Thinning: fix param_name_2_layer_name

This fix does not change the behavior. 
The previous code worked correctly because 'weights' and '.weight' have the same length.

distiller/thinning.py

@@ -84,7 +84,14 @@ def get_normalized_recipe(recipe):
 
 
 def param_name_2_layer_name(param_name):
-    return param_name[:-len('weights')]
+    """Convert a weights tensor's name to the name of the layer using the tensor.
+    
+    By convention, PyTorch modules name their weights parameters as self.weight
+    (see for example: torch.nn.modules.conv) which means that their fully-qualified 
+    name when enumerating a model's parameters is the modules name followed by '.weight'.
+    We exploit this convention to convert a weights tensor name to the fully-qualified 
+    module name."""
+    return param_name[:-len('.weight')]
 
 
 def directives_equal(d1, d2):