Fix weights access when pruning in pytorch 1.4

Patch based on discussion at:
https://discuss.pytorch.org/t/forward-pre-hooks-not-called-after-applying-nn-uitls-prune-methods/69497

This patch has been tested by Richard Schoonhoven.

Author: Allard Hendriksen

msd_pytorch/msd_block.py

@@ -176,6 +176,6 @@ def forward(self, input):
         # `self.parameters()`, but this returns the original
         # (unmasked) parameters.
         bias = self.bias
-        weights = (self.__getattr__("weight{}".format(i)) for i in range(len(self.weights)))
+        weights = (getattr(self, "weight{}".format(i)) for i in range(len(self.weights)))
 
         return MSDBlockImpl2d.apply(input, self.dilations, bias, *weights)