pytme 0.2.0b0__cp311-cp311-macosx_14_0_arm64.whl → 0.2.2__cp311-cp311-macosx_14_0_arm64.whl

scripts/match_template_filters.py

@@ -8,20 +8,19 @@
 import os
 import argparse
 import warnings
-import importlib.util
 from sys import exit
 from time import time
 from typing import Tuple
 from copy import deepcopy
-from os.path import abspath
+from os.path import abspath, exists
 
 import numpy as np
 
 from tme import Density, __version__
 from tme.matching_utils import (
     get_rotation_matrices,
+    get_rotations_around_vector,
     compute_parallelization_schedule,
-    euler_from_rotationmatrix,
     scramble_phases,
     generate_tempfile_name,
     write_pickle,
@@ -32,9 +31,9 @@ from tme.analyzer import (
     MaxScoreOverRotations,
     PeakCallerMaximumFilter,
 )
+from tme.backends import backend as be
 from tme.preprocessing import Compose
-from tme.backends import backend
-
+from tme.scoring import flc_scoring2
 
 def get_func_fullname(func) -> str:
     """Returns the full name of the given function, including its module."""
@@ -51,7 +50,7 @@ def print_block(name: str, data: dict, label_width=20) -> None:
 
 def print_entry() -> None:
     width = 80
-    text = f" pyTME v{__version__} "
+    text = f" pytme v{__version__} "
     padding_total = width - len(text) - 2
     padding_left = padding_total // 2
     padding_right = padding_total - padding_left
@@ -152,8 +151,200 @@ def crop_data(data: Density, cutoff: float, data_mask: Density = None) -> bool:
     return True
 
