Merge branch 'patch_stack_one_channel' into 'staging'

Patch stack one channel

See merge request !13

model_server/base/models.py

@@ -3,7 +3,7 @@ from math import floor
 
 import numpy as np
 
-from model_server.base.accessors import GenericImageDataAccessor, InMemoryDataAccessor
+from model_server.base.accessors import GenericImageDataAccessor, InMemoryDataAccessor, PatchStack
 
 
 class Model(ABC):
@@ -88,6 +88,16 @@ class InstanceSegmentationModel(ImageToImageModel):
         if not img.shape == mask.shape:
             raise InvalidInputImageError('Expect input image and mask to be the same shape')
 
+    def label_patch_stack(self, img: PatchStack, mask: PatchStack, **kwargs):
+        """
+        Iterative over a patch stack, call inference on each patch
+        :return: PatchStack of same shape in input
+        """
+        res_data = np.zeros(img.shape, dtype='uint16')
+        for i in range(0, img.count): # interpret as PYXCZ
+            res_data[i, :, :, :, :] = self.label_instance_class(img.iat(i), mask.iat(i), **kwargs).data
+        return PatchStack(res_data)
+
 
 class DummySemanticSegmentationModel(SemanticSegmentationModel):
 

model_server/base/roiset.py

@@ -279,7 +279,7 @@ class RoiSet(object):
     def classify_by(self, name: str, channel: int, object_classification_model: InstanceSegmentationModel, ):
 
         # do this on a patch basis, i.e. only one object per frame
-        obmap_patches = object_classification_model.label_instance_class(
+        obmap_patches = object_classification_model.label_patch_stack(
             self.get_raw_patches(channel=channel),
             self.get_patch_masks()
         )

model_server/extensions/ilastik/models.py

@@ -252,7 +252,12 @@ class IlastikObjectClassifierFromPixelPredictionsModel(IlastikModel, ImageToImag
             def label_instance_class(
                     self, img: GenericImageDataAccessor, mask: GenericImageDataAccessor, **kwargs
             ) -> GenericImageDataAccessor:
-                return super().label_instance_class(img, mask, pixel_classification_channel=px_ch)
+                if mask.dtype == 'bool':
+                    norm_mask = 1.0 * mask.data
+                else:
+                    norm_mask = mask.data / np.iinfo(mask.dtype).max
+                norm_mask_acc = InMemoryDataAccessor(norm_mask.astype('float32'))
+                return super().label_instance_class(img, norm_mask_acc, pixel_classification_channel=px_ch)
         return _Mod(params={'project_file': self.project_file})