TensorFlow：将float64张量强制转换为float32

Question

我正在尝试使用：train = optimizer.minimize(loss)，但标准优化器不适用于tf.float64。因此，我想将loss从tf.float64截断为仅tf.float32。

Traceback (most recent call last):
  File "q4.py", line 85, in <module>
    train = optimizer.minimize(loss)
  File "/Library/Python/2.7/site-packages/tensorflow/python/training/optimizer.py", line 190, in minimize
    colocate_gradients_with_ops=colocate_gradients_with_ops)
  File "/Library/Python/2.7/site-packages/tensorflow/python/training/optimizer.py", line 229, in compute_gradients
    self._assert_valid_dtypes([loss])
  File "/Library/Python/2.7/site-packages/tensorflow/python/training/optimizer.py", line 354, in _assert_valid_dtypes
    dtype, t.name, [v for v in valid_dtypes]))
ValueError: Invalid type tf.float64 for Add_1:0, expected: [tf.float32].

Answer 1

简短的回答是，您可以使用tf.cast()操作将张量从tf.float64转换为tf.float32：

loss = tf.cast(loss, tf.float32)

答案越长，这将无法解决优化器的所有问题。 （缺少对tf.float64的支持是known issue。）优化程序要求您尝试优化的所有 tf.Variable 对象也必须具有类型tf.float32。