Feature/jax unit tests

autolens/analysis/positions.py

@@ -224,10 +224,12 @@ def log_likelihood_penalty_base_from(
                 residual_map=residual_map, noise_map=dataset.noise_map
             )
 
-            chi_squared = aa.util.fit.chi_squared_from(chi_squared_map=chi_squared_map)
+            chi_squared = aa.util.fit.chi_squared_from(
+                chi_squared_map=chi_squared_map.array
+            )
 
             noise_normalization = aa.util.fit.noise_normalization_from(
-                noise_map=dataset.noise_map
+                noise_map=dataset.noise_map.array
             )
 
         else:

autolens/imaging/fit_imaging.py

@@ -104,6 +104,7 @@ def tracer_to_inversion(self) -> TracerToInversion:
             noise_map=self.noise_map,
             grids=self.grids,
             psf=self.dataset.psf,
+            convolver=self.dataset.convolver,
             w_tilde=self.w_tilde,
         )
 

autolens/lens/to_inversion.py

@@ -181,6 +181,7 @@ def lp_linear_func_list_galaxy_dict(
                 noise_map=self.dataset.noise_map,
                 grids=grids,
                 psf=self.psf,
+                convolver=self.dataset.convolver,
                 transformer=self.transformer,
                 w_tilde=self.dataset.w_tilde,
             )

autolens/lens/tracer.py

@@ -1,6 +1,5 @@
 from abc import ABC
 import numpy as np
-from functools import wraps
 from scipy.interpolate import griddata
 from typing import Dict, List, Optional, Type, Union
 
@@ -549,9 +548,9 @@ def image_2d_via_input_plane_image_from(
         )[plane_index]
 
         image = griddata(
-            points=plane_grid,
-            values=plane_image,
-            xi=traced_grid.over_sampled,
+            points=plane_grid.array,
+            values=plane_image.array,
+            xi=traced_grid.over_sampled.array,
             fill_value=0.0,
             method="linear",
         )

autolens/point/fit/positions/source/separations.py

@@ -1,4 +1,4 @@
-from autoarray.numpy_wrapper import numpy as npw
+import jax.numpy as jnp
 import numpy as np
 from typing import Optional
 
@@ -126,8 +126,8 @@ def noise_normalization(self) -> float:
         """
         Returns the normalization of the noise-map, which is the sum of the noise-map values squared.
         """
-        return npw.sum(
-            npw.log(
+        return jnp.sum(
+            jnp.log(
                 2
                 * np.pi
                 * (self.magnifications_at_positions**-2.0 * self.noise_map**2.0)

autolens/point/solver/point_solver.py

@@ -1,10 +1,7 @@
 import logging
 from typing import Tuple, Optional
 
-from autoarray.numpy_wrapper import np
-
 import autoarray as aa
-from autoarray.numpy_wrapper import use_jax
 from autoarray.structures.triangles.shape import Point
 
 from autofit.jax_wrapper import jit, register_pytree_node_class
@@ -56,23 +53,23 @@ def solve(
         filtered_means = self._filter_low_magnification(
             tracer=tracer, points=kept_triangles.means
         )
-        if use_jax:
-            return aa.Grid2DIrregular([pair for pair in filtered_means])
-
-        filtered_means = [
-            pair for pair in filtered_means if not np.any(np.isnan(pair)).all()
-        ]
 
-        difference = len(kept_triangles.means) - len(filtered_means)
-        if difference > 0:
-            logger.debug(
-                f"Filtered one multiple-image with magnification below threshold."
-            )
-        elif difference > 1:
-            logger.warning(
-                f"Filtered {difference} multiple-images with magnification below threshold."
-            )
+        return aa.Grid2DIrregular([pair for pair in filtered_means])
 
-        return aa.Grid2DIrregular(
-            [pair for pair in filtered_means if not np.isnan(pair).all()]
-        )
+        # filtered_means = [
+        #     pair for pair in filtered_means if not np.any(np.isnan(pair)).all()
+        # ]
+        #
+        # difference = len(kept_triangles.means) - len(filtered_means)
+        # if difference > 0:
+        #     logger.debug(
+        #         f"Filtered one multiple-image with magnification below threshold."
+        #     )
+        # elif difference > 1:
+        #     logger.warning(
+        #         f"Filtered {difference} multiple-images with magnification below threshold."
+        #     )
+        #
+        # return aa.Grid2DIrregular(
+        #     [pair for pair in filtered_means if not np.isnan(pair).all()]
+        # )

autolens/point/solver/shape_solver.py

@@ -1,3 +1,5 @@
+import jax.numpy as jnp
+from jax import jit
 import logging
 import math
 