 
+def parse_rotation_logic(args, ndim):
+    if args.angular_sampling is not None:
+        rotations = get_rotation_matrices(
+            angular_sampling=args.angular_sampling,
+            dim=ndim,
+            use_optimized_set=not args.no_use_optimized_set,
+        )
+        if args.angular_sampling >= 180:
+            rotations = np.eye(ndim).reshape(1, ndim, ndim)
+        return rotations
+
+    if args.axis_sampling is None:
+        args.axis_sampling = args.cone_sampling
+
+    rotations = get_rotations_around_vector(
+        cone_angle=args.cone_angle,
+        cone_sampling=args.cone_sampling,
+        axis_angle=args.axis_angle,
+        axis_sampling=args.axis_sampling,
+        n_symmetry=args.axis_symmetry,
+    )
+    return rotations
+
+
+# TODO: Think about whether wedge mask should also be added to target
+# For now leave it at the cost of incorrect upper bound on the scores
+def setup_filter(args, template: Density, target: Density) -> Tuple[Compose, Compose]:
+    from tme.preprocessing import LinearWhiteningFilter, BandPassFilter
+    from tme.preprocessing.tilt_series import (
+        Wedge,
+        WedgeReconstructed,
+        ReconstructFromTilt,
+    )
+
+    template_filter, target_filter = [], []
+    if args.tilt_angles is not None:
+        try:
+            wedge = Wedge.from_file(args.tilt_angles)
+            wedge.weight_type = args.tilt_weighting
+            if args.tilt_weighting in ("angle", None) and args.ctf_file is None:
+                wedge = WedgeReconstructed(
+                    angles=wedge.angles, weight_wedge=args.tilt_weighting == "angle"
+                )
+        except FileNotFoundError:
+            tilt_step, create_continuous_wedge = None, True
+            tilt_start, tilt_stop = args.tilt_angles.split(",")
+            if ":" in tilt_stop:
+                create_continuous_wedge = False
+                tilt_stop, tilt_step = tilt_stop.split(":")
+            tilt_start, tilt_stop = float(tilt_start), float(tilt_stop)
+            tilt_angles = (tilt_start, tilt_stop)
+            if tilt_step is not None:
+                tilt_step = float(tilt_step)
+                tilt_angles = np.arange(
+                    -tilt_start, tilt_stop + tilt_step, tilt_step
+                ).tolist()
+
+            if args.tilt_weighting is not None and tilt_step is None:
+                raise ValueError(
+                    "Tilt weighting is not supported for continuous wedges."
+                )
+            if args.tilt_weighting not in ("angle", None):
+                raise ValueError(
+                    "Tilt weighting schemes other than 'angle' or 'None' require "
+                    "a specification of electron doses via --tilt_angles."
+                )
+
+            wedge = Wedge(
+                angles=tilt_angles,
+                opening_axis=args.wedge_axes[0],
+                tilt_axis=args.wedge_axes[1],
+                shape=None,
+                weight_type=None,
+                weights=np.ones_like(tilt_angles),
+            )
+            if args.tilt_weighting in ("angle", None) and args.ctf_file is None:
+                wedge = WedgeReconstructed(
+                    angles=tilt_angles,
+                    weight_wedge=args.tilt_weighting == "angle",
+                    create_continuous_wedge=create_continuous_wedge,
+                )
+
+        wedge.opening_axis = args.wedge_axes[0]
+        wedge.tilt_axis = args.wedge_axes[1]
+        wedge.sampling_rate = template.sampling_rate
+        template_filter.append(wedge)
+        if not isinstance(wedge, WedgeReconstructed):
+            template_filter.append(
+                ReconstructFromTilt(
+                    reconstruction_filter=args.reconstruction_filter,
+                    interpolation_order=args.reconstruction_interpolation_order,
+                )
+            )
+
+    if args.ctf_file is not None or args.defocus is not None:
+        from tme.preprocessing.tilt_series import CTF
+
+        needs_reconstruction = True
+        if args.ctf_file is not None:
+            ctf = CTF.from_file(args.ctf_file)
+            n_tilts_ctfs, n_tils_angles = len(ctf.defocus_x), len(wedge.angles)
+            if n_tilts_ctfs != n_tils_angles:
+                raise ValueError(
+                    f"CTF file contains {n_tilts_ctfs} micrographs, but match_template "
+                    f"recieved {n_tils_angles} tilt angles. Expected one angle "
+                    "per micrograph."
+                )
+            ctf.angles = wedge.angles
+            ctf.opening_axis, ctf.tilt_axis = args.wedge_axes
+        else:
+            needs_reconstruction = False
+            ctf = CTF(
+                defocus_x=[args.defocus],
+                phase_shift=[args.phase_shift],
+                defocus_y=None,
+                angles=[0],
+                shape=None,
+                return_real_fourier=True,
+            )
+        ctf.sampling_rate = template.sampling_rate
+        ctf.flip_phase = args.no_flip_phase
+        ctf.amplitude_contrast = args.amplitude_contrast
+        ctf.spherical_aberration = args.spherical_aberration
+        ctf.acceleration_voltage = args.acceleration_voltage * 1e3
+        ctf.correct_defocus_gradient = args.correct_defocus_gradient
+
+        if not needs_reconstruction:
+            template_filter.append(ctf)
+        elif isinstance(template_filter[-1], ReconstructFromTilt):
+            template_filter.insert(-1, ctf)
+        else:
+            template_filter.insert(0, ctf)
+            template_filter.insert(
+                1,
+                ReconstructFromTilt(
+                    reconstruction_filter=args.reconstruction_filter,
+                    interpolation_order=args.reconstruction_interpolation_order,
+                ),
+            )
+
+    if args.lowpass or args.highpass is not None:
+        lowpass, highpass = args.lowpass, args.highpass
+        if args.pass_format == "voxel":
+            if lowpass is not None:
+                lowpass = np.max(np.multiply(lowpass, template.sampling_rate))
+            if highpass is not None:
+                highpass = np.max(np.multiply(highpass, template.sampling_rate))
+        elif args.pass_format == "frequency":
+            if lowpass is not None:
+                lowpass = np.max(np.divide(template.sampling_rate, lowpass))
+            if highpass is not None:
+                highpass = np.max(np.divide(template.sampling_rate, highpass))
+
+        try:
+            if args.lowpass >= args.highpass:
+                warnings.warn("--lowpass should be smaller than --highpass.")
+        except Exception:
+            pass
+
+        bandpass = BandPassFilter(
+            use_gaussian=args.no_pass_smooth,
+            lowpass=lowpass,
+            highpass=highpass,
+            sampling_rate=template.sampling_rate,
+        )
+        template_filter.append(bandpass)
+
+        if not args.no_filter_target:
+            target_filter.append(bandpass)
+
+    if args.whiten_spectrum:
+        whitening_filter = LinearWhiteningFilter()
+        template_filter.append(whitening_filter)
+        target_filter.append(whitening_filter)
+
+    needs_reconstruction = any(
+        [isinstance(t, ReconstructFromTilt) for t in template_filter]
+    )
+    if needs_reconstruction and args.reconstruction_filter is None:
+        warnings.warn(
+            "Consider using a --reconstruction_filter such as 'ramp' to avoid artifacts."
+        )
+
+    template_filter = Compose(template_filter) if len(template_filter) else None
+    target_filter = Compose(target_filter) if len(target_filter) else None
+
+    return template_filter, target_filter
+
+
 def parse_args():
-    parser = argparse.ArgumentParser(description="Perform template matching.")
+    parser = argparse.ArgumentParser(
+        description="Perform template matching.",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
 
