tf.gather

Gather slices from params axis axis according to indices. (deprecated arguments)

Gather slices from params axis axis according to indices. indices must be an integer tensor of any dimension (often 1-D).

Tensor.getitem works for scalars, tf.newaxis, and python slices

tf.gather extends indexing to handle tensors of indices.

In the simplest case it's identical to scalar indexing:

params = tf.constant(['p0', 'p1', 'p2', 'p3', 'p4', 'p5'])
params[3].numpy()
b'p3'
tf.gather(params, 3).numpy()
b'p3'

The most common case is to pass a single axis tensor of indices (this can't be expressed as a python slice because the indices are not sequential):

indices = [2, 0, 2, 5]
tf.gather(params, indices).numpy()
array([b'p2', b'p0', b'p2', b'p5'], dtype=object)

The indices can have any shape. When the params has 1 axis, the output shape is equal to the input shape:

tf.gather(params, [[2, 0], [2, 5]]).numpy()
array([[b'p2', b'p0'],
       [b'p2', b'p5']], dtype=object)

The params may also have any shape. gather can select slices across any axis depending on the axis argument (which defaults to 0). Below it is used to gather first rows, then columns from a matrix:

params = tf.constant([[0, 1.0, 2.0],
                      [10.0, 11.0, 12.0],
                      [20.0, 21.0, 22.0],
                      [30.0, 31.0, 32.0]])
tf.gather(params, indices=[3,1]).numpy()
array([[30., 31., 32.],
       [10., 11., 12.]], dtype=float32)
tf.gather(params, indices=[2,1], axis=1).numpy()
array([[ 2.,  1.],
       [12., 11.],
       [22., 21.],
       [32., 31.]], dtype=float32)

More generally: The output shape has the same shape as the input, with the indexed-axis replaced by the shape of the indices.

def result_shape(p_shape, i_shape, axis=0):
  return p_shape[:axis] + i_shape + p_shape[axis+1:]

result_shape([1, 2, 3], [], axis=1)
[1, 3]
result_shape([1, 2, 3], [7], axis=1)
[1, 7, 3]
result_shape([1, 2, 3], [7, 5], axis=1)
[1, 7, 5, 3]

Here are some examples:

params.shape.as_list()
[4, 3]
indices = tf.constant([[0, 2]])
tf.gather(params, indices=indices, axis=0).shape.as_list()
[1, 2, 3]
tf.gather(params, indices=indices, axis=1).shape.as_list()
[4, 1, 2]
params = tf.random.normal(shape=(5, 6, 7, 8))
indices = tf.random.uniform(shape=(10, 11), maxval=7, dtype=tf.int32)
result = tf.gather(params, indices, axis=2)
result.shape.as_list()
[5, 6, 10, 11, 8]

This is because each index takes a slice from params, and places it at the corresponding location in the output. For the above example

# For any location in indices
a, b = 0, 1
tf.reduce_all(
    # the corresponding slice of the result
    result[:, :, a, b, :] ==
    # is equal to the slice of `params` along `axis` at the index.
    params[:, :, indices[a, b], :]
).numpy()
True

Batching:

The batch_dims argument lets you gather different items from each element of a batch.

Using batch_dims=1 is equivalent to having an outer loop over the first axis of params and indices:

params = tf.constant([
    [0, 0, 1, 0, 2],
    [3, 0, 0, 0, 4],
    [0, 5, 0, 6, 0]])
indices = tf.constant([
    [2, 4],
    [0, 4],
    [1, 3]])
tf.gather(params, indices, axis=1, batch_dims=1).numpy()
array([[1, 2],
       [3, 4],
       [5, 6]], dtype=int32)

This is is equivalent to:

def manually_batched_gather(params, indices, axis):
  batch_dims=1
  result = []
  for p,i in zip(params, indices):
    r = tf.gather(p, i, axis=axis-batch_dims)
    result.append(r)
  return tf.stack(result)
manually_batched_gather(params, indices, axis=1).numpy()
array([[1, 2],
       [3, 4],
       [5, 6]], dtype=int32)

Higher values of batch_dims are equivalent to multiple nested loops over the outer axes of params and indices. So the overall shape function is

def batched_result_shape(p_shape, i_shape, axis=0, batch_dims=0):
  return p_shape[:axis] + i_shape[batch_dims:] + p_shape[axis+1:]

batched_result_shape(
    p_shape=params.shape.as_list(),
    i_shape=indices.shape.as_list(),
    axis=1,
    batch_dims=1)
[3, 2]
tf.gather(params, indices, axis=1, batch_dims=1).shape.as_list()
[3, 2]

This comes up naturally if you need to use the indices of an operation like tf.argsort, or tf.math.top_k where the last dimension of the indices indexes into the last dimension of input, at the corresponding location. In this case you can use tf.gather(values, indices, batch_dims=-1).

See also:

  • tf.Tensor.getitem: The direct tensor index operation (t[]), handles scalars and python-slices tensor[..., 7, 1:-1]
  • tf.scatter: A collection of operations similar to __setitem__ (t[i] = x)
  • tf.gather_nd: An operation similar to tf.gather but gathers across multiple axis at once (it can gather elements of a matrix instead of rows or columns)
  • tf.boolean_mask, tf.where: Binary indexing.
  • tf.slice and tf.strided_slice: For lower level access to the implementation of __getitem__'s python-slice handling (t[1:-1:2])

params The Tensor from which to gather values. Must be at least rank axis + 1.
indices The index Tensor. Must be one of the following types: int32, int64. The values must be in range [0, params.shape[axis]).
validate_indices Deprecated, does nothing. Indices are always validated on CPU, never validated on GPU.

axis A Tensor. Must be one of the following types: int32, int64. The axis in params to gather indices from. Must be greater than or equal to batch_dims. Defaults to the first non-batch dimension. Supports negative indexes.
batch_dims An integer. The number of batch dimensions. Must be less than or equal to rank(indices).
name A name for the operation (optional).

A Tensor. Has the same type as params.