transform

commit b1e638bb2a94157b60866c12a0538457b83d6c59
parent 754e753c3b20c82954d5e60d953aa20c01410ab5
Author: Étienne Simon <esimon@esimon.eu>
Date:   Fri, 18 Apr 2014 13:59:36 +0200

Add embeddings normaliser

Diffstat:
M
embeddings.py
 | 
4
++++
M
model.py
 | 
2
++

2 files changed, 6 insertions(+), 0 deletions(-)
diff --git a/embeddings.py b/embeddings.py
@@ -47,3 +47,7 @@ class Embeddings(object):
         learning_rate -- The learning rate used for gradient descent.
         """
         return [(self.E, self.E - learning_rate * T.grad(cost=cost, wrt=self.E))]
+
+    def normalise_updates(self):
+        """ Normalise the embeddings' L2 norm to 1. """
+        return [(self.E, self.E / T.sqrt(T.sum(self.E **2, axis=1)]
diff --git a/model.py b/model.py
@@ -89,6 +89,7 @@ class Model(object):
         criterion = T.mean(violating_margin*score)
 
         self.train_function = theano.function(inputs=list(inputs), outputs=[criterion], updates=self.updates(criterion))
+        self.normalise_function = theano.function(inputs=[], outputs=[], updates=self.embeddings.normalise_updates())
 
         relation = T.addbroadcast(relation, 0)
         broadcasted_left = T.addbroadcast(positive_left, 0)
@@ -124,6 +125,7 @@ class Model(object):
             for (relation, left_positive, right_positive, left_negative, right_negative) in self.dataset.training_minibatch(batch_size):
                 c1=self.train_function(relation, left_positive, right_positive, left_positive, right_negative)
                 c2=self.train_function(relation, left_positive, right_positive, left_negative, right_positive)
+                self.normalise_function()
 
     def error(self, name):
         """ Compute the mean rank and top 10 on a given data. """