@@ -6,23 +8,11 @@
 import autoarray as aa
 
 from autoarray.structures.triangles.shape import Shape
-from autofit.jax_wrapper import jit, use_jax, numpy as np, register_pytree_node_class
-
-try:
-    if use_jax:
-        from autoarray.structures.triangles.coordinate_array.jax_coordinate_array import (
-            CoordinateArrayTriangles,
-        )
-    else:
-        from autoarray.structures.triangles.coordinate_array.coordinate_array import (
-            CoordinateArrayTriangles,
-        )
-
-except ImportError:
-    from autoarray.structures.triangles.coordinate_array.coordinate_array import (
-        CoordinateArrayTriangles,
-    )
+from autofit.jax_wrapper import register_pytree_node_class
 
+from autoarray.structures.triangles.coordinate_array.jax_coordinate_array import (
+    CoordinateArrayTriangles,
+)
 from autoarray.structures.triangles.abstract import AbstractTriangles
 
 from autogalaxy import OperateDeflections
@@ -278,13 +268,13 @@ def _filter_low_magnification(
         -------
         The points with an absolute magnification above the threshold.
         """
-        points = np.array(points)
+        points = jnp.array(points)
         magnifications = tracer.magnification_2d_via_hessian_from(
             grid=aa.Grid2DIrregular(points),
             buffer=self.scale,
         )
-        mask = np.abs(magnifications.array) > self.magnification_threshold
-        return np.where(mask[:, None], points, np.nan)
+        mask = jnp.abs(magnifications.array) > self.magnification_threshold
+        return jnp.where(mask[:, None], points, jnp.nan)
 
     def _source_triangles(
         self,

docs/installation/conda.rst

@@ -105,7 +105,6 @@ For interferometer analysis there are two optional dependencies that must be ins
 .. code-block:: bash
 
     pip install pynufft
-    pip install pylops==2.3.1
 
 **PyAutoLens** will run without these libraries and it is recommended that you only install them if you intend to
 do interferometer analysis.

docs/installation/overview.rst

@@ -66,6 +66,4 @@ Dependencies
 
 And the following optional dependencies:
 
-**pynufft**: https://github.com/jyhmiinlin/pynufft
-
-**PyLops**: https://github.com/PyLops/pylops
+**pynufft**: https://github.com/jyhmiinlin/pynufft

docs/installation/pip.rst

@@ -86,7 +86,6 @@ For interferometer analysis there are two optional dependencies that must be ins
 .. code-block:: bash
 
     pip install pynufft
-    pip install pylops==2.3.1
 
 **PyAutoLens** will run without these libraries and it is recommended that you only install them if you intend to
 do interferometer analysis.

docs/installation/source.rst

@@ -59,7 +59,6 @@ For unit tests to pass you will also need the following optional requirements:
 .. code-block:: bash
 
     pip install pynufft
-    pip install pylops==2.3.1
 
 If you are using a ``conda`` environment, add the source repository as follows:
 

files/citations.bib

@@ -33,17 +33,6 @@ @article{astropy2
 Bdsk-Url-1 = {https://doi.org/10.3847/1538-3881/aabc4f}
 }
 
-@article{PyLops,
-abstract = {Linear operators and optimisation are at the core of many algorithms used in signal and image processing, remote sensing, and inverse problems. For small to medium-scale problems, existing software packages (e.g., MATLAB, Python numpy and scipy) allow for explicitly building dense (or sparse) matrices and performing algebraic operations (e.g., computation of matrix-vector products and manipulation of matrices) with syntax that closely represents their corresponding analytical forms. However, many real application, large-scale operators do not lend themselves to explicit matrix representations, usually forcing practitioners to forego of the convenient linear-algebra syntax available for their explicit-matrix counterparts. PyLops is an open-source Python library providing a flexible and scalable framework for the creation and combination of so-called linear operators, class-based entities that represent matrices and inherit their associated syntax convenience, but do not rely on the creation of explicit matrices. We show that PyLops operators can dramatically reduce the memory load and CPU computations compared to explicit-matrix calculations, while still allowing users to seamlessly use their existing knowledge of compact matrix-based syntax that scales to any problem size because no explicit matrices are required.},
-archivePrefix = {arXiv},
-arxivId = {1907.12349},
-author = {Ravasi, Matteo and Vasconcelos, Ivan},
-eprint = {1907.12349},
-file = {:home/jammy/Documents/Papers/Software/PyLops.pdf:pdf},
-title = {{PyLops -- A Linear-Operator Python Library for large scale optimization}},
-url = {http://arxiv.org/abs/1907.12349},
-year = {2019}
-}
 