     io_group = parser.add_argument_group("Input / Output")
     io_group.add_argument(
@@ -223,14 +414,65 @@ def parse_args():
         choices=list(MATCHING_EXHAUSTIVE_REGISTER.keys()),
         help="Template matching scoring function.",
     )
-    scoring_group.add_argument(
+
+    angular_group = parser.add_argument_group("Angular Sampling")
+    angular_exclusive = angular_group.add_mutually_exclusive_group(required=True)
+
+    angular_exclusive.add_argument(
         "-a",
         dest="angular_sampling",
         type=check_positive,
-        default=40.0,
-        help="Angular sampling rate for template matching. "
+        default=None,
+        help="Angular sampling rate using optimized rotational sets."
         "A lower number yields more rotations. Values >= 180 sample only the identity.",
     )
+    angular_exclusive.add_argument(
+        "--cone_angle",
+        dest="cone_angle",
+        type=check_positive,
+        default=None,
+        help="Half-angle of the cone to be sampled in degrees. Allows to sample a "
+        "narrow interval around a known orientation, e.g. for surface oversampling.",
+    )
+    angular_group.add_argument(
+        "--cone_sampling",
+        dest="cone_sampling",
+        type=check_positive,
+        default=None,
+        help="Sampling rate of the cone in degrees.",
+    )
+    angular_group.add_argument(
+        "--axis_angle",
+        dest="axis_angle",
+        type=check_positive,
+        default=360.0,
+        required=False,
+        help="Sampling angle along the z-axis of the cone.",
+    )
+    angular_group.add_argument(
+        "--axis_sampling",
+        dest="axis_sampling",
+        type=check_positive,
+        default=None,
+        required=False,
+        help="Sampling rate along the z-axis of the cone. Defaults to --cone_sampling.",
+    )
+    angular_group.add_argument(
+        "--axis_symmetry",
+        dest="axis_symmetry",
+        type=check_positive,
+        default=1,
+        required=False,
+        help="N-fold symmetry around z-axis of the cone.",
+    )
+    angular_group.add_argument(
+        "--no_use_optimized_set",
+        dest="no_use_optimized_set",
+        action="store_true",
+        default=False,
+        required=False,
+        help="Whether to use random uniform instead of optimized rotation sets.",
+    )
 
     computation_group = parser.add_argument_group("Computation")
     computation_group.add_argument(
@@ -273,23 +515,7 @@ def parse_args():
         required=False,
         type=float,
         default=0.85,
-        help="Fraction of available memory that can be used. Defaults to 0.85 and is "
-        "ignored if --ram is set",
-    )
-    computation_group.add_argument(
-        "--use_mixed_precision",
-        dest="use_mixed_precision",
-        action="store_true",
-        default=False,
-        help="Use float16 for real values operations where possible.",
-    )
-    computation_group.add_argument(
-        "--use_memmap",
-        dest="use_memmap",
-        action="store_true",
-        default=False,
-        help="Use memmaps to offload large data objects to disk. "
-        "Particularly useful for large inputs in combination with --use_gpu.",
+        help="Fraction of available memory to be used. Ignored if --ram is set."
     )
     computation_group.add_argument(
         "--temp_directory",
@@ -297,6 +523,13 @@ def parse_args():
         default=None,
         help="Directory for temporary objects. Faster I/O improves runtime.",
     )
+    computation_group.add_argument(
+        "--backend",
+        dest="backend",
+        default=None,
+        choices=be.available_backends(),
+        help="[Expert] Overwrite default computation backend.",
+    )
 
