e-7)
392 def test_grad_sersic_ellipse_params(self):
393 """Finite-difference check of grad_sersic for the ellipse
394 parameters. Index 5 (d/dn) is covered by
395 ``test_grad_sersic_n_index``; this covers indices 0-4:
396 d/d{y0, x0, sigma_y, sigma_x, theta}. ``bn(n)`` does not
397 depend on these parameters, so the FD evaluation can use
398 ``models.sersic`` directly.
399 """
400 bbox = Box((33, 33), origin=(0, 0))
401 spectrum = np.array([1.0])
402 n = 1.5
403 n_params = np.array([n])
404 y0, x0, sigma_y, sigma_x, theta = 16.0, 16.0, 5.0, 4.0, 0.3
406 ellipse = EllipseFrame(y0, x0, sigma_y, sigma_x, theta, bbox)
407 morph = models.sersic(n_params, ellipse)
408 input_grad = np.ones((1,) + morph.shape)
409 params = np.array([y0, x0, sigma_y, sigma_x, theta, n])
410 d_analytical = models.grad_sersic(input_grad, params, morph, spectrum, ellipse)[:5]
412 eps_list = [1e-4, 1e-4, 1e-4, 1e-4, 1e-5]
413 d_fd = np.zeros(5)
414 ellipse_params = np.array([y0, x0, sigma_y, sigma_x, theta])
415 for i, eps in enumerate(eps_list):
416 perturb = np.zeros(5)
417 perturb[i] = eps
418 yp, xp, syp, sxp, tp = ellipse_params + perturb
419 ym, xm, sym, sxm, tm = ellipse_params - perturb
420 f_plus = np.sum(models.sersic(n_params, EllipseFrame(yp, xp, syp, sxp, tp, bbox)))
421 f_minus = np.sum(models.sersic(n_params, EllipseFrame(ym, xm, sym, sxm, tm, bbox)))
422 d_fd[i] = (f_plus - f_minus) / (2 * eps)
424 np.testing.assert_allclose(d_analytical, d_fd, rtol=1e-3, atol=1e-7)
426 def test_parametric_component(self):
427 observation = self.observation
428 bands = observation.bands
429 spectrum = np.ones((observation.n_bands,), dtype=float)
430 frame = CartesianFrame(observation.bbox)
432 # Test integrated Gaussian PSF and sersic
433 sources = []
434 for idx, center in enumerate(self.centers):
435 # Get the integer center of the source
436 cy, cx = int(np.round(center[0])), int(np.round(center[1]))
437 # For now we use a fixed bounding box that is the size of
438 # the observed PSF image
439 bbox = Box((41, 41), origin=(cy - 20, cx - 20)) & observation.bbox
440 # Keep track of the initial positions
441 yi, xi = cy, cx
442 # Restrict the values of the parameters
443 _proxmin = np.array([yi - 2, xi - 2, 1e-1, 1e-1, -np.pi / 2, 1.1])
444 _proxmax = np.array([yi + 2, xi + 2, frame.shape[-2] / 2, frame.shape[-1] / 2, np.pi / 2, 3])
446 __proxmin = np.array([yi - 2, xi - 2, 0.8])
447 __proxmax = np.array([yi + 2, xi + 2, 1.2])
449 # Initialize a PSF-like component of the source using a
450 # non-pixel integrated gaussian
451 component = ParametricComponent(
452 bands,
453 bbox,
454 spectrum=parameter(spectrum.copy()),
455 morph_params=parameter(np.array([center[0], center[1], 0.8])),
456 morph_func=models.integrated_gaussian,
457 morph_grad=models.grad_integrated_gaussian,
458 morph_prox=partial(models.bounded_prox, proxmin=__proxmin, proxmax=__proxmax),
459 morph_step=np.array([1e-2, 1e-2, 1e-2]),
460 prox_spectrum=lambda x: x,
461 )
462 # Define the component to use ADAPROX as the optimizer
463 components = [component]
465 # Initialize an n=1 sersic component
466 component = EllipticalParametricComponent(
467 bands,
468 bbox,
469 spectrum=parameter(spectrum.copy()),
470 morph_params=parameter(np.array([center[0], center[1], 2 * 1.2**2, 2 * 1.2**2, 0.0, 1])),
471 morph_func=models.sersic,
472 morph_grad=models.grad_sersic,
473 morph_prox=partial(models.bounded_prox, proxmin=_proxmin, proxmax=_proxmax),
474 morph_step=np.array([1e-2, 1e-2, 1e-3, 1e-3, 1e-2, 1e-2]),
475 )
476 # Define the component to use ADAPROX as the optimizer
477 components.append(component)
479 component = EllipticalParametricComponent(
480 bands,
481 bbox,
482 spectrum=parameter(spectrum.copy()),
483 morph_params=parameter(np.array([center[0], center[1], 2 * 1.2**2, 2 * 1.2**2, 0.0, 1])),
484 morph_func=models.sersic,
485 morph_grad=models.grad_sersic,
486 morph_prox=partial(models.bounded_prox, proxmin=_proxmin, proxmax=_proxmax),
487 morph_step=np.array([1e-2, 1e-2, 1e-3, 1e-3, 1e-2, 1e-2]),
488 )
489 # Define the component to use ADAPROX as the optimizer
490 components.append(component)
492 # Create a new source using the two components
493 sources.append(Source(components))
495 # Fit the models
496 blend = Blend(sources, observation)
497 blend.parameterize(parameterize)
