tensorflow - 值错误: Shape must be rank 0 but is rank 1 for 'cond_11/Switch' (op: 'Switch' )

Question

实际上我们不能在 if 语句中使用 tf.var 作为 bool 来代替使用 tf.cond。我为规范化输入数据编写了这段代码，但出现了令人困惑的错误，我哪里做错了？

   def global_contrast_normalize(X, scale=1., subtract_mean=True,use_std=False,
       sqrt_bias=0., min_divisor=1e-8):
       mean=tf.reduce_mean(X,axis=1)
       if subtract_mean:
         X = X - mean[:, numpy.newaxis]  # Makes a copy.
      else:
         X = tf.copy.copy(X)
      if X.get_shape()[1]==1:
        #ddof = 0
        mean, var = tf.nn.moments(X, axes=[1])

        normalizers = tf.sqrt(sqrt_bias + var) / scale

      else:
        normalizers = tf.sqrt(sqrt_bias + tf.reduce_sum((X ** 2),axis=1)) / scale
        Normalizers= tf.Variable(normalizers,'float32')
        M=tf.Variable(min_divisor,'float32')

      tf.cond( tf.less_equal(Normalizers,M),lambda:tf.assign(Normalizers,  [1]),lambda:tf.assign(Normalizers,normalizers))
      X /= Normalizers[:, tf.newaxis]  # Does not make a copy.
      return X

错误:

in _call_cpp_shape_fn_impl raise ValueError(err.message)

ValueError: Shape must be rank 0 but is rank 1 for 'cond_11/Switch' (op: 'Switch') with input shapes: [1], [1].

Answer 1

错误指出预期输入是标量(等级 0)，但形状为 ([1],[1])。通常，您可以通过将输入 reshape 为标量值(使用 tf.reshape(Normalizers, []) )来解决此问题。

对于这种情况，您似乎想要根据规范化器是否 <= M 有条件地设置它们的值。tf.where 正是这样做的。

(注意，您不必将标准化器或 min_divisor 转换为 tf.Variable)

tf.where 的用法示例:

def global_contrast_normalize(...):
  ...   
  comparison = tf.less_equal(normalizers,M)
  normalizers = tf.where(comparison, tf.ones_like(normalizers), normalizers
  X /= normalizers[:, tf.newaxis]
  return X