Merge pull request #60 from JonasHell/main

Add sampler that checks if there are a minimum number of instances in the volume

torch_em/data/sampler.py

@@ -29,3 +29,30 @@ def __call__(self, x, y=None):
             return True
         else:
             return np.random.rand() > self.p_reject
+
+
+class MinInstanceSampler:
+    def __init__(self, min_num_instances=2, p_reject=1.0):
+        self.min_num_instances = min_num_instances
+        self.p_reject = p_reject
+
+    def __call__(self, x, y):
+        uniques = np.unique(y)
+        if len(uniques) >= self.min_num_instances:
+            return True
+        else:
+            return np.random.rand() > self.p_reject
+
+
+class MinTwoInstanceSampler:
+    # for the case of min_num_instances=2 this is roughly 10x faster
+    # than using MinInstanceSampler since np.unique is slow
+    def __init__(self, p_reject=1.0):
+        self.p_reject = p_reject
+
+    def __call__(self, x, y):
+        sample_value = y.flat[0]
+        if (y != sample_value).any():
+            return True
+        else:
+            return np.random.rand() > self.p_reject

torch_em/shallow2deep/prepare_shallow2deep.py

@@ -88,7 +88,8 @@ def _load_rf_segmentation_dataset(
     raw_paths, raw_key, label_paths, label_key, patch_shape_min, patch_shape_max, **kwargs
 ):
     rois = kwargs.pop("rois", None)
-    sampler = torch_em.data.MinForegroundSampler(min_fraction=0.01)
+    sampler = kwargs.pop("sampler", None)
+    sampler = sampler if sampler else torch_em.data.MinForegroundSampler(min_fraction=0.01)
     if isinstance(raw_paths, str):
         if rois is not None:
             assert len(rois) == 3 and all(isinstance(roi, slice) for roi in rois)
@@ -299,6 +300,7 @@ def prepare_shallow2deep(
     is_seg_dataset=None,
     balance_labels=True,
     filter_config=None,
+    sampler=None,
     **rf_kwargs,
 ):
     assert len(patch_shape_min) == len(patch_shape_max)
@@ -312,7 +314,7 @@ def prepare_shallow2deep(
         ds = _load_rf_segmentation_dataset(raw_paths, raw_key, label_paths, label_key,
                                            patch_shape_min, patch_shape_max,
                                            raw_transform=raw_transform, label_transform=label_transform,
-                                           rois=rois, n_samples=n_forests)
+                                           rois=rois, n_samples=n_forests, sampler=sampler)
     else:
         ds = _load_rf_image_collection_dataset(raw_paths, raw_key, label_paths, label_key,
                                                patch_shape_min, patch_shape_max, roi=rois,