File size: 26,370 Bytes
7f5c4ef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ec6b668
 
 
7f5c4ef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ec6b668
 
7f5c4ef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ec6b668
 
7f5c4ef
ec6b668
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7f5c4ef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
"""
Utility functions for Gradio demos

Provides reusable components for:
- Data loading from OME-Zarr stores
- Image normalization and processing
- Slice extraction from xarray DataArrays
- Phase reconstruction and optimization

Design Notes
------------
All image processing functions work with xarray.DataArray to maintain
labeled dimensions and coordinate information as long as possible.
Only convert to numpy arrays at the final display step.
"""

from pathlib import Path
from typing import Generator

import numpy as np
import torch
import xarray as xr
from numpy.typing import NDArray
from xarray_ome import open_ome_dataset

from waveorder import util
from waveorder.models import isotropic_thin_3d
from waveorder.cli.compute_transfer_function import (
    _position_list_from_shape_scale_offset,
)

# Type alias for device specification
Device = torch.device | str | None


def get_device(device: Device = None) -> torch.device:
    """
    Get torch device with smart defaults.

    Parameters
    ----------
    device : torch.device | str | None
        If None, auto-selects cuda if available, else cpu.
        If str, converts to torch.device.
        If torch.device, returns as-is.

    Returns
    -------
    torch.device
        Validated device ready for use

    Examples
    --------
    >>> get_device()  # Auto-detect
    device(type='cuda', index=0)  # if GPU available

    >>> get_device("cpu")  # Force CPU
    device(type='cpu')

    >>> get_device(torch.device("cuda:1"))  # Specific GPU
    device(type='cuda', index=1)
    """
    if device is None:
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        if device.type == "cuda":
            print(f"🚀 Using GPU: {torch.cuda.get_device_name(device)}")
            gpu_mem_gb = torch.cuda.get_device_properties(device).total_memory / 1e9
            print(f"   GPU Memory: {gpu_mem_gb:.2f} GB")
        else:
            print("💻 Using CPU (GPU not available)")
        return device

    if isinstance(device, str):
        return torch.device(device)

    return device


# === HCS Plate Loading with iohub ===
def get_plate_metadata(zarr_path: Path | str, allowed_fovs: list[str]) -> dict:
    """
    Extract HCS plate metadata for FOV selection using iohub.

    Optimized to only load metadata for specified FOVs.

    Parameters
    ----------
    zarr_path : Path | str
        Path to the HCS plate zarr store
    allowed_fovs : list[str]
        List of allowed FOV names (e.g., ['002026', '002027', '002028'])

    Returns
    -------
    dict
        Metadata with keys:
        - 'rows': list of row names (e.g., ['A'])
        - 'columns': list of column names (e.g., ['1', '2', '3'])
        - 'wells': dict mapping (row, col) to list of field names
        - 'plate': iohub Plate object for later access
        - 'zarr_path': stored path for data loading
    """
    from iohub import open_ome_zarr

    # Open HCS plate with iohub (fast - doesn't load data)
    plate = open_ome_zarr(str(zarr_path), mode="r", layout="hcs")

    # Hardcoded metadata for known structure (avoids iterating 1000s of positions)
    rows = ["A"]
    columns = ["1", "2", "3"]

    # Only return the allowed FOVs for each well
    wells = {
        ("A", "1"): allowed_fovs,
        ("A", "2"): allowed_fovs,
        ("A", "3"): allowed_fovs,
    }

    return {
        "rows": rows,
        "columns": columns,
        "wells": wells,
        "plate": plate,
        "zarr_path": str(zarr_path),
    }


def load_fov_from_plate(
    plate, row: str, column: str, field: str, resolution: int = 0
) -> xr.DataArray:
    """
    Load a specific FOV from HCS plate using hybrid iohub + xarray-ome approach.

    Uses iohub for navigation, then xarray-ome for fast data loading.