     filter_group = parser.add_argument_group("Filters")
     filter_group.add_argument(
@@ -315,11 +548,27 @@ def parse_args():
         help="Resolution to highpass filter template and target to in the same unit "
         "as the sampling rate of template and target (typically Ångstrom).",
     )
+    filter_group.add_argument(
+        "--no_pass_smooth",
+        dest="no_pass_smooth",
+        action="store_false",
+        default=True,
+        help="Whether a hard edge filter should be used for --lowpass and --highpass.",
+    )
+    filter_group.add_argument(
+        "--pass_format",
+        dest="pass_format",
+        type=str,
+        required=False,
+        default="sampling_rate",
+        choices=["sampling_rate", "voxel", "frequency"],
+        help="How values passed to --lowpass and --highpass should be interpreted. ",
+    )
     filter_group.add_argument(
         "--whiten_spectrum",
         dest="whiten_spectrum",
         action="store_true",
-        default=False,
+        default=None,
         help="Apply spectral whitening to template and target based on target spectrum.",
     )
     filter_group.add_argument(
@@ -327,7 +576,7 @@ def parse_args():
         dest="wedge_axes",
         type=str,
         required=False,
-        default="0,2",
+        default=None,
         help="Indices of wedge opening and tilt axis, e.g. 0,2 for a wedge that is open "
         "in z-direction and tilted over the x axis.",
     )
@@ -337,10 +586,10 @@ def parse_args():
         type=str,
         required=False,
         default=None,
-        help="Path to a file with angles and corresponding doses, or comma separated "
-        "start and stop stage tilt angle, e.g. 50,45, which yields a continuous wedge "
-        "mask. Alternatively, a tilt step size can be specified like 50,45:5.0 to "
-        "sample 5.0 degree tilt angle steps.",
+        help="Path to a tab-separated file containing the column angles and optionally "
+        " weights, or comma separated start and stop stage tilt angle, e.g. 50,45, which "
+        " yields a continuous wedge mask. Alternatively, a tilt step size can be "
+        "specified like 50,45:5.0 to sample 5.0 degree tilt angle steps.",
     )
     filter_group.add_argument(
         "--tilt_weighting",
@@ -351,17 +600,100 @@ def parse_args():
         default=None,
         help="Weighting scheme used to reweight individual tilts. Available options: "
         "angle (cosine based weighting), "
-        "relion (relion formalism for wedge weighting ),"
+        "relion (relion formalism for wedge weighting) requires,"
         "grigorieff (exposure filter as defined in Grant and Grigorieff 2015)."
-        "",
+        "relion and grigorieff require electron doses in --tilt_angles weights column.",
+    )
+    filter_group.add_argument(
+        "--reconstruction_filter",
+        dest="reconstruction_filter",
+        type=str,
+        required=False,
+        choices=["ram-lak", "ramp", "ramp-cont", "shepp-logan", "cosine", "hamming"],
+        default=None,
+        help="Filter applied when reconstructing (N+1)-D from N-D filters.",
+    )
+    filter_group.add_argument(
+        "--reconstruction_interpolation_order",
+        dest="reconstruction_interpolation_order",
+        type=int,
+        default=1,
+        required=False,
+        help="Analogous to --interpolation_order but for reconstruction.",
     )
     filter_group.add_argument(
+        "--no_filter_target",
+        dest="no_filter_target",
+        action="store_true",
+        default=False,
+        help="Whether to not apply potential filters to the target.",
+    )
+
+    ctf_group = parser.add_argument_group("Contrast Transfer Function")
+    ctf_group.add_argument(
         "--ctf_file",
         dest="ctf_file",
         type=str,
         required=False,
         default=None,
-        help="Path to a file with CTF parameters.",
+        help="Path to a file with CTF parameters from CTFFIND4. Each line will be "
+        "interpreted as tilt obtained at the angle specified in --tilt_angles. ",
+    )
+    ctf_group.add_argument(
+        "--defocus",
+        dest="defocus",
+        type=float,
+        required=False,
+        default=None,
+        help="Defocus in units of sampling rate (typically Ångstrom). "
+        "Superseded by --ctf_file.",
+    )
+    ctf_group.add_argument(
+        "--phase_shift",
+        dest="phase_shift",
+        type=float,
+        required=False,
+        default=0,
+        help="Phase shift in degrees. Superseded by --ctf_file.",
+    )
+    ctf_group.add_argument(
+        "--acceleration_voltage",
+        dest="acceleration_voltage",
+        type=float,
+        required=False,
+        default=300,
+        help="Acceleration voltage in kV.",
+    )
+    ctf_group.add_argument(
+        "--spherical_aberration",
+        dest="spherical_aberration",
+        type=float,
+        required=False,
+        default=2.7e7,
+        help="Spherical aberration in units of sampling rate (typically Ångstrom).",
+    )
+    ctf_group.add_argument(
+        "--amplitude_contrast",
+        dest="amplitude_contrast",
+        type=float,
+        required=False,
+        default=0.07,
+        help="Amplitude contrast.",
+    )
+    ctf_group.add_argument(
+        "--no_flip_phase",
+        dest="no_flip_phase",
+        action="store_false",
+        required=False,
+        help="Perform phase-flipping CTF correction.",
+    )
+    ctf_group.add_argument(
+        "--correct_defocus_gradient",
+        dest="correct_defocus_gradient",
+        action="store_true",
+        required=False,
+        help="[Experimental] Whether to compute a more accurate 3D CTF incorporating "
+        "defocus gradients.",
     )
 
     performance_group = parser.add_argument_group("Performance")
@@ -404,14 +736,37 @@ def parse_args():
         "for numerical stability. When working with very large targets, e.g. tomograms, "
         "it is safe to use this flag and benefit from the performance gain.",
     )
+    performance_group.add_argument(
+        "--no_filter_padding",
+        dest="no_filter_padding",
+        action="store_true",
+        default=False,
+        help="Omits padding of optional template filters. Particularly effective when "
+        "the target is much larger than the template. However, for fast osciliating "
+        "filters setting this flag can introduce aliasing effects.",
+    )
     performance_group.add_argument(
         "--interpolation_order",
         dest="interpolation_order",
         required=False,
         type=int,
         default=3,
-        help="Spline interpolation used for template rotations. If less than zero "
-        "no interpolation is performed.",
+        help="Spline interpolation used for rotations.",
+    )
+    performance_group.add_argument(
+        "--use_mixed_precision",
+        dest="use_mixed_precision",
+        action="store_true",
+        default=False,
+        help="Use float16 for real values operations where possible.",
+    )
+    performance_group.add_argument(
+        "--use_memmap",
+        dest="use_memmap",
+        action="store_true",
+        default=False,
+        help="Use memmaps to offload large data objects to disk. "
+        "Particularly useful for large inputs in combination with --use_gpu.",
     )
 
