TensorFlow中tf.concat()对同形同值的t1、t2输出不同结果的原因

Great question! The root cause here is how TensorFlow interprets Python nested lists vs. NumPy arrays when passed to tf.concat()—even if their shape and values look identical, the handling logic is totally different:

• How t1 (Python list) is processed
  When you pass the Python nested list t1 to tf.concat(), TensorFlow treats it as a list containing 2 separate tensors:

  • Each sub-list [[0],[1],[2]] and [[3],[4],[5]] gets converted to a tensor with shape [3, 1]
  • tf.concat(t1, 1) then concatenates these two [3,1] tensors along the 1st axis, resulting in a [3,2] tensor (the [[0 3],[1 4],[2 5]] you see)

• How t2 (NumPy array) is processed
  When you pass the NumPy array t2, TensorFlow first converts the entire array into a single 3D tensor with shape [2,3,1].
  Now, tf.concat() expects its first argument to be a list/tuple of tensors. If you pass a single tensor, it automatically wraps it into a list with one element (like tf.concat([t2], 1)). Concatenating a list with only one tensor does nothing—you get back the original tensor, hence the output matching t2's structure exactly.

If you want t2 to produce the same result as t1, you need to split it into the two [3,1] sub-tensors first, then pass them as a list to tf.concat():

print(sess.run(tf.concat([t2[0], t2[1]], 1)))

This will give you the same [3,2] output as the t1 case.

内容的提问来源于stack exchange，提问作者ji cocaer