[BUG] NaN when training with selection initializer

How to reproduce:
Run the `glvq_spiral.py` file under `examples/`.

The error seems to occur when using a lot of prototypes in combination with the
`StratifiedSelectionInitializer`. Using only a prototype per class, or using
another initializer like the `StratifiedMeanInitializer` seems to make the
problem go away.

examples/glvq_spiral.py

@@ -0,0 +1,56 @@
+"""GLVQ example using the spiral dataset."""
+
+import pytorch_lightning as pl
+import torch
+from prototorch.components import initializers as cinit
+from prototorch.datasets.abstract import NumpyDataset
+from prototorch.datasets.spiral import make_spiral
+from prototorch.models.callbacks.visualization import VisGLVQ2D
+from prototorch.models.glvq import GLVQ
+from torch.utils.data import DataLoader
+
+
+class StopOnNaN(pl.Callback):
+    def __init__(self, param):
+        super().__init__()
+        self.param = param
+
+    def on_epoch_end(self, trainer, pl_module, logs={}):
+        if torch.isnan(self.param).any():
+            raise ValueError("NaN encountered. Stopping.")
+
+
+if __name__ == "__main__":
+    # Dataset
+    x_train, y_train = make_spiral(n_samples=600, noise=0.6)
+    train_ds = NumpyDataset(x_train, y_train)
+
+    # Dataloaders
+    train_loader = DataLoader(train_ds, num_workers=0, batch_size=256)
+
+    # Hyperparameters
+    hparams = dict(
+        nclasses=2,
+        prototypes_per_class=20,
+        # prototype_initializer=cinit.SSI(torch.Tensor(x_train),
+        prototype_initializer=cinit.SMI(torch.Tensor(x_train),
+                                        torch.Tensor(y_train)),
+        lr=0.01,
+    )
+
+    # Initialize the model
+    model = GLVQ(hparams)
+
+    # Callbacks
+    vis = VisGLVQ2D(x_train, y_train)
+    # vis = VisGLVQ2D(x_train, y_train, show_last_only=True, block=True)
+    snan = StopOnNaN(model.proto_layer.components)
+
+    # Setup trainer
+    trainer = pl.Trainer(
+        max_epochs=200,
+        callbacks=[vis, snan],
+    )
+
+    # Training loop
+    trainer.fit(model, train_loader)

prototorch/models/callbacks/visualization.py

@@ -261,20 +261,29 @@ class VisPointProtos(VisWeights):
         self._show_and_save(epoch)
 
 
-class VisGLVQ2D(pl.Callback):
+class Vis2DAbstract(pl.Callback):
     def __init__(self,
                  x_train,
                  y_train,
                  title="Prototype Visualization",
-                 cmap="viridis"):
+                 cmap="viridis",
+                 show_last_only=False,
+                 block=False):
         super().__init__()
         self.x_train = x_train
         self.y_train = y_train
         self.title = title
         self.fig = plt.figure(self.title)
         self.cmap = cmap
+        self.show_last_only = show_last_only
+        self.block = block
 
+
+class VisGLVQ2D(Vis2DAbstract):
     def on_epoch_end(self, trainer, pl_module):
+        if self.show_last_only:
+            if trainer.current_epoch != trainer.max_epochs - 1:
+                return
         protos = pl_module.prototypes
         plabels = pl_module.prototype_labels
         x_train, y_train = self.x_train, self.y_train
@@ -306,22 +315,13 @@ class VisGLVQ2D(pl.Callback):
         ax.contourf(xx, yy, y_pred, cmap=self.cmap, alpha=0.35)
         ax.set_xlim(left=x_min + 0, right=x_max - 0)
         ax.set_ylim(bottom=y_min + 0, top=y_max - 0)
-        plt.pause(0.1)
+        if not self.block:
+            plt.pause(0.01)
+        else:
+            plt.show(block=True)
 
 
-class VisSiameseGLVQ2D(pl.Callback):
+class VisSiameseGLVQ2D(Vis2DAbstract):
-    def __init__(self,
-                 x_train,
-                 y_train,
-                 title="Prototype Visualization",
-                 cmap="viridis"):
-        super().__init__()
-        self.x_train = x_train
-        self.y_train = y_train
-        self.title = title
-        self.fig = plt.figure(self.title)
-        self.cmap = cmap
-
     def on_epoch_end(self, trainer, pl_module):
         protos = pl_module.prototypes
         plabels = pl_module.prototype_labels
@@ -361,22 +361,14 @@ class VisSiameseGLVQ2D(pl.Callback):
             global_step=trainer.current_epoch,
             close=False,
         )
-        plt.pause(0.1)
+
+        if not self.block:
+            plt.pause(0.01)
+        else:
+            plt.show(block=True)
 
 
-class VisNG2D(pl.Callback):
+class VisNG2D(Vis2DAbstract):
-    def __init__(self,
-                 x_train,
-                 y_train,
-                 title="Neural Gas Visualization",
-                 cmap="viridis"):
-        super().__init__()
-        self.x_train = x_train
-        self.y_train = y_train
-        self.title = title
-        self.fig = plt.figure(self.title)
-        self.cmap = cmap
-
     def on_epoch_end(self, trainer, pl_module):
         protos = pl_module.prototypes
         cmat = pl_module.topology_layer.cmat.cpu().numpy()
@@ -410,4 +402,7 @@ class VisNG2D(pl.Callback):
                         "k-",
                     )
 
+        if not self.block:
             plt.pause(0.01)
+        else:
+            plt.show(block=True)