tf.gather

TensorFlow 1 version View source on GitHub

Gather slices from params axis axis according to indices.

Used in the notebooks

Used in the tutorials

Gather slices from params axis axis according to indices. indices must be an integer tensor of any dimension (usually 0-D or 1-D).

For 0-D (scalar) indices:

$$\begin{align*} output[p_0, ..., p_{axis-1}, && &&& p_{axis + 1}, ..., p_{N-1}] = \\ params[p_0, ..., p_{axis-1}, && indices, &&& p_{axis + 1}, ..., p_{N-1}] \end{align*}$$

Where N = ndims(params).

For 1-D (vector) indices with batch_dims=0:

$$\begin{align*} output[p_0, ..., p_{axis-1}, && &i, &&p_{axis + 1}, ..., p_{N-1}] =\\ params[p_0, ..., p_{axis-1}, && indices[&i], &&p_{axis + 1}, ..., p_{N-1}] \end{align*}$$

In the general case, produces an output tensor where:

$$\begin{align*} output[p_0, &..., p_{axis-1}, & &i_{B}, ..., i_{M-1}, & p_{axis + 1}, &..., p_{N-1}] = \\ params[p_0, &..., p_{axis-1}, & indices[p_0, ..., p_{B-1}, &i_{B}, ..., i_{M-1}], & p_{axis + 1}, &..., p_{N-1}] \end{align*}$$

Where N = ndims(params), M = ndims(indices), and B = batch_dims. Note that params.shape[:batch_dims] must be identical to indices.shape[:batch_dims].

The shape of the output tensor is:

output.shape = params.shape[:axis] + indices.shape[batch_dims:] + params.shape[axis + 1:].

Note that on CPU, if an out of bound index is found, an error is returned. On GPU, if an out of bound index is found, a 0 is stored in the corresponding output value.

See also tf.gather_nd.

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. Must be in range [0, params.shape[axis]).
validate_indices Deprecated, does nothing.
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.