Prevent use-before-init bug in LSML

As identified in gh-62.

Author: perimosocordiae

metric_learn/lsml.py

@@ -68,9 +68,11 @@ def fit(self, X, constraints, weights=None):
         scale factor for each constraint
     """
     self._prepare_inputs(X, constraints, weights)
+    step_sizes = np.logspace(-10, 0, 10)
     prior_inv = scipy.linalg.inv(self.M_)
+    # Keep track of the best step size and the loss at that step.
+    l_best = 0
     s_best = self._total_loss(self.M_, prior_inv)
-    step_sizes = np.logspace(-10, 0, 10)
     if self.verbose:
       print('initial loss', s_best)
     for it in xrange(1, self.max_iter+1):