Support for unequal prototype distributions

prototorch/components/components.py

@@ -4,8 +4,10 @@ import warnings
 from typing import Tuple
 
 import torch
-from prototorch.components.initializers import (ComponentsInitializer,
-                                                EqualLabelInitializer,
+from prototorch.components.initializers import (ClassAwareInitializer,
+                                                ComponentsInitializer,
+                                                EqualLabelsInitializer,
+                                                UnequalLabelsInitializer,
                                                 ZeroReasoningsInitializer)
 from prototorch.functions.initializers import get_initializer
 from torch.nn.parameter import Parameter
@@ -30,12 +32,15 @@ class Components(torch.nn.Module):
         else:
             self._initialize_components(number_of_components, initializer)
 
-    def _initialize_components(self, number_of_components, initializer):
+    def _precheck_initializer(self, initializer):
         if not isinstance(initializer, ComponentsInitializer):
             emsg = f"`initializer` has to be some subtype of " \
                 f"{ComponentsInitializer}. " \
                 f"You have provided: {initializer=} instead."
             raise TypeError(emsg)
+
+    def _initialize_components(self, number_of_components, initializer):
+        self._precheck_initializer(initializer)
         self._components = Parameter(
             initializer.generate(number_of_components))
 
@@ -57,7 +62,7 @@ class LabeledComponents(Components):
     Every Component has a label assigned.
     """
     def __init__(self,
-                 labels=None,
+                 distribution=None,
                  initializer=None,
                  *,
                  initialized_components=None):
@@ -65,15 +70,27 @@ class LabeledComponents(Components):
             super().__init__(initialized_components=initialized_components[0])
             self._labels = initialized_components[1]
         else:
-            self._initialize_labels(labels)
+            self._initialize_labels(distribution)
             super().__init__(number_of_components=len(self._labels),
                              initializer=initializer)
 
-    def _initialize_labels(self, labels):
-        if type(labels) == tuple:
-            num_classes, prototypes_per_class = labels
-            labels = EqualLabelInitializer(num_classes, prototypes_per_class)
+    def _initialize_components(self, number_of_components, initializer):
+        if isinstance(initializer, ClassAwareInitializer):
+            self._precheck_initializer(initializer)
+            self._components = Parameter(
+                initializer.generate(number_of_components, self.distribution))
+        else:
+            super()._initialize_components(self, number_of_components,
+                                           initializer)
 
+    def _initialize_labels(self, distribution):
+        if type(distribution) == tuple:
+            num_classes, prototypes_per_class = distribution
+            labels = EqualLabelsInitializer(num_classes, prototypes_per_class)
+        elif type(distribution) == list:
+            labels = UnequalLabelsInitializer(distribution)
+
+        self.distribution = labels.distribution
         self._labels = labels.generate()
 
     @property

prototorch/components/initializers.py

@@ -1,6 +1,7 @@
 """ProtoTroch Initializers."""
 import warnings
 from collections.abc import Iterable
+from itertools import chain
 
 import torch
 from torch.utils.data import DataLoader, Dataset
@@ -91,6 +92,15 @@ class ClassAwareInitializer(ComponentsInitializer):
         self.clabels = torch.unique(self.labels)
         self.num_classes = len(self.clabels)
 
+    def _get_samples_from_initializer(self, length, dist):
+        if not dist:
+            per_class = length // self.num_classes
+            dist = self.num_classes * [per_class]
+        samples_list = [
+            init.generate(n) for init, n in zip(self.initializers, dist)
+        ]
+        return torch.vstack(samples_list)
+
 
 class StratifiedMeanInitializer(ClassAwareInitializer):
     def __init__(self, arg):
@@ -102,10 +112,9 @@ class StratifiedMeanInitializer(ClassAwareInitializer):
             class_initializer = MeanInitializer(class_data)
             self.initializers.append(class_initializer)
 
-    def generate(self, length):
-        per_class = length // self.num_classes
-        samples_list = [init.generate(per_class) for init in self.initializers]
-        return torch.vstack(samples_list)
+    def generate(self, length, dist=[]):
+        samples = self._get_samples_from_initializer(length, dist)
+        return samples
 
 
 class StratifiedSelectionInitializer(ClassAwareInitializer):
@@ -126,10 +135,8 @@ class StratifiedSelectionInitializer(ClassAwareInitializer):
         mask = torch.bernoulli(n1) - torch.bernoulli(n2)
         return x + (self.noise * mask)
 
-    def generate(self, length):
-        per_class = length // self.num_classes
-        samples_list = [init.generate(per_class) for init in self.initializers]
-        samples = torch.vstack(samples_list)
+    def generate(self, length, dist=[]):
+        samples = self._get_samples_from_initializer(length, dist)
         if self.noise is not None:
             # samples = self.add_noise(samples)
             samples = samples + self.noise
@@ -142,11 +149,29 @@ class LabelsInitializer:
         raise NotImplementedError("Subclasses should implement this!")
 
 
-class EqualLabelInitializer(LabelsInitializer):
+class UnequalLabelsInitializer(LabelsInitializer):
+    def __init__(self, dist):
+        self.dist = dist
+
+    @property
+    def distribution(self):
+        return self.dist
+
+    def generate(self):
+        clabels = range(len(self.dist))
+        labels = list(chain(*[[i] * n for i, n in zip(clabels, self.dist)]))
+        return torch.tensor(labels)
+
+
+class EqualLabelsInitializer(LabelsInitializer):
     def __init__(self, classes, per_class):
         self.classes = classes
         self.per_class = per_class
 
+    @property
+    def distribution(self):
+        return self.classes * [self.per_class]
+
     def generate(self):
         return torch.arange(self.classes).repeat(self.per_class, 1).T.flatten()