Switch loss to mean (#67)

Closes Should we use mean rather than sum for batchwise loss reduction? #63

sparse_autoencoder/train/train_autoencoder.py

@@ -88,7 +88,7 @@ def train_autoencoder(
             learned_activations_fired_count.add_(fired.sum(dim=0))
 
             # Backwards pass
-            total_loss.sum().backward()
+            total_loss.mean().backward()
 
             optimizer.step()