     analyzer_group = parser.add_argument_group("Analyzer")
@@ -430,7 +785,15 @@ def parse_args():
         default=False,
         help="Perform peak calling instead of score aggregation.",
     )
+    analyzer_group.add_argument(
+        "--number_of_peaks",
+        dest="number_of_peaks",
+        action="store_true",
+        default=1000,
+        help="Number of peaks to call, 1000 by default..",
+    )
     args = parser.parse_args()
+    args.version = __version__
 
     if args.interpolation_order < 0:
         args.interpolation_order = None
@@ -467,94 +830,22 @@ def parse_args():
             int(x) for x in os.environ["CUDA_VISIBLE_DEVICES"].split(",")
         ]
 
-    if args.wedge_axes is not None:
-        args.wedge_axes = [int(x) for x in args.wedge_axes.split(",")]
-
-    if args.tilt_angles is not None and args.wedge_axes is None:
-        raise ValueError("Wedge axes have to be specified with tilt angles.")
-
-    if args.ctf_file is not None and args.wedge_axes is None:
-        raise ValueError("Wedge axes have to be specified with CTF parameters.")
-    if args.ctf_file is not None and args.tilt_angles is None:
-        raise ValueError("Angles have to be specified with CTF parameters.")
-
-    return args
-
-
-def setup_filter(args, template: Density, target: Density) -> Tuple[Compose, Compose]:
-    from tme.preprocessing import LinearWhiteningFilter, BandPassFilter
-
-    template_filter, target_filter = [], []
     if args.tilt_angles is not None:
-        from tme.preprocessing.tilt_series import (
-            Wedge,
-            WedgeReconstructed,
-            ReconstructFromTilt,
-        )
-
-        try:
-            wedge = Wedge.from_file(args.tilt_angles)
-            wedge.weight_type = args.tilt_weighting
-        except FileNotFoundError:
-            tilt_step = None
-            tilt_start, tilt_stop = args.tilt_angles.split(",")
-            if ":" in tilt_stop:
-                tilt_stop, tilt_step = tilt_stop.split(":")
-            tilt_start, tilt_stop = float(tilt_start), float(tilt_stop)
-            tilt_angles = None
-            if tilt_step is not None:
-                tilt_step = float(tilt_step)
-                tilt_angles = np.arange(
-                    -tilt_start, tilt_stop + tilt_step, tilt_step
-                ).tolist()
-            wedge = WedgeReconstructed(
-                angles=tilt_angles,
-                start_tilt=tilt_start,
-                stop_tilt=tilt_stop,
-            )
-        wedge.opening_axis = args.wedge_axes[0]
-        wedge.tilt_axis = args.wedge_axes[1]
-        wedge.sampling_rate = template.sampling_rate
-        template_filter.append(wedge)
-        if not isinstance(wedge, WedgeReconstructed):
-            template_filter.append(ReconstructFromTilt())
-
-    if args.ctf_file is not None:
-        from tme.preprocessing.tilt_series import CTF
-
-        ctf = CTF.from_file(args.ctf_file)
-        ctf.tilt_axis = args.wedge_axes[1]
-        ctf.opening_axis = args.wedge_axes[0]
-        template_filter.append(ctf)
-        if isinstance(template_filter[-1], ReconstructFromTilt):
-            template_filter.insert(-1, ctf)
-        else:
-            template_filter.insert(0, ctf)
-            template_filter.isnert(1, ReconstructFromTilt())
-
-    if args.lowpass or args.highpass is not None:
-        from tme.preprocessing import BandPassFilter
-
-        bandpass = BandPassFilter(
-            use_gaussian=True,
-            lowpass=args.lowpass,
-            highpass=args.highpass,
-            sampling_rate=template.sampling_rate,
-        )
-        template_filter.append(bandpass)
-        target_filter.append(bandpass)
-
-    if args.whiten_spectrum:
-        from tme.preprocessing import LinearWhiteningFilter
+        if args.wedge_axes is None:
+            raise ValueError("Need to specify --wedge_axes when --tilt_angles is set.")
+        if not exists(args.tilt_angles):
+            try:
+                float(args.tilt_angles.split(",")[0])
+            except ValueError:
+                raise ValueError(f"{args.tilt_angles} is not a file nor a range.")
 
-        whitening_filter = LinearWhiteningFilter()
-        template_filter.append(whitening_filter)
-        target_filter.append(whitening_filter)
+    if args.ctf_file is not None and args.tilt_angles is None:
+        raise ValueError("Need to specify --tilt_angles when --ctf_file is set.")
 
