ranking_fusion.py 文件源码

def _generate_train_model_function(self, scores):
       u = T.lvector('u')
       i = T.lvector('i')
       j = T.lvector('j')
       self.W = theano.shared(numpy.zeros((self._dim)).astype('float32'), name='W');
       self.S = theano.shared(scores, name='S');
       x_ui  = T.dot(self.W, self.S[u,i,:].T);
       x_uj  = T.dot(self.W, self.S[u,j,:].T);
       x_uij = x_ui - x_uj;
       obj = T.sum(
               T.log(T.nnet.sigmoid(x_uij)).sum() - \
               self._lambda_w * 0.5 * (self.W ** 2).sum()
               )
       cost = -obj
       g_cost_W = T.grad(cost=cost, wrt=self.W)
       updates = [
               (self.W, self.W - self._learning_rate * g_cost_W)
               ]
       self.train_model = theano.function(inputs=[u,i,j], outputs=cost, updates=updates);