tntn

commit 5b3138429c544454203f99b6ac938d3937e851ba
parent 6ccff23ecf2f6f601ea69d3fef7eec1821e74a07
Author: Étienne Simon <esimon@esimon.eu>
Date:   Thu, 10 Apr 2014 12:02:44 +0200

Fix bias broadcasting

Diffstat:
M
model.py
 | 
6
++++--

1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/model.py b/model.py
@@ -1,4 +1,5 @@
 #!/usr/bin/env python2
+
 import numpy
 import scipy
 import theano
@@ -40,14 +41,15 @@ class Relation(object):
         self.W = ip(tag+".W", (n_in, n_in, n_hid), wbound)
         self.V = ip(tag+".V", (n_hid, n_in*2), vbound)
         self.u = ip(tag+".u", (n_hid,), ubound)
-        self.b = theano.shared(name=tag+".b", value=numpy.zeros(shape=(n_hid,1), dtype=theano.config.floatX), broadcastable=[False, True])
+        self.b = theano.shared(name=tag+".b", value=numpy.zeros(shape=(n_hid,), dtype=theano.config.floatX))
         self.params = [ self.W, self.V, self.u, self.b ]
 
     def score(self, inputl, inputr):
         """ Compute the score on given embeddings. """
         bilinear = ((inputr.transpose().reshape((inputr.shape[1], inputr.shape[0], 1))) * T.tensordot(inputl, self.W, axes=([0], [0]))).sum(1).transpose()
         linear = T.dot(self.V, T.concatenate([inputl, inputr]))
-        return T.dot(self.u, self.act(bilinear + linear + self.b))
+        bias = self.b.dimshuffle(0, 'x')
+        return T.dot(self.u, self.act(bilinear + linear + bias))
 
     def regularizer(self):
         """ Compute the squared L2-norm of the relation's parameters. """