-    template_filter = Compose(template_filter) if len(template_filter) else None
-    target_filter = Compose(target_filter) if len(target_filter) else None
+    if args.wedge_axes is not None:
+        args.wedge_axes = tuple(int(i) for i in args.wedge_axes.split(","))
 
-    return template_filter, target_filter
+    return args
 
 
 def main():
@@ -566,17 +857,20 @@ def main():
     try:
         template = Density.from_file(args.template)
     except Exception:
+        drop = target.metadata.get("batch_dimension", ())
+        keep = [i not in drop for i in range(target.data.ndim)]
         template = Density.from_structure(
             filename_or_structure=args.template,
-            sampling_rate=target.sampling_rate,
+            sampling_rate=target.sampling_rate[keep],
         )
 
-    if not np.allclose(target.sampling_rate, template.sampling_rate):
-        print(
-            f"Resampling template to {target.sampling_rate}. "
-            "Consider providing a template with the same sampling rate as the target."
-        )
-        template = template.resample(target.sampling_rate, order=3)
+    if target.sampling_rate.size == template.sampling_rate.size:
+        if not np.allclose(target.sampling_rate, template.sampling_rate):
+            print(
+                f"Resampling template to {target.sampling_rate}. "
+                "Consider providing a template with the same sampling rate as the target."
+            )
+            template = template.resample(target.sampling_rate, order=3)
 
     template_mask = load_and_validate_mask(
         mask_target=template, mask_path=args.template_mask
@@ -668,72 +962,104 @@ def main():
             template.data, noise_proportion=1.0, normalize_power=True
         )
 
-    available_memory = backend.get_available_memory()
-    if args.use_gpu:
-        args.cores = len(args.gpu_indices)
-        has_torch = importlib.util.find_spec("torch") is not None
-        has_cupy = importlib.util.find_spec("cupy") is not None
-
-        if not has_torch and not has_cupy:
+    # Determine suitable backend for the selected operation
+    available_backends = be.available_backends()
+    if args.backend is not None:
+        req_backend = args.backend
+        if req_backend not in available_backends:
             raise ValueError(
-                "Found neither CuPy nor PyTorch installation. You need to install"
-                " either to enable GPU support."
+                "Requested backend is not available."
             )
+        available_backends = [req_backend,]
 
-        if args.peak_calling:
-            preferred_backend = "pytorch"
-            if not has_torch:
-                preferred_backend = "cupy"
-            backend.change_backend(backend_name=preferred_backend, device="cuda")
-        else:
-            preferred_backend = "cupy"
-            if not has_cupy:
-                preferred_backend = "pytorch"
-            backend.change_backend(backend_name=preferred_backend, device="cuda")
-            if args.use_mixed_precision and preferred_backend == "pytorch":
+    be_selection = ("numpyfftw", "pytorch", "jax", "mlx")
+    if args.use_gpu:
+        args.cores = len(args.gpu_indices)
+        be_selection = ("pytorch", "cupy", "jax")
+    if args.use_mixed_precision:
+        be_selection = tuple(x for x in be_selection if x in ("cupy", "numpyfftw"))
+
+    available_backends = [x for x in available_backends if x in be_selection]
+    if args.peak_calling:
+        if "jax" in available_backends:
+            available_backends.remove("jax")
+        if args.use_gpu and "pytorch" in available_backends:
+            available_backends = ("pytorch",)
+            if args.interpolation_order == 3:
                 raise NotImplementedError(
-                    "pytorch backend does not yet support mixed precision."
-                    " Consider installing CuPy to enable this feature."
-                )
-            elif args.use_mixed_precision:
-                backend.change_backend(
-                    backend_name="cupy",
-                    default_dtype=backend._array_backend.float16,
-                    complex_dtype=backend._array_backend.complex64,
-                    default_dtype_int=backend._array_backend.int16,
+                    "Pytorch does not support --interpolation_order 3, 1 is supported."
                 )
-        available_memory = backend.get_available_memory() * args.cores
-        if preferred_backend == "pytorch" and args.interpolation_order == 3:
-            args.interpolation_order = 1
+    ndim = len(template.shape)
+    if len(target.shape) == ndim and ndim <= 3 and args.use_gpu:
+        available_backends = ["jax", ]
+
+    backend_preference = ("numpyfftw", "pytorch", "jax", "mlx")
+    if args.use_gpu:
+        backend_preference = ("cupy", "jax", "pytorch")
+    for pref in backend_preference:
+        if pref not in available_backends:
+            continue
+        be.change_backend(pref)
+        if pref == "pytorch":
+            be.change_backend(pref, device = "cuda" if args.use_gpu else "cpu")
+
+        if args.use_mixed_precision:
+            be.change_backend(
+                backend_name=pref,
+                default_dtype=be._array_backend.float16,
+                complex_dtype=be._array_backend.complex64,
+                default_dtype_int=be._array_backend.int16,
+            )
+        break
 
