GPU deskew

mantis/analysis/deskew.py

@@ -1,5 +1,7 @@
 import numpy as np
-import scipy
+import torch
+
+from monai.transforms.spatial.array import Affine
 
 
 def _average_n_slices(data, average_window_width=1):
@@ -139,53 +141,47 @@ def deskew_data(
         If false, only compute the deskewed volume within a cuboid region.
     average_n_slices : int, optional
         after deskewing, averages every n slices (default = 1 applies no averaging)
-    order : int, optional
-        interpolation order (default 1 is linear interpolation)
-    cval : float, optional
-        fill value area outside of the measured volume (default None fills
-        with the minimum value of the input array)
     Returns
     -------
     deskewed_data : NDArray with ndim == 3
         axis 0 is the Z axis, normal to the coverslip
         axis 1 is the Y axis, input axis 2 in the plane of the coverslip
         axis 2 is the X axis, the scanning axis
     """
-    if cval is None:
-        cval = np.min(np.ravel(raw_data))
-
     # Prepare transforms
     Z, Y, X = raw_data.shape
 
     ct = np.cos(ls_angle_deg * np.pi / 180)
-    Z_shift = 0
-    if not keep_overhang:
-        Z_shift = int(np.floor(Y * ct * px_to_scan_ratio))
 
     matrix = np.array(
         [
             [
                 -px_to_scan_ratio * ct,
                 0,
                 px_to_scan_ratio,
-                Z_shift,
+                0,
             ],
-            [-1, 0, 0, Y - 1],
-            [0, -1, 0, X - 1],
+            [-1, 0, 0, 0],
+            [0, -1, 0, 0],
+            [0, 0, 0, 1],
         ]
     )
     output_shape, _ = get_deskewed_data_shape(
         raw_data.shape, ls_angle_deg, px_to_scan_ratio, keep_overhang
     )
 
-    # Apply transforms
-    deskewed_data = scipy.ndimage.affine_transform(
-        raw_data,
-        matrix,
-        output_shape=output_shape,
-        order=order,
-        cval=cval,
-    )
+    # to tensor on GPU
+    if torch.cuda.is_available():
+        raw_data = torch.tensor(raw_data).to("cuda")
+
+    # Returns callable
+    affine_func = Affine(affine=matrix, padding_mode="zeros", image_only=True)
+
+    # affine_func accepts CZYX array, so for ZYX input we need [None] and for ZYX output we need [0]
+    deskewed_data = affine_func(raw_data[None], mode="bilinear", spatial_size=output_shape)[0]
+
+    # to numpy array on CPU
+    deskewed_data = deskewed_data.cpu().numpy()
 
     # Apply averaging
     averaged_deskewed_data = _average_n_slices(

mantis/cli/deskew.py

@@ -4,6 +4,7 @@
 from typing import List
 
 import click
+import torch
 
 from iohub.ngff import open_ome_zarr
 
@@ -13,6 +14,10 @@
 from mantis.cli.parsing import config_filepath, input_position_dirpaths, output_dirpath
 from mantis.cli.utils import yaml_to_model
 
+# Needed for multiprocessing with GPUs
+# https://github.com/pytorch/pytorch/issues/40403#issuecomment-1422625325
+torch.multiprocessing.set_start_method('spawn', force=True)
+
 
 @click.command()
 @input_position_dirpaths()

pyproject.toml

@@ -30,6 +30,7 @@ dependencies = [
   "ndtiff>=2.0",
   "nidaqmx",
   "numpy",
+  "monai",
   "pandas~=2.1",
   "pycromanager==0.28.1",
   "pydantic",