Transforms elems by applying fn to each element unstacked on axis 0. (deprecated arguments)

Used in the notebooks

Used in the tutorials

See also tf.scan.

map_fn unstacks elems on axis 0 to obtain a sequence of elements; calls fn to transform each element; and then stacks the transformed values back together.

Mapping functions with single-Tensor inputs and outputs

If elems is a single tensor and fn's signature is tf.Tensor->tf.Tensor, then map_fn(fn, elems) is equivalent to tf.stack([fn(elem) for elem in tf.unstack(elems)]). E.g.:

tf.map_fn(fn=lambda t: tf.range(t, t + 3), elems=tf.constant([3, 5, 2]))
<tf.Tensor: shape=(3, 3), dtype=int32, numpy=
  array([[3, 4, 5],
         [5, 6, 7],
         [2, 3, 4]], dtype=int32)>

map_fn(fn, elems).shape = [elems.shape[0]] + fn(elems[0]).shape.

Mapping functions with multi-arity inputs and outputs

map_fn also supports functions with multi-arity inputs and outputs:

  • If elems is a tuple (or nested structure) of tensors, then those tensors must all have the same outer-dimension size (num_elems); and fn is used to transform each tuple (or structure) of corresponding slices from elems. E.g., if elems is a tuple (t1, t2, t3), then fn is used to transform each tuple of slices (t1[i], t2[i], t3[i]) (where 0 <= i < num_elems).

  • If fn returns a tuple (or nested structure) of tensors, then the result is formed by stacking corresponding elements from those structures.

Specifying fn's output signature

If fn's input and output signatures are different, then the output signature must be specified using fn_output_signature. (The input and output signatures are differ if their structures, dtypes, or tensor types do not match). E.g.:

tf.map_fn(fn=tf.strings.length,  # input & output have different dtypes
          elems=tf.constant(["hello", "moon"]),
<tf.Tensor: shape=(2,), dtype=int32, numpy=array([5, 4], dtype=int32)>
tf.map_fn(fn=tf.strings.join,  # input & output have different structures
          elems=[tf.constant(['The', 'A']), tf.constant(['Dog', 'Cat'])],
<tf.Tensor: shape=(2,), dtype=string,
 numpy=array([b'TheDog', b'ACat'], dtype=object)>

fn_output_signature can be specified using any of the following:


map_fn supports tf.RaggedTensor inputs and outputs. In particular:

  • If elems is a RaggedTensor, then fn will be called with each row of that ragged tensor.

    • If elems has only one ragged dimension, then the values passed to fn will be tf.Tensors.
    • If elems has multiple ragged dimensions, then the values passed to fn will be tf.RaggedTensors with one fewer ragged dimension.
  • If the result of map_fn should be a RaggedTensor, then use a tf.RaggedTensorSpec to specify fn_output_signature.

    • If fn returns tf.Tensors with varying sizes, then use a tf.RaggedTensorSpec with ragged_rank=0 to combine them into a single ragged tensor (which will have ragged_rank=1).
    • If fn returns