解决 ValueError: Cannot convert a partially known TensorShape to a Tensor: (?, 256)

• 既然说 paritially known，那就想办法直接告诉后台shape到底什么样
• tf.shape 就可以获取当前张量的shape
• 样例代码

class Sampling(layers.Layer):
  def call(self, inputs):
    z_mean, z_log_var = inputs
    batch = tf.shape(z_mean)[0] # 获取第一维
    dim = tf.shape(z_mean)[1] # 获取第二维
    epsilon = tf.keras.backend.random_normal(shape=(batch, dim))
    return z_mean + tf.exp(0.5 * z_log_var) * epsilon

参考