Tensorflow 2.0 中的 tf.function 和 tf.while 循环

Question

I am trying to parallelize loop using tf.while_loop . As suggested here , the parallel_iterations argument doesn't make a difference in the eager mode. So I attempted to wrap tf.while_loop with tf.function . However, after adding the decorator,the behavior of the iteration variable changes.

For example, this piece of code works.

result = np.zeros(10)
iteration = tf.constant(0)
c = lambda i: tf.less(i, 10)
def print_fun(iteration):
    result[iteration] = iteration
    iteration += 1
    return (iteration,)
tf.while_loop(c, print_fun, [iteration])

If I add the decorator, bug occurs.

result = np.zeros(10)
iteration = tf.constant(0)
c = lambda i: tf.less(i, 10)
def print_fun(iteration):
    result[iteration] = iteration
    iteration += 1
    return (iteration,)

@tf.function
def run_graph():
    iteration = tf.constant(0)
    tf.while_loop(c, print_fun, [iteration])

run_graph()

From my debugging process, I found that variable iteration changes from a tensor to a placeholder. Why is that? How should I modify the code to eliminate the bug?

Thanks.

Answer 1

The code in your first snippet (the one without the @tf.function ) takes advantage of TensorFlow 2's eager execution to manipulate a numpy array (ie, your outer iteration object) directly. With @tf.function , this doesn't work because @tf.function tries to compile your code into a tf.Graph, which cannot operate on a numpy array directly (it can only process tensorflow tensors). To get around this issue, use a tf.Variable and keep assigning value into its slices.

With @tf.function , what you are trying to do is actually achievable with simpler code, by taking advantage of @tf.function 's automatic Python-to-graph transformation feature (known as AutoGraph). You just write a normal Python while loop (using tf.less() in lieu of the < operator), and the while loop will be compiled by AutoGraph into a tf.while_loop under the hood.

The code looks something like:

result = tf.Variable(np.zeros([10], dtype=np.int32))

@tf.function
def run_graph():
  i = tf.constant(0, dtype=tf.int32)
  while tf.less(i, 10):
    result[i].assign(i)  # Performance may require tuning here.
    i += 1

run_graph()
print(result.read_value())