498 blend.fit_spectra()
500 # Check properties of a component
501 component = cast(ParametricComponent, blend.components[0])
502 self.assertTupleEqual(component.peak, tuple(self.centers[0]))
503 self.assertEqual(component.y0, component._params.x[0])
504 self.assertEqual(component.x0, component._params.x[1])
505 assert_array_equal(component.morph_step, np.array([1e-2, 1e-2, 1e-2]))
507 component = cast(EllipticalParametricComponent, blend.components[1])
508 self.assertTupleEqual(component.peak, tuple(self.centers[0]))
509 self.assertEqual(component.y0, component._params.x[0])
510 self.assertEqual(component.x0, component._params.x[1])
511 assert_array_equal(component.semi_major, component._params.x[2])
512 assert_array_equal(component.semi_minor, component._params.x[3])
513 assert_array_equal(component.theta, component._params.x[4])
514 assert_array_equal(component.ellipse_params, component._params.x[:5])
515 assert_array_equal(component.radial_params, component._params.x[5:])
516 self.assertIsNotNone(component.morph_prox)
517 self.assertIsNotNone(component.morph_grad)
519 blend.fit(12)
521 # Test elliptical and circular Gaussian models
522 sources = []
523 for idx, center in enumerate(self.centers):
524 # Get the integer center of the source
525 cy, cx = int(np.round(center[0])), int(np.round(center[1]))
526 # For now we use a fixed bounding box that is the size of
527 # the observed PSF image
528 bbox = Box((41, 41), origin=(cy - 20, cx - 20)) & observation.bbox
529 # Keep track of the initial positions
530 yi, xi = cy, cx
531 # Restrict the values of the parameters
532 _proxmin = np.array([yi - 2, xi - 2, 1e-1, 1e-1, -np.pi / 2])
533 _proxmax = np.array([yi + 2, xi + 2, frame.shape[-2] / 2, frame.shape[-1] / 2, np.pi / 2])
535 __proxmin = np.array([yi - 2, xi - 2])
536 __proxmax = np.array([yi + 2, xi + 2])
538 # Initialize a PSF-like component of the source using a
539 # non-pixel integrated gaussian
540 component = ParametricComponent(
541 bands,
542 bbox,
543 spectrum=parameter(spectrum.copy()),
544 morph_params=parameter(np.array([center[0], center[1]])),
545 morph_func=partial(models.circular_gaussian, sigma=0.8),
546 morph_grad=partial(models.grad_circular_gaussian, sigma=0.8),
547 morph_prox=partial(models.bounded_prox, proxmin=__proxmin, proxmax=__proxmax),
548 morph_step=np.array([1e-2, 1e-2, 1e-2]),
549 )
550 # Define the component to use ADAPROX as the optimizer
551 components = [component]
553 # Initialize an n=1 sersic component
554 component = EllipticalParametricComponent(
555 bands,
556 bbox,
557 spectrum=parameter(spectrum.copy()),
558 morph_params=parameter(np.array([center[0], center[1], 2 * 1.2**2, 2 * 1.2**2, 0.0])),
559 morph_func=models.gaussian2d,
560 morph_grad=models.grad_gaussian2,
561 morph_prox=partial(models.bounded_prox, proxmin=_proxmin, proxmax=_proxmax),
562 morph_step=np.array([1e-2, 1e-2, 1e-3, 1e-3, 1e-2, 1e-2]),
563 )
564 # Define the component to use ADAPROX as the optimizer
565 components.append(component)
566 # Create a new source using the two components
567 sources.append(Source(components))
569 # Fit the models
570 blend = Blend(sources, observation)
571 blend.parameterize(parameterize)
572 blend.fit_spectra()
573 blend.fit(12)
575 def test_psf_fitting(self):
576 # Use flat weights for FISTA optimization
577 weights = np.ones(self.data["images"].shape)
579 monotonicity = Monotonicity((101, 101))
581 observation = FittedPsfObservation(
582 self.data["images"],
583 self.data["variance"],
584 weights,
585 self.data["psfs"],
586 self.model_psf[None],
587 bands=self.data["filters"],
588 )
590 def fista_parameterization(component: Component):
591 if isinstance(component, FactorizedComponent):
592 component._spectrum = FistaParameter(component.spectrum, step=0.5)
593 component._morph = FistaParameter(component.morph, step=0.5)
594 else:
595 if isinstance(component, FittedPsfObservation):
596 component._fitted_kernel = FistaParameter(component._fitted_kernel.x, step=1e-2)
598 init = FactorizedInitialization(observation, self.centers, monotonicity=monotonicity)
599 blend = FittedPsfBlend(init.sources, observation).fit_spectra()
600 blend.parameterize(fista_parameterization)
601 assert isinstance(cast(FittedPsfObservation, blend.observation)._fitted_kernel, FistaParameter)
602 blend.fit(12, e_rel=1e-4)