Add multi-frame micrograph support to predict and erase CLIs (#29)

* add multi-frame micrograph support to fidder

* add multi-frame support to fidder erase

* force float in erase cli

* remove stupid bug

* pack properly

* pack properly again

* add fix from andriko

src/fidder/erase/cli.py

@@ -1,11 +1,12 @@
 from pathlib import Path
 
+import einops
 import mrcfile
 import numpy as np
 import torch
 from typer import Option
 
-from .erase import erase_masked_region
+from .erase import erase_masked_region as _erase_masked_region
 from ..utils import get_pixel_spacing_from_header
 from .._cli import cli, OPTION_PROMPT_KWARGS as PKWARGS
 
@@ -29,18 +30,27 @@ def erase_masked_region(
     ),
 ):
     """Erase a masked region in a cryo-EM image."""
-    image = torch.as_tensor(mrcfile.read(input_image)).squeeze()
-    mask = torch.as_tensor(mrcfile.read(input_mask), dtype=torch.bool).squeeze()
-    inpainted_image = erase_masked_region(
-        image=image,
-        mask=mask,
-        background_intensity_model_resolution=(8, 8),
-        background_intensity_model_samples=25000,
-    )
+    images = torch.as_tensor(mrcfile.read(input_image)).squeeze().float()
+    masks = torch.as_tensor(mrcfile.read(input_mask), dtype=torch.bool).squeeze()
+    if images.shape != masks.shape:
+        raise ValueError('Shape mismatch between data in image and mask files.')
+    images, ps = einops.pack([images], pattern='* h w')
+    masks, ps = einops.pack([masks], pattern='* h w')
+
+    erased_images = torch.empty_like(images, dtype=torch.float32)
+    for idx, (image, mask) in enumerate(zip(images, masks)):
+        _erased_image = _erase_masked_region(
+            image=image,
+            mask=mask,
+            background_intensity_model_resolution=(8, 8),
+            background_intensity_model_samples=25000,
+        )
+        erased_images[idx] = _erased_image
+    [erased_images] = einops.unpack(erased_images, pattern='* h w', packed_shapes=ps)
     pixel_spacing = get_pixel_spacing_from_header(input_image)
     mrcfile.write(
         name=output_image,
-        data=np.array(inpainted_image, dtype=np.float32),
+        data=np.array(erased_images, dtype=np.float32),
         voxel_size=pixel_spacing,
         overwrite=True,
     )

src/fidder/predict/cli.py

@@ -1,6 +1,7 @@
 from pathlib import Path
 from typing import Optional
 
+import einops
 import mrcfile
 import numpy as np
 import torch
@@ -36,20 +37,29 @@ def predict_fiducial_mask(
     ),
 ):
     """Predict a fiducial mask using a pretrained model."""
-    image = torch.tensor(mrcfile.read(input_image))
+    images = torch.tensor(mrcfile.read(input_image)).float()
+    images, ps = einops.pack([images], pattern='* h w')
     if pixel_spacing is None:
         pixel_spacing = get_pixel_spacing_from_header(input_image)
-    mask, probabilities = _predict_fiducial_mask(
-        image=image,
-        pixel_spacing=pixel_spacing,
-        probability_threshold=probability_threshold,
-        model_checkpoint_file=model_checkpoint_file,
-    )
-    mask = mask.cpu().numpy().astype(np.int8)
-    probabilities = probabilities.cpu().numpy()
+
+    masks = torch.empty_like(images, dtype=torch.int8)
+    probabilities = torch.empty_like(images)
+    for idx, image in enumerate(images):
+        _mask, _probabilities = _predict_fiducial_mask(
+            image=image,
+            pixel_spacing=pixel_spacing,
+            probability_threshold=probability_threshold,
+            model_checkpoint_file=model_checkpoint_file,
+        )
+        masks[idx] = _mask
+        probabilities[idx] = _probabilities
+    masks = masks.cpu().numpy().astype(np.int8)
+    probabilities = probabilities.float().cpu().numpy()
+    [masks] = einops.unpack(masks, pattern='* h w', packed_shapes=ps)
+    [probabilities] = einops.unpack(probabilities, pattern='* h w', packed_shapes=ps)
     output_pixel_spacing = (1, pixel_spacing, pixel_spacing)
     mrcfile.write(
-        name=output_mask, data=mask, voxel_size=output_pixel_spacing, overwrite=True
+        name=output_mask, data=masks, voxel_size=output_pixel_spacing, overwrite=True
     )
     if output_probabilities is not None:
         probabilities = probabilities.astype(np.float32)