24小时热门版块排行榜

返回列表

【悬赏金币】回答本帖问题，作者yang_12345将赠送您 5 个金币

yang_12345

新虫 (初入文坛)

应助: 0 (幼儿园)
金币: 18
帖子: 5
在线: 2.5小时
虫号: 21035355
注册: 2020-02-21
专业: 人工智能与知识工程

[求助] Tensorflow自伴特征分解不成功，输入可能无效

Self-adjoint eigen decomposition was not successful. The input might not be valid(Tensorflow自伴特征分解不成功，输入可能无效)

def cca_loss(outdim_size, use_all_singular_values):
def inner_cca_objective(y_true, y_pred):
      r1 = 1e-4
      r2 = 1e-4
      eps = 1e-12
      print("y_pred:",y_pred.shape[1])
      o1 = o2 = int(y_pred.shape[1] // 2)
      print(y_pred)
      print("aaaa",o1,o2)
      # unpack (separate) the output of networks for view 1 and view 2
      H1 = tf.transpose(y_pred[:, 0

1])
H2 = tf.transpose(y_pred[:, o1

1 + o2])
      m = tf.shape(H1)[1]
      H1bar = H1 - tf.cast(tf.divide(1, m), tf.float32) * tf.matmul(H1, tf.ones([m, m]))
      H2bar = H2 - tf.cast(tf.divide(1, m), tf.float32) * tf.matmul(H2, tf.ones([m, m]))
      SigmaHat12 = tf.cast(tf.divide(1, m - 1), tf.float32) * tf.matmul(H1bar, H2bar, transpose_b=True)  # [dim, dim]
      SigmaHat11 = tf.cast(tf.divide(1, m - 1), tf.float32) * tf.matmul(H1bar, H1bar, transpose_b=True) + r1 * tf.eye(
         o1)
      SigmaHat22 = tf.cast(tf.divide(1, m - 1), tf.float32) * tf.matmul(H2bar, H2bar, transpose_b=True) + r2 * tf.eye(
         o2)
      D1, V1 = tf.self_adjoint_eig(SigmaHat11) #求特征值  特征向量
      D2, V2 = tf.self_adjoint_eig(SigmaHat22)  # Added to increase stability
      posInd1 = tf.where(tf.greater(D1, eps))
      D1 = tf.gather_nd(D1, posInd1)  # get eigen values that are larger than eps  选出大于eps的特征值
      V1 = tf.transpose(tf.nn.embedding_lookup(tf.transpose(V1), tf.squeeze(posInd1)))
      posInd2 = tf.where(tf.greater(D2, eps))
      D2 = tf.gather_nd(D2, posInd2)
      V2 = tf.transpose(tf.nn.embedding_lookup(tf.transpose(V2), tf.squeeze(posInd2)))

      SigmaHat11RootInv = tf.matmul(tf.matmul(V1, tf.diag(D1 ** -0.5)), V1, transpose_b=True)  # [dim, dim]
      SigmaHat22RootInv = tf.matmul(tf.matmul(V2, tf.diag(D2 ** -0.5)), V2, transpose_b=True)

      Tval = tf.matmul(tf.matmul(SigmaHat11RootInv, SigmaHat12), SigmaHat22RootInv)

      print("use_all_singular_values",use_all_singular_values)
      if use_all_singular_values:
         corr = tf.sqrt(tf.trace(tf.matmul(Tval, Tval, transpose_a=True)))
      else:
         [U, V] = tf.self_adjoint_eig(tf.matmul(Tval, Tval, transpose_a=True))
         U = tf.gather_nd(U, tf.where(tf.greater(U, eps)))
         kk = tf.reshape(tf.cast(tf.shape(U), tf.int32), [])
         K = tf.minimum(kk, outdim_size)
         w, _ = tf.nn.top_k(U, k=K)
         corr = tf.reduce_sum(tf.sqrt(w))

      print("corr:",corr)
      return -corr

return inner_cca_objective

回复此楼

» 猜你喜欢

1楼 2020-12-14 14:02:22

已阅回复此楼关注TA 给TA发消息送TA红花 TA的回帖

相关版块跳转我要订阅楼主 yang_12345 的主题更新

返回列表

不应助 确定回帖应助 (注意：应助才可能被奖励，但不允许灌水，必须填写15个字符以上)

普通表情龙兔虎猫