Repeat body while the condition cond is true. (deprecated argument values)

def sumSquare(n):
  i, result = tf.constant(0), tf.constant(0)
  while i < n: # AutoGraph converts while-loop to tf.while_loop().
    result += i * i
    i += 1
  return result
def sumSquare2(n):
  i, result = tf.constant(0), tf.constant(0)
  c = lambda i, _: tf.less(i, n)
  b = lambda i, result: (i + 1, result + i * i)
  return tf.while_loop(c, b, [i, result])[1]

For more information, see tf.function and AutoGraph guide .

cond is a callable returning a boolean scalar tensor. body is a callable returning a (possibly nested) tuple, namedtuple or list of tensors of the same arity (length and structure) and types as loop_vars. loop_vars is a (possibly nested) tuple, namedtuple or list of tensors that is passed to both cond and body. cond and body both take as many arguments as there are loop_vars.

In addition to regular Tensors or IndexedSlices, the body may accept and return TensorArray objects. The flows of the TensorArray objects will be appropriately forwarded between loops and during gradient calculations.

Note that while_loop calls cond and body exactly once (inside the call to while_loop, and not at all during while_loop stitches together the graph fragments created during the cond and body calls with some additional graph nodes to create the graph flow that repeats body until cond returns false.

For correctness, tf.while_loop() strictly enforces shape invariants for the loop variables. A shape invariant is a (possibly partial) shape that is unchanged across the iterations of the loop. An error will be raised if the shape of a loop variable after an iteration is determined to be more general than or incompatible with its shape invariant. For example, a shape of [11, None] is more general than a shape of [11, 17], and [11, 21] is not compatible with [11, 17]. By default (if the argument shape_invariants is not specified), it is assumed that the initial shape of each tensor in loop_vars is the same in every iteration. The shape_invariants argument allows the caller to specify a less specific shape invariant for each loop variable, which is needed if the shape varies between iterations. The tf.Tensor.set_shape function may also be used in the body function to indicate that the output loop variable has a particular shape. The shape invariant for SparseTensor and IndexedSlices are treated specially as follows:

a) If a loop variable is a SparseTensor, the shape invariant must be TensorShape([r]) where r is the rank of the dense tensor represented by the sparse tensor. It means the shapes of the three tensors of the SparseTensor are ([None], [None, r], [r]). NOTE: The shape invariant here is the shape of the SparseTensor.dense_shape property. It must be the shape of a vector.

b) If a loop variable is an IndexedSlices, the shape invariant must be a shape invariant of the values tensor of the IndexedSlices. It means the shapes of the three tensors of the IndexedSlices are (shape, [shape[0]], [shape.ndims]).

while_loop implements non-strict semantics, enabling multiple iterations to run in parallel. The maximum number of parallel iterations can be controlled by parallel_iterations, which gives users some control over memory consumption and execution order. For correct programs, while_loop should return the same result for any parallel_iterations > 0.

For training, TensorFlow stores the tensors that are produced in the forward inference and are needed in back propagation. These tensors are a main source of memory consumption and often cause OOM errors when training on GPUs. When the flag swap_memory is true, we swap out these tensors from GPU to CPU. This for example allows us to train RNN models with very long sequences and large batches.

cond A callable that represents the termination condition of the loop.
body A callable that represents the loop body.
loop_vars A (possibly nested) tuple, namedtuple or list of numpy array, Tensor, and TensorArray objects.
shape_invariants The shape invariants for the loop variables.
parallel_iterations The number of iterations allowed to run in parallel. It must be a positive integer.
back_prop (optional) Deprecated. False disables support for back propagation. Prefer using tf.stop_gradient instead.
swap_memory Whether GPU-CPU memory swap is enabled for this loop.
maximum_iterations Optional maximum number of iterations of the while loop to run. If provided, the cond output is AND-ed with an additional condition ensuring the number of iterations executed is no greater than maximum_iterations.
name Optional name prefix for the returned tensors.

The output tensors for the loop variables after the loop. The return value has the same structure as loop_vars.

TypeError if cond or body is not callable.
ValueError if loop_vars is empty.


i = tf.constant(0)
c = lambda i: tf.less(i, 10)
b = lambda i: (tf.add(i, 1), )
r = tf.while_loop(c, b, [i])[0]

Example with nesting and a namedtuple:

import collections
Pair = collections.namedtuple('Pair', 'j, k')
ijk_0 = (tf.constant(0), Pair(tf.constant(1), tf.constant(2)))
c = lambda i, p: i < 10
b = lambda i, p: (i + 1, Pair((p.j + p.k), (p.j - p.k)))
ijk_final = tf.while_loop(c, b, ijk_0)[1]
ijk_final[0].numpy(), ijk_final[1].numpy()
(32, 64)

Example using shape_invariants:

i0 = tf.constant(0)
m0 = tf.ones([2, 2])
c = lambda i, m: i < 10
b = lambda i, m: [i+1, tf.concat([m, m], axis=0)]
    c, b, loop_vars=[i0, m0],
    shape_invariants=[i0.get_shape(), tf.TensorShape([None, 2])])[1]
<tf.Tensor: shape=(2048, 2), dtype=float32, numpy=...>

Example which demonstrates non-strict semantics: In the following example, the final value of counter does not depend on x. So the while_loop can increment the counter parallel to updates of x. However, because the loop counter at one loop iteration depends on the value at the previous iteration, the loop counter itself cannot be incremented in parallel. Hence if we just want the final value of the counter (which we print on the line print(, then x will never be incremented, but the counter will be updated on a single thread. Conversely, if we want the value of the output (which we print on the line print(, then the counter may be incremented on its own thread, while x can be incremented in parallel on a separate thread. In the extreme case, it is conceivable that the thread incrementing the counter runs until completion before x is incremented even a single time. The only thing that can never happen is that the thread updating x can never get ahead of the counter thread because the thread incrementing x depends on the value of the counter.

with tf.compat.v1.Session() as sess:
  n = 10
  c = lambda i, x: i < n
  b = lambda i, x: (
      tf.compat.v1.Print(i + 1, [i], "Updating i based on i == "),
      # Let x depend on i
      tf.compat.v1.Print(x + i, [i], "Updating x based on i == "))

  # Make x to be a big matrix so its updating thread would run slowly
  x = tf.zeros([1000, 100], dtype=tf.int32)
  counter = tf.constant(0)
  counter_out, x_out = tf.while_loop(c, b, (counter, x))

  # The following line may increment the counter and x in parallel.
  # The counter thread may get ahead of the x thread, but not the
  # other way around. For example, the log may contain these messages:

... # Updating i based on i == [9] ... # Updating x based on i == [3] ... # ... # meaning that the counter(i) thread is on iteration 9, ... # while the x thread is on iteration 3. ... print( (1000, 100)