Fix divide-by-zero in example

examples/glvq_iris.py

@@ -23,15 +23,16 @@ class Model(torch.nn.Module):
     def __init__(self, **kwargs):
         """GLVQ model."""
         super().__init__()
-        self.p1 = Prototypes1D(input_dim=2,
-                               prototypes_per_class=3,
-                               nclasses=3,
-                               prototype_initializer='stratified_random',
-                               data=[x_train, y_train])
+        self.proto_layer = Prototypes1D(
+            input_dim=2,
+            prototypes_per_class=3,
+            nclasses=3,
+            prototype_initializer='stratified_random',
+            data=[x_train, y_train])
 
     def forward(self, x):
-        protos = self.p1.prototypes
-        plabels = self.p1.prototype_labels
+        protos = self.proto_layer.prototypes
+        plabels = self.proto_layer.prototype_labels
         dis = euclidean_distance(x, protos)
         return dis, plabels
 
@@ -61,7 +62,10 @@ for epoch in range(70):
     optimizer.step()
 
     # Get the prototypes form the model
-    protos = model.p1.prototypes.data.numpy()
+    protos = model.proto_layer.prototypes.data.numpy()
+    if np.isnan(np.sum(protos)):
+        print(f'Stopping because of `nan` in prototypes.')
+        break
 
     # Visualize the data and the prototypes
     ax = fig.gca()

prototorch/functions/initializers.py

@@ -76,7 +76,11 @@ def stratified_mean(x_train, y_train, prototype_distribution, one_hot=True):
 
 
 @register_initializer
-def stratified_random(x_train, y_train, prototype_distribution, one_hot=True):
+def stratified_random(x_train,
+                      y_train,
+                      prototype_distribution,
+                      one_hot=True,
+                      epsilon=1e-7):
     nprotos = torch.sum(prototype_distribution)
     pdim = x_train.shape[1]
     protos = torch.empty(nprotos, pdim)
@@ -89,7 +93,7 @@ def stratified_random(x_train, y_train, prototype_distribution, one_hot=True):
         xl = x_train[matcher]
         rand_index = torch.zeros(1).long().random_(0, xl.shape[0] - 1)
         random_xl = xl[rand_index]
-        protos[i] = random_xl
+        protos[i] = random_xl + epsilon
     plabels = labels_from(prototype_distribution, one_hot=one_hot)
     return protos, plabels
 

prototorch/modules/losses.py

@@ -15,6 +15,6 @@ class GLVQLoss(torch.nn.Module):
 
     def forward(self, outputs, targets):
         distances, plabels = outputs
-        mu = glvq_loss(distances, targets, plabels)
+        mu = glvq_loss(distances, targets, prototype_labels=plabels)
         batch_loss = self.squashing(mu + self.margin, beta=self.beta)
         return torch.sum(batch_loss, dim=0)