    Parameters
    ----------
    plate : iohub.Plate
        Plate loaded with open_ome_zarr(..., layout="hcs")
    row : str
        Row name (e.g., 'A')
    column : str
        Column name (e.g., '1')
    field : str
        Field/position name (e.g., '002026')
    resolution : int, optional
        Resolution level to load, by default 0

    Returns
    -------
    xr.DataArray
        Image data with labeled dimensions (T, C, Z, Y, X)
    """
    # Navigate to position using iohub (fast)
    position_key = f"{row}/{column}/{field}"
    position = plate[position_key]

    # Get full zarr path from position (handle both Zarr V2 and V3)
    store = position.zgroup.store
    if hasattr(store, 'path'):
        base_path = Path(store.path)  # Zarr V2 (DirectoryStore)
    elif hasattr(store, 'root'):
        base_path = Path(store.root)  # Zarr V3 (LocalStore)
    else:
        raise RuntimeError(f"Unknown store type: {type(store)}")

    position_path = base_path / position.zgroup.path

    # Load with xarray-ome (fast and reliable)
    fov_dataset = open_ome_dataset(position_path, resolution=resolution, validate=False)
    data_xr = fov_dataset["image"]

    return data_xr


# === Data Loading ===
def load_ome_zarr_fov(
    zarr_path: Path | str, fov_path: Path | str, resolution: int = 0
) -> xr.DataArray:
    """
    Load a field of view from an OME-Zarr store as an xarray DataArray.

    Parameters
    ----------
    zarr_path : Path | str
        Path to the root OME-Zarr store
    fov_path : Path | str
        Relative path to the FOV (e.g., "A/1/001007")
    resolution : int, optional
        Resolution level to load (0 is full resolution), by default 0

    Returns
    -------
    xr.DataArray
        Image data with labeled dimensions (T, C, Z, Y, X)
    """
    zarr_path = Path(zarr_path)
    fov_path = Path(fov_path)

    print(f"Loading zarr store from: {zarr_path}")
    print(f"Accessing FOV: {fov_path}")

    # Load as xarray Dataset
    fov_dataset: xr.Dataset = open_ome_dataset(
        zarr_path / fov_path, resolution=resolution, validate=False
    )

    # Extract the image DataArray
    data_xr = fov_dataset["image"]

    print(f"Loaded data shape: {dict(data_xr.sizes)}")
    print(f"Dimensions: {list(data_xr.dims)}")
    print(f"Data type: {data_xr.dtype}")

    return data_xr


# === Image Processing ===
def normalize_for_display(
    img_2d: xr.DataArray,
    percentiles: tuple[float, float] = (1, 99),
    clip_to_uint8: bool = True,
) -> np.ndarray:
    """
    Normalize a 2D microscopy image using percentile clipping.

    Uses robust percentile-based normalization to handle outliers
    common in microscopy data. Works with xarray DataArrays to maintain
    labeled dimensions through the processing pipeline.

    Parameters
    ----------
    img_2d : xr.DataArray
        2D image DataArray to normalize
    percentiles : tuple[float, float], optional
        Lower and upper percentiles for clipping, by default (1, 99)
    clip_to_uint8 : bool, optional
        If True, convert to uint8 (0-255), otherwise keep as float (0-1),
        by default True

    Returns
    -------
    np.ndarray
        Normalized numpy array (uint8 if clip_to_uint8=True, else float32)

    Notes
    -----
    Expects xarray.DataArray input. For raw numpy arrays,
    wrap in xarray first: xr.DataArray(array, dims=["Y", "X"])
    """
    # Calculate percentiles using xarray
    p_low = float(img_2d.quantile(percentiles[0] / 100.0).values)
    p_high = float(img_2d.quantile(percentiles[1] / 100.0).values)

    # Handle edge case: no intensity variation
    if p_high - p_low < 1e-10:
        return np.zeros(img_2d.shape, dtype=np.uint8 if clip_to_uint8 else np.float32)

    # Clip and normalize using xarray operations
    img_clipped = img_2d.clip(min=p_low, max=p_high)
    img_normalized = (img_clipped - p_low) / (p_high - p_low)

    # Convert to numpy array
    result = img_normalized.values

    # Convert to requested output format
    if clip_to_uint8:
        result = (result * 255).astype(np.uint8)
    else:
        result = result.astype(np.float32)

    return result


# === Slice Extraction ===
def extract_2d_slice(
    data_xr: xr.DataArray,
    t: int | None = None,
    c: int | None = None,
    z: int | None = None,
    normalize: bool = True,
    verbose: bool = True,
) -> np.ndarray:
    """
    Extract and optionally normalize a 2D slice from xarray data.

    Flexibly handles different dimension specifications. If a dimension
    index is None, it will be squeezed out if size=1 or raise an error
    if size>1.

