8cef69efb71c8f26931f6e448057942ea2533676,metric_learn/itml.py,ITML,fit,#ITML#Any#Any#Any#Any#,68

Before Change

        _lambda[i] -= alpha
        beta = alpha/(1 - alpha*wtw)
        pos_bhat[i] = 1./((1 / pos_bhat[i]) + (alpha / gamma))
        A += beta * A.dot(np.outer(v,v)).dot(A)

      // update negatives
      vv = self.X[c] - self.X[d]
      for i,v in enumerate(vv):
        wtw = v.dot(A).dot(v)  // scalar
        alpha = min(_lambda[i+num_pos],gamma_proj*(1./neg_bhat[i] - 1./wtw))
        _lambda[i+num_pos] -= alpha
        beta = -alpha/(1 + alpha*wtw)
        neg_bhat[i] = 1./((1 / neg_bhat[i]) - (alpha / gamma))
        A += beta * A.dot(np.outer(v,v)).dot(A)

      normsum = np.linalg.norm(_lambda) + np.linalg.norm(lambdaold)

After Change

        beta = alpha/(1 - alpha*wtw)
        pos_bhat[i] = 1./((1 / pos_bhat[i]) + (alpha / gamma))
        Av = A.dot(v)
        A += beta * np.outer(Av, Av)

      // update negatives
      vv = self.X[c] - self.X[d]
      for i,v in enumerate(vv):
        wtw = v.dot(A).dot(v)  // scalar
        alpha = min(_lambda[i+num_pos],gamma_proj*(1./neg_bhat[i] - 1./wtw))
        _lambda[i+num_pos] -= alpha
        beta = -alpha/(1 + alpha*wtw)
        neg_bhat[i] = 1./((1 / neg_bhat[i]) - (alpha / gamma))
        Av = A.dot(v)
        A += beta * np.outer(Av, Av)

Instances