如何动态更新 tf.ones_like() 的形状？

Question

我在 tensorflow 中的权重矩阵维度遇到问题。

#outputs = tf.reshape(outputs, [batch_size, seq_length, num_classes])

outputs = tf.reshape(outputs, [-1, seq_length, num_classes])

output_dim = outputs.get_shape().as_list()

weights = tf.ones([output_dim[0], seq_length], tf.int32) #TODO: change the dimension

sequence_loss = tf.contrib.seq2seq.sequence_loss(logits=outputs, targets=Y, weights=weights)

所以，我有一个在最后一个 epoch 发生变化的 batch_size，而权重的维度在最后一个 epoch 时会引起麻烦。

weights = tf.ones([output_dim[0], seq_length], tf.int32) 导致以下错误：

"Cannot convert a partially known TensorShape to a Tensor: %s" % s)
ValueError: Cannot convert a partially known TensorShape to a Tensor: (?, 25)

你会如何解决这个问题？我尝试使用tf.ones_like(outputs)，但这似乎不起作用，tf.ones 似乎需要一个固定值作为其维度。

Answer 1

使用tf.fill，支持动态形状：

a = tf.placeholder(tf.float32, shape=[None, 25, 10])
b = tf.fill(tf.shape(a)[:-1], 1)  # shape=[None, 25]

with tf.Session() as sess:
  print(sess.run(b, feed_dict={a: np.zeros([10, 25, 10])}))

# Prints:
# [[1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
#  [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
#  [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
#  [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
#  [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
#  [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
#  [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
#  [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
#  [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
#  [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]]