    Parameters
    ----------
    data_xr : xr.DataArray
        Image data with dimensions (T, C, Z, Y, X)
    t : int | None, optional
        Timepoint index, by default None
    c : int | None, optional
        Channel index, by default None
    z : int | None, optional
        Z-slice index, by default None
    normalize : bool, optional
        Whether to normalize for display, by default True
    verbose : bool, optional
        Whether to print slice information, by default True

    Returns
    -------
    np.ndarray
        2D numpy array (normalized uint8 if normalize=True, else raw values)

    Raises
    ------
    ValueError
        If result is empty or not 2D after slicing and squeezing
    """
    # Build selection dictionary for indexed dimensions
    sel_dict = {}
    if t is not None:
        sel_dict["T"] = int(t)
    if c is not None:
        sel_dict["C"] = int(c)
    if z is not None:
        sel_dict["Z"] = int(z)

    # Extract slice using xarray's labeled indexing
    slice_xr = data_xr.isel(**sel_dict) if sel_dict else data_xr

    # Compute if Dask-backed (load from disk)
    if hasattr(slice_xr.data, "compute"):
        slice_xr = slice_xr.compute()

    # Squeeze singleton dimensions (e.g., single channel, single Z)
    slice_xr = slice_xr.squeeze()

    # Validation: ensure non-empty result
    if slice_xr.size == 0:
        raise ValueError(
            f"Empty array after slicing. Selection: {sel_dict}, "
            f"Original shape: {data_xr.shape}"
        )

    # Validation: ensure 2D result
    if slice_xr.ndim != 2:
        raise ValueError(
            f"Expected 2D array after slicing, got shape {slice_xr.shape}. "
            f"Selection: {sel_dict}"
        )

    # Verbose output: print slice information
    if verbose:
        sel_str = (
            ", ".join(f"{k}={v}" for k, v in sel_dict.items())
            if sel_dict
            else "full array"
        )
        print(
            f"Extracted slice: {sel_str}, Shape={slice_xr.shape}, "
            f"Range=[{float(slice_xr.min()):.1f}, {float(slice_xr.max()):.1f}]"
        )

    # Normalize or convert to numpy
    if normalize:
        slice_2d = normalize_for_display(slice_xr)
    else:
        slice_2d = slice_xr.values

    return slice_2d


# === Slice Extraction Factory ===
def create_slice_extractor(
    data_xr: xr.DataArray,
    normalize: bool = True,
    channel: int = 0,
):
    """
    Create a closure function for extracting slices from a specific dataset.

    This factory function is useful for Gradio callbacks where the data
    is loaded once and the same extraction function is called multiple times.

    Parameters
    ----------
    data_xr : xr.DataArray
        Image data to extract slices from
    normalize : bool, optional
        Whether to normalize for display, by default True
    channel : int, optional
        Default channel to use, by default 0

    Returns
    -------
    callable
        Function with signature (t: int, z: int) -> np.ndarray that extracts
        and normalizes 2D slices
    """

    def get_slice(t: int, z: int) -> np.ndarray:
        """Extract and normalize a 2D slice at timepoint t and z-slice z."""
        return extract_2d_slice(
            data_xr,
            t=int(t),
            c=channel,
            z=int(z),
            normalize=normalize,
            verbose=True,
        )

    return get_slice


# === Metadata Helpers ===
def get_dimension_info(data_xr: xr.DataArray) -> dict:
    """
    Extract dimension information from xarray DataArray.

    Parameters
    ----------
    data_xr : xr.DataArray
        Image data with dimensions

    Returns
    -------
    dict
        Dictionary with keys: 'sizes', 'dims', 'coords', 'dtype'
    """
    return {
        "sizes": dict(data_xr.sizes),
        "dims": list(data_xr.dims),
        "coords": {dim: data_xr.coords[dim].values.tolist() for dim in data_xr.dims},
        "dtype": str(data_xr.dtype),
    }


def print_data_summary(data_xr: xr.DataArray) -> None:
    """
    Print a formatted summary of xarray DataArray.

    Parameters
    ----------
    data_xr : xr.DataArray
        Image data to summarize
    """
    info = get_dimension_info(data_xr)

    print("\n" + "=" * 60)
    print("DATA SUMMARY")
    print("=" * 60)
    print(f"Shape: {info['sizes']}")
    print(f"Dimensions: {info['dims']}")
    print(f"Data type: {info['dtype']}")

    # Print coordinate ranges
    print("\nCoordinate Ranges:")
    for dim in info["dims"]:
        coords = info["coords"][dim]
        if len(coords) > 0:
            print(f"  {dim}: [{coords[0]:.2f} ... {coords[-1]:.2f}] (n={len(coords)})")

    # Print memory size estimate
    total_elements = np.prod(list(info["sizes"].values()))
    dtype_size = np.dtype(data_xr.dtype).itemsize
    size_mb = (total_elements * dtype_size) / (1024**2)
    print(f"\nEstimated size: {size_mb:.1f} MB")
    print("=" * 60 + "\n")


# === Phase Reconstruction Functions ===
def run_reconstruction(zyx_tile: torch.Tensor, recon_args: dict) -> torch.Tensor:
    """
    Run phase reconstruction on a Z-stack.