 @article{colossus,
 abstract = {This paper introduces Colossus, a public, open-source python package for calculations related to cosmology, the large-scale structure (LSS) of matter in the universe, and the properties of dark matter halos. The code is designed to be fast and easy to use, with a coherent, well-documented user interface. The cosmology module implements Friedman-Lemaitre-Robertson-Walker cosmologies including curvature, relativistic species, and different dark energy equations of state, and provides fast computations of the linear matter power spectrum, variance, and correlation function. The LSS module is concerned with the properties of peaks in Gaussian random fields and halos in a statistical sense, including their peak height, peak curvature, halo bias, and mass function. The halo module deals with spherical overdensity radii and masses, density profiles, concentration, and the splashback radius. To facilitate the rapid exploration of these quantities, Colossus implements more than 40 different fitting functions from the literature. I discuss the core routines in detail, with particular emphasis on their accuracy. Colossus is available at bitbucket.org/bdiemer/colossus.},

files/citations.md

@@ -19,7 +19,6 @@ This work uses the following software packages:
 - `PyAutoFit` https://github.com/rhayes777/PyAutoFit [@pyautofit]
 - `PyAutoGalaxy` https://github.com/Jammy2211/PyAutoGalaxy [@Nightingale2018] [@pyautogalaxy]
 - `PyAutoLens` https://github.com/Jammy2211/PyAutoLens [@Nightingale2015] [@Nightingale2018] [@pyautolens]
-- `PyLops` https://github.com/equinor/pylops [@pylops]
 - `PyNUFFT` https://github.com/jyhmiinlin/pynufft [@pynufft]
 - `PySwarms` https://github.com/ljvmiranda921/pyswarms [@pyswarms]
 - `Python` https://www.python.org/ [@python]

files/citations.tex

@@ -54,9 +54,6 @@ \section*{Software Citations}
 \href{https://github.com/Jammy2211/PyAutoLens}{\textt{PyAutoLens}}
 \citep{Nightingale2015, Nightingale2018, pyautolens}
 
-\item
-\href{https://github.com/equinor/pylops}{\textt{PyLops}}
-\citep{pylops}
 
 \item
 \href{https://github.com/jyhmiinlin/pynufft}{\textt{PyNUFFT}}

optional_requirements.txt

@@ -1,5 +1,4 @@
 numba
-pylops>=1.10.0,<=2.3.1
 pynufft
 zeus-mcmc==2.5.4
 getdist==1.4

paper/paper.bib

@@ -30,17 +30,7 @@ @article{astropy2
 Volume = {156},
 Year = 2018,
 Bdsk-Url-1 = {https://doi.org/10.3847/1538-3881/aabc4f}}
-@article{PyLops,
-abstract = {Linear operators and optimisation are at the core of many algorithms used in signal and image processing, remote sensing, and inverse problems. For small to medium-scale problems, existing software packages (e.g., MATLAB, Python numpy and scipy) allow for explicitly building dense (or sparse) matrices and performing algebraic operations (e.g., computation of matrix-vector products and manipulation of matrices) with syntax that closely represents their corresponding analytical forms. However, many real application, large-scale operators do not lend themselves to explicit matrix representations, usually forcing practitioners to forego of the convenient linear-algebra syntax available for their explicit-matrix counterparts. PyLops is an open-source Python library providing a flexible and scalable framework for the creation and combination of so-called linear operators, class-based entities that represent matrices and inherit their associated syntax convenience, but do not rely on the creation of explicit matrices. We show that PyLops operators can dramatically reduce the memory load and CPU computations compared to explicit-matrix calculations, while still allowing users to seamlessly use their existing knowledge of compact matrix-based syntax that scales to any problem size because no explicit matrices are required.},
-archivePrefix = {arXiv},
-arxivId = {1907.12349},
-author = {Ravasi, Matteo and Vasconcelos, Ivan},
-eprint = {1907.12349},
-file = {:home/jammy/Documents/Papers/Software/PyLops.pdf:pdf},
-title = {{PyLops -- A Linear-Operator Python Library for large scale optimization}},
-url = {http://arxiv.org/abs/1907.12349},
-year = {2019}
-}
+
 @article{colossus,
 abstract = {This paper introduces Colossus, a public, open-source python package for calculations related to cosmology, the large-scale structure (LSS) of matter in the universe, and the properties of dark matter halos. The code is designed to be fast and easy to use, with a coherent, well-documented user interface. The cosmology module implements Friedman-Lemaitre-Robertson-Walker cosmologies including curvature, relativistic species, and different dark energy equations of state, and provides fast computations of the linear matter power spectrum, variance, and correlation function. The LSS module is concerned with the properties of peaks in Gaussian random fields and halos in a statistical sense, including their peak height, peak curvature, halo bias, and mass function. The halo module deals with spherical overdensity radii and masses, density profiles, concentration, and the splashback radius. To facilitate the rapid exploration of these quantities, Colossus implements more than 40 different fitting functions from the literature. I discuss the core routines in detail, with particular emphasis on their accuracy. Colossus is available at bitbucket.org/bdiemer/colossus.},
 archivePrefix = {arXiv},