+    available_memory = be.get_available_memory() * be.device_count()
     if args.memory is None:
         args.memory = int(args.memory_scaling * available_memory)
 
-    target_padding = np.zeros_like(template.shape)
-    if args.pad_target_edges:
-        target_padding = template.shape
-
-    template_box = template.shape
-    if not args.pad_fourier:
-        template_box = np.ones(len(template_box), dtype=int)
-
     callback_class = MaxScoreOverRotations
     if args.peak_calling:
         callback_class = PeakCallerMaximumFilter
 
+    matching_data = MatchingData(
+        target=target,
+        template=template.data,
+        target_mask=target_mask,
+        template_mask=template_mask,
+        invert_target=args.invert_target_contrast,
+        rotations=parse_rotation_logic(args=args, ndim=template.data.ndim),
+    )
+
+    matching_data.template_filter, matching_data.target_filter  = setup_filter(
+        args, template, target
+    )
+
+    target_dims = target.metadata.get("batch_dimension", None)
+    matching_data._set_matching_dimension(target_dims=target_dims, template_dims=None)
+    args.score = "FLC" if target_dims is not None else args.score
+    args.target_batch, args.template_batch = target_dims, None
+
+    template_box = matching_data._output_template_shape
+    if not args.pad_fourier:
+        template_box = tuple(0 for _ in range(len(template_box)))
+
+    target_padding = tuple(0 for _ in range(len(template_box)))
+    if args.pad_target_edges:
+        target_padding = matching_data._output_template_shape
+
     splits, schedule = compute_parallelization_schedule(
         shape1=target.shape,
-        shape2=template_box,
-        shape1_padding=target_padding,
+        shape2=tuple(int(x) for x in template_box),
+        shape1_padding=tuple(int(x) for x in target_padding),
         max_cores=args.cores,
         max_ram=args.memory,
         split_only_outer=args.use_gpu,
         matching_method=args.score,
         analyzer_method=callback_class.__name__,
-        backend=backend._backend_name,
-        float_nbytes=backend.datatype_bytes(backend._default_dtype),
-        complex_nbytes=backend.datatype_bytes(backend._complex_dtype),
-        integer_nbytes=backend.datatype_bytes(backend._default_dtype_int),
+        backend=be._backend_name,
+        float_nbytes=be.datatype_bytes(be._float_dtype),
+        complex_nbytes=be.datatype_bytes(be._complex_dtype),
+        integer_nbytes=be.datatype_bytes(be._int_dtype),
+        split_axes=target_dims,
     )
 
     if splits is None:
@@ -743,66 +1069,87 @@ def main():
         )
         exit(-1)
 
-    analyzer_args = {
-        "score_threshold": args.score_threshold,
-        "number_of_peaks": 1000,
-        "convolution_mode": "valid",
-        "use_memmap": args.use_memmap,
-    }
-
     matching_setup, matching_score = MATCHING_EXHAUSTIVE_REGISTER[args.score]
-    matching_data = MatchingData(target=target, template=template.data)
-    matching_data.rotations = get_rotation_matrices(
-        angular_sampling=args.angular_sampling, dim=target.data.ndim
-    )
-    if args.angular_sampling >= 180:
-        ndim = target.data.ndim
-        matching_data.rotations = np.eye(ndim).reshape(1, ndim, ndim)
-
-    template_filter, target_filter = setup_filter(args, template, target)
-    matching_data.template_filter = template_filter
-    matching_data.target_filter = target_filter
-
-    matching_data._invert_target = args.invert_target_contrast
-    if target_mask is not None:
-        matching_data.target_mask = target_mask
-    if template_mask is not None:
-        matching_data.template_mask = template_mask.data
-
+    if target_dims is not None:
+        matching_score = flc_scoring2
     n_splits = np.prod(list(splits.values()))
     target_split = ", ".join(
         [":".join([str(x) for x in axis]) for axis in splits.items()]
     )
     gpus_used = 0 if args.gpu_indices is None else len(args.gpu_indices)
     options = {
-        "CPU Cores": args.cores,
-        "Run on GPU": f"{args.use_gpu} [N={gpus_used}]",
-        "Use Mixed Precision": args.use_mixed_precision,
-        "Assigned Memory [MB]": f"{args.memory // 1e6} [out of {available_memory//1e6}]",
-        "Temporary Directory": args.temp_directory,
+        "Angular Sampling": f"{args.angular_sampling}"
+        f" [{matching_data.rotations.shape[0]} rotations]",
+        "Center Template": not args.no_centering,
+        "Scramble Template": args.scramble_phases,
+        "Invert Contrast": args.invert_target_contrast,
         "Extend Fourier Grid": not args.no_fourier_padding,
         "Extend Target Edges": not args.no_edge_padding,
         "Interpolation Order": args.interpolation_order,
-        "Score": f"{args.score}",
         "Setup Function": f"{get_func_fullname(matching_setup)}",
         "Scoring Function": f"{get_func_fullname(matching_score)}",
-        "Angular Sampling": f"{args.angular_sampling}"
-        f" [{matching_data.rotations.shape[0]} rotations]",
-        "Scramble Template": args.scramble_phases,
-        "Target Splits": f"{target_split} [N={n_splits}]",
     }
 