    Uses waveorder's official _position_list_from_shape_scale_offset
    to ensure proper z-position calculation and correct phase sign.

    Parameters
    ----------
    zyx_tile : torch.Tensor
        Input Z-stack data with shape (Z, Y, X). Can be on CPU or GPU.
    recon_args : dict
        Reconstruction arguments including wavelength, NA, pixel sizes, etc.
        All tensor values should be on the same device as zyx_tile.

    Returns
    -------
    torch.Tensor
        Reconstructed 2D phase image with shape (Y, X), on same device as input.

    Notes
    -----
    All intermediate tensors are created on the same device as the input
    to ensure efficient computation without device transfers.
    """
    # Infer device from input tensor
    device = zyx_tile.device

    # Prepare transfer function arguments
    tf_args = recon_args.copy()
    Z, _, _ = zyx_tile.shape

    # Extract z_offset value (keep as tensor if it is one, for gradient flow)
    z_offset_value = recon_args["z_offset"]
    if torch.is_tensor(z_offset_value):
        # For optimization: extract scalar value for _position_list function
        z_offset_scalar = z_offset_value.item()
    else:
        z_offset_scalar = z_offset_value

    # Use waveorder's official function (returns torch.Tensor on CPU)
    z_position_list_cpu = _position_list_from_shape_scale_offset(
        shape=Z,
        scale=recon_args["z_scale"],
        offset=z_offset_scalar,
    )

    # Move to device and ensure gradient connection if z_offset is a parameter
    if torch.is_tensor(z_offset_value) and z_offset_value.requires_grad:
        # Recompute on device to maintain gradient connection
        # Uses same formula as waveorder: -arange(Z) + (Z // 2) + offset
        z_position_list = (
            -torch.arange(Z, dtype=torch.float32, device=device) + (Z // 2) + z_offset_value
        ) * recon_args["z_scale"]
    else:
        # No gradient needed, just move to device
        z_position_list = z_position_list_cpu.to(device)

    tf_args["z_position_list"] = z_position_list
    tf_args.pop("z_offset")
    tf_args.pop("z_scale")

    # Core reconstruction calls (all on same device)
    tf_abs, tf_phase = isotropic_thin_3d.calculate_transfer_function(**tf_args)
    system = isotropic_thin_3d.calculate_singular_system(tf_abs, tf_phase)
    _, yx_phase_recon = isotropic_thin_3d.apply_inverse_transfer_function(
        zyx_tile, system, regularization_strength=1e-2
    )
    return yx_phase_recon


def compute_midband_power(
    yx_array: torch.Tensor,
    NA_det: float,
    lambda_ill: float,
    pixel_size: float,
    band: tuple[float, float] = (0.125, 0.25),
) -> torch.Tensor:
    """
    Compute midband power metric for optimization loss.

    Parameters
    ----------
    yx_array : torch.Tensor
        2D reconstructed image (on CPU or GPU)
    NA_det : float
        Numerical aperture of detection
    lambda_ill : float
        Illumination wavelength
    pixel_size : float
        Pixel size in same units as wavelength
    band : tuple[float, float], optional
        Frequency band as fraction of cutoff, by default (0.125, 0.25)

    Returns
    -------
    torch.Tensor
        Scalar power value in the specified frequency band, on same device as input.

    Notes
    -----
    All operations are performed on the same device as the input tensor
    for efficient GPU computation.
    """
    device = yx_array.device

    # Generate frequency coordinates (returns numpy arrays)
    _, _, fxx, fyy = util.gen_coordinate(yx_array.shape, pixel_size)

    # Convert to torch tensor on same device
    frr = torch.tensor(np.sqrt(fxx**2 + fyy**2), dtype=torch.float32, device=device)

    # FFT and frequency masking (all on device)
    xy_abs_fft = torch.abs(torch.fft.fftn(yx_array))
    cutoff = 2 * NA_det / lambda_ill
    mask = torch.logical_and(frr > cutoff * band[0], frr < cutoff * band[1])

    return torch.sum(xy_abs_fft[mask])


def prepare_optimizer(
    optimizable_params: dict[str, tuple[bool, float, float]],
    device: torch.device,
) -> tuple[dict[str, torch.nn.Parameter], torch.optim.Optimizer]:
    """
    Prepare optimization parameters and Adam optimizer.

