Added an option to return a latent embedding to the ConvAutoencoder class + unit test for this case

@heusch,

Following our today's discussion, here is the update to the autoencoder class allowing to avoid prediction_function in the configs. Please merge if fine with you. Thanks!

bob/learn/pytorch/architectures/ConvAutoencoder.py

@@ -12,9 +12,22 @@ from torch import nn
 # Define the network:
 
 class ConvAutoencoder(nn.Module):
+    """
+    A class defining a simple convolutional autoencoder.
 
-    def __init__(self):
+    Attributes
+    ----------
+    return_latent_embedding : bool
+        If set to ``True`` forward() method returns a latent
+        emebedding (encoder output), otherwise a reconstructed
+        image is returned. Default: ``False``
+    """
+
+    def __init__(self, return_latent_embedding = False):
         super(ConvAutoencoder, self).__init__()
+
+        self.return_latent_embedding = return_latent_embedding
+
         self.encoder = nn.Sequential(nn.Conv2d(3, 16, 5, padding=2),
                                      nn.ReLU(True),
                                      nn.MaxPool2d(2),
@@ -45,5 +58,9 @@ class ConvAutoencoder(nn.Module):
         """
         x = self.encoder(x)
         x = self.decoder(x)
+
+        if self.return_latent_embedding:
+            return self.encoder(x)
+
         return x