     print_block(
-        name="Template Matching Options",
+        name="Template Matching",
         data=options,
-        label_width=max(len(key) for key in options.keys()) + 2,
+        label_width=max(len(key) for key in options.keys()) + 3,
+    )
+
+    compute_options = {
+        "Backend" :be._BACKEND_REGISTRY[be._backend_name],
+        "Compute Devices" : f"CPU [{args.cores}], GPU [{gpus_used}]",
+        "Use Mixed Precision": args.use_mixed_precision,
+        "Assigned Memory [MB]": f"{args.memory // 1e6} [out of {available_memory//1e6}]",
+        "Temporary Directory": args.temp_directory,
+        "Target Splits": f"{target_split} [N={n_splits}]",
+    }
+    print_block(
+        name="Computation",
+        data=compute_options,
+        label_width=max(len(key) for key in options.keys()) + 3,
     )
 
-    options = {"Analyzer": callback_class, **analyzer_args}
+    filter_args = {
+        "Lowpass": args.lowpass,
+        "Highpass": args.highpass,
+        "Smooth Pass": args.no_pass_smooth,
+        "Pass Format": args.pass_format,
+        "Spectral Whitening": args.whiten_spectrum,
+        "Wedge Axes": args.wedge_axes,
+        "Tilt Angles": args.tilt_angles,
+        "Tilt Weighting": args.tilt_weighting,
+        "Reconstruction Filter": args.reconstruction_filter,
+    }
+    if args.ctf_file is not None or args.defocus is not None:
+        filter_args["CTF File"] = args.ctf_file
+        filter_args["Defocus"] = args.defocus
+        filter_args["Phase Shift"] = args.phase_shift
+        filter_args["Flip Phase"] = args.no_flip_phase
+        filter_args["Acceleration Voltage"] = args.acceleration_voltage
+        filter_args["Spherical Aberration"] = args.spherical_aberration
+        filter_args["Amplitude Contrast"] = args.amplitude_contrast
+        filter_args["Correct Defocus"] = args.correct_defocus_gradient
+
+    filter_args = {k: v for k, v in filter_args.items() if v is not None}
+    if len(filter_args):
+        print_block(
+            name="Filters",
+            data=filter_args,
+            label_width=max(len(key) for key in options.keys()) + 3,
+        )
+
+    analyzer_args = {
+        "score_threshold": args.score_threshold,
+        "number_of_peaks": args.number_of_peaks,
+        "min_distance" : max(template.shape) // 2,
+        "min_boundary_distance" : max(template.shape) // 2,
+        "use_memmap": args.use_memmap,
+    }
     print_block(
-        name="Score Analysis Options",
-        data=options,
-        label_width=max(len(key) for key in options.keys()) + 2,
+        name="Analyzer",
+        data={"Analyzer": callback_class, **analyzer_args},
+        label_width=max(len(key) for key in options.keys()) + 3,
     )
     print("\n" + "-" * 80)
 
@@ -823,6 +1170,7 @@ def main():
         target_splits=splits,
         pad_target_edges=args.pad_target_edges,
         pad_fourier=args.pad_fourier,
+        pad_template_filter=not args.no_filter_padding,
         interpolation_order=args.interpolation_order,
     )
 
@@ -832,19 +1180,19 @@ def main():
             candidates[0] *= target_mask.data
         with warnings.catch_warnings():
             warnings.simplefilter("ignore", category=UserWarning)
+            nbytes = be.datatype_bytes(be._float_dtype)
+            dtype = np.float32 if nbytes == 4 else np.float16
+            rot_dim = matching_data.rotations.shape[1]
             candidates[3] = {
-                x: euler_from_rotationmatrix(
-                    np.frombuffer(i, dtype=matching_data.rotations.dtype).reshape(
-                        candidates[0].ndim, candidates[0].ndim
-                    )
-                )
+                x: np.frombuffer(i, dtype=dtype).reshape(rot_dim, rot_dim)
                 for i, x in candidates[3].items()
             }
-
-    candidates.append((target.origin, template.origin, target.sampling_rate, args))
+    print(np.where(candidates[0] == candidates[0].max()), candidates[0].max())
+    candidates.append((target.origin, template.origin, template.sampling_rate, args))
     write_pickle(data=candidates, filename=args.output)
 
     runtime = time() - start
+    print("\n" + "-" * 80)
     print(f"\nRuntime real: {runtime:.3f}s user: {(runtime * args.cores):.3f}s.")