    Parameters
    ----------
    optimizable_params : dict
        Dict mapping param names to (enabled, initial_value, learning_rate)
    device : torch.device
        Device to create parameters on (CPU or GPU)

    Returns
    -------
    tuple[dict, Optimizer]
        optimization_params dict and configured optimizer

    Notes
    -----
    All parameters are created on the specified device for efficient
    GPU-accelerated optimization if available.
    """
    optimization_params: dict[str, torch.nn.Parameter] = {}
    optimizer_config = []
    for name, (enabled, initial, lr) in optimizable_params.items():
        if enabled:
            param = torch.nn.Parameter(
                torch.tensor([initial], dtype=torch.float32, device=device),
                requires_grad=True,
            )
            optimization_params[name] = param
            optimizer_config.append({"params": [param], "lr": lr})

    optimizer = torch.optim.Adam(optimizer_config)
    return optimization_params, optimizer


def run_reconstruction_single(
    zyx_stack: np.ndarray,
    pixel_scales: tuple[float, float, float],
    fixed_params: dict,
    param_values: dict,
    device: Device = None,
) -> np.ndarray:
    """
    Run a single phase reconstruction with specified parameters (no optimization).

    Parameters
    ----------
    zyx_stack : np.ndarray
        Input Z-stack with shape (Z, Y, X)
    pixel_scales : tuple[float, float, float]
        (z_scale, y_scale, x_scale) in micrometers
    fixed_params : dict
        Fixed reconstruction parameters (wavelength, index, etc.)
    param_values : dict
        Parameter values to use (z_offset, numerical_aperture_detection, etc.)
    device : torch.device | str | None, optional
        Computing device. If None, auto-selects GPU if available, else CPU.

    Returns
    -------
    np.ndarray
        Normalized uint8 array of reconstructed phase image (for display)
    """
    # Resolve device (will print GPU info if available)
    device = get_device(device)

    # Convert to torch tensor on target device
    zyx_tile = torch.tensor(zyx_stack, dtype=torch.float32, device=device)

    # Prepare reconstruction arguments
    z_scale, y_scale, x_scale = pixel_scales
    recon_args = fixed_params.copy()

    # Remove non-reconstruction parameters from fixed_params
    recon_args.pop("num_iterations", None)
    recon_args.pop("use_tiling", None)
    recon_args.pop("device", None)

    recon_args["yx_shape"] = zyx_tile.shape[1:]
    recon_args["yx_pixel_size"] = y_scale
    recon_args["z_scale"] = z_scale

    # Set parameter values (convert to tensors on device)
    for name, value in param_values.items():
        recon_args[name] = torch.tensor([value], dtype=torch.float32, device=device)

    # Run reconstruction
    yx_recon = run_reconstruction(zyx_tile, recon_args)

    # Transfer to CPU and normalize for display
    recon_numpy = yx_recon.detach().cpu().numpy()
    # Wrap in xarray for normalize_for_display (expects xr.DataArray)
    recon_normalized = normalize_for_display(xr.DataArray(recon_numpy))

    return recon_normalized


def run_optimization_streaming(
    zyx_stack: np.ndarray,
    pixel_scales: tuple[float, float, float],
    fixed_params: dict,
    optimizable_params: dict,
    num_iterations: int = 10,
    device: Device = None,
) -> Generator[dict, None, None]:
    """
    Run phase reconstruction optimization with streaming updates.

    Generator that yields reconstruction results and loss after each iteration.
    Supports GPU acceleration for significant speedup (15-25x on typical hardware).