paper/paper.md

@@ -160,8 +160,7 @@ effects like the telescope optics and background sky subtraction in the model-fi
 performed directly on the observed visibilities in their native Fourier space, circumventing issues associated with the 
 incomplete sampling of the uv-plane that give rise to artefacts that can bias the inferred mass model and source 
 reconstruction in real-space. To make feasible the analysis of millions of visibilities, `PyAutoLens` 
-uses `PyNUFFT` [@pynufft] to fit the visibilities via a non-uniform fast Fourier transform and `PyLops` [@PyLops] to 
-express the memory-intensive linear algebra calculations as efficient linear operators [@Powell2020]. Creating 
+uses `PyNUFFT` [@pynufft] to fit the visibilities via a non-uniform fast Fourier transform. Creating 
 realistic simulations of imaging and interferometer strong lensing datasets is possible, as performed 
 by [@Alexander2019] [@Hermans2019] who used `PyAutoLens` to train neural networks to detect strong lenses.
 
@@ -198,7 +197,6 @@ taken without a local `PyAutoLens` installation.
 - `numba` [@numba]
 - `NumPy` [@numpy]
 - `PyAutoFit` [@pyautofit]
-- `PyLops` [@PyLops]
 - `PyMultiNest` [@pymultinest] [@multinest]
 - `PyNUFFT` [@pynufft]
 - `pyprojroot` (https://github.com/chendaniely/pyprojroot)

test_autolens/config/notation.yaml

@@ -62,7 +62,6 @@ label:
     weight_power: W_{\rm p}
   superscript:
     ExternalShear: ext
-    InputDeflections: defl
     Pixelization: pix
     Point: point
     Redshift: ''

test_autolens/imaging/model/test_analysis_imaging.py

@@ -67,8 +67,8 @@ def test__positions__resample__raises_exception(masked_imaging_7x7):
 
 
 def test__positions__likelihood_overwrites__changes_likelihood(masked_imaging_7x7):
-    lens = al.Galaxy(redshift=0.5, mass=al.mp.IsothermalSph())
-    source = al.Galaxy(redshift=1.0, light=al.lp.SersicSph())
+    lens = al.Galaxy(redshift=0.5, mass=al.mp.IsothermalSph(centre=(0.05, 0.05)))
+    source = al.Galaxy(redshift=1.0, light=al.lp.SersicSph(centre=(0.05, 0.05)))
 
     model = af.Collection(galaxies=af.Collection(lens=lens, source=source))
 
@@ -82,7 +82,7 @@ def test__positions__likelihood_overwrites__changes_likelihood(masked_imaging_7x
     fit = al.FitImaging(dataset=masked_imaging_7x7, tracer=tracer)
 
     assert fit.log_likelihood == pytest.approx(analysis_log_likelihood, 1.0e-4)
-    assert analysis_log_likelihood == pytest.approx(-6258.043397009, 1.0e-4)
+    assert analysis_log_likelihood == pytest.approx(-14.79034680979, 1.0e-4)
 
     positions_likelihood = al.PositionsLHPenalty(
         positions=al.Grid2DIrregular([(1.0, 100.0), (200.0, 2.0)]), threshold=0.01

test_autolens/imaging/model/test_result_imaging.py

@@ -9,7 +9,7 @@
 def test___linear_light_profiles_in_result(analysis_imaging_7x7):
 
     galaxies = af.ModelInstance()
-    galaxies.galaxy = al.Galaxy(redshift=0.5, bulge=al.lp_linear.Sersic())
+    galaxies.galaxy = al.Galaxy(redshift=0.5, bulge=al.lp_linear.Sersic(centre=(0.05, 0.05)))
 
     instance = af.ModelInstance()
     instance.galaxies = galaxies
@@ -24,4 +24,4 @@ def test___linear_light_profiles_in_result(analysis_imaging_7x7):
     )
     assert result.max_log_likelihood_tracer.galaxies[
         0
-    ].bulge.intensity == pytest.approx(0.002310735, 1.0e-4)
+    ].bulge.intensity == pytest.approx(0.1868684644, 1.0e-4)