    Parameters
    ----------
    zyx_stack : np.ndarray
        Input Z-stack with shape (Z, Y, X)
    pixel_scales : tuple[float, float, float]
        (z_scale, y_scale, x_scale) in micrometers
    fixed_params : dict
        Fixed reconstruction parameters (wavelength, index, etc.)
    optimizable_params : dict
        Parameters to optimize with (enabled, initial, lr) tuples
    num_iterations : int, optional
        Number of optimization iterations, by default 10
    device : torch.device | str | None, optional
        Computing device. If None, auto-selects GPU if available, else CPU.
        Examples: "cuda", "cpu", "cuda:0", torch.device("cuda")
        By default None

    Yields
    ------
    dict
        Dictionary with keys:
        - 'reconstructed_image': normalized uint8 array (on CPU for display)
        - 'loss': float loss value
        - 'iteration': int iteration number (1-indexed)
        - 'params': dict of current parameter values

    Notes
    -----
    All computation is performed on the specified device (GPU if available).
    Only final results are transferred to CPU for display, minimizing
    transfer overhead.
    """
    # Resolve device (will print GPU info if available)
    device = get_device(device)

    # Convert to torch tensor on target device (single transfer)
    zyx_tile = torch.tensor(zyx_stack, dtype=torch.float32, device=device)

    # Prepare reconstruction arguments
    z_scale, y_scale, x_scale = pixel_scales
    recon_args = fixed_params.copy()

    # Remove non-reconstruction parameters from fixed_params
    recon_args.pop("num_iterations", None)
    recon_args.pop("use_tiling", None)
    recon_args.pop("device", None)  # Remove device if present

    recon_args["yx_shape"] = zyx_tile.shape[1:]
    recon_args["yx_pixel_size"] = y_scale
    recon_args["z_scale"] = z_scale

    # Initialize optimizable parameters on device
    for name, (enabled, initial, lr) in optimizable_params.items():
        recon_args[name] = torch.tensor([initial], dtype=torch.float32, device=device)

    # Prepare optimizer with parameters on device
    optimization_params, optimizer = prepare_optimizer(optimizable_params, device)

    # Optimization loop (all on device)
    for step in range(num_iterations):
        # Update parameters
        for name, param in optimization_params.items():
            recon_args[name] = param

        # Run reconstruction (all on device)
        yx_recon = run_reconstruction(zyx_tile, recon_args)

        # Compute loss (all on device, negative midband power - we want to maximize)
        loss = -compute_midband_power(
            yx_recon,
            NA_det=0.15,
            lambda_ill=recon_args["wavelength_illumination"],
            pixel_size=recon_args["yx_pixel_size"],
            band=(0.1, 0.2),
        )

        # Backward pass and optimizer step (on device)
        loss.backward()
        optimizer.step()
        optimizer.zero_grad()

        # Transfer to CPU ONLY for display (single transfer per iteration)
        recon_numpy = yx_recon.detach().cpu().numpy()
        # Wrap in xarray for normalize_for_display (expects xr.DataArray)
        recon_normalized = normalize_for_display(xr.DataArray(recon_numpy))

        # Extract current parameter values (scalars, already on CPU)
        param_values = {
            name: param.item() for name, param in optimization_params.items()
        }

        # Yield results
        yield {
            "reconstructed_image": recon_normalized,
            "loss": loss.item(),
            "iteration": step + 1,
            "params": param_values,
        }


def extract_tiles(
    zyx_data: np.ndarray, num_tiles: tuple[int, int], overlap_pct: float
) -> tuple[dict[str, np.ndarray], dict[str, tuple[int, int, int]]]:
    """
    Extract overlapping tiles from a Z-stack for processing.

    Parameters
    ----------
    zyx_data : np.ndarray
        Input data with shape (Z, Y, X)
    num_tiles : tuple[int, int]
        Number of tiles in (Y, X) dimensions
    overlap_pct : float
        Overlap percentage between tiles (0.0 to 1.0)

    Returns
    -------
    tuple[dict, dict]
        tiles: dict mapping tile names to arrays
        translations: dict mapping tile names to (z, y, x) positions
    """
    Z, Y, X = zyx_data.shape
    tile_height = int(np.ceil(Y / (num_tiles[0] - (num_tiles[0] - 1) * overlap_pct)))
    tile_width = int(np.ceil(X / (num_tiles[1] - (num_tiles[1] - 1) * overlap_pct)))
    stride_y = int(tile_height * (1 - overlap_pct))
    stride_x = int(tile_width * (1 - overlap_pct))

    tiles = {}
    translations = {}
    for yi in range(num_tiles[0]):
        for xi in range(num_tiles[1]):
            y0, x0 = yi * stride_y, xi * stride_x
            y1, x1 = min(y0 + tile_height, Y), min(x0 + tile_width, X)
            tile_name = f"0/0/{yi:03d}{xi:03d}"
            tiles[tile_name] = zyx_data[:, y0:y1, x0:x1]
            translations[tile_name] = (0, y0, x0)
    return tiles, translations