GatherND

Description

Collects data slices to the output tensor with the rank of (q-1) + (r - indices_shape[-1]) for the input tensor data with the rank of r ≥ 1 and the input tensor indices with the rank of q ≥ 1. indices is a q-dimensional integer tensor, which can be considered as a q-1-dimensional special tensor consisting of index pairs (each index pair is a one-dimensional tensor whose length is indices_shape[-1], and each index pair points to a slice in data).

Input

• data: input tensor of type int8, uint8, float16, float, double, int32, or int64, with rank r >= 1.
• indices: input tensor of type int64, with rank q >= 1.

Attribute

batch_dims: int, indicating the number of batch processing axes. Currently, this parameter can only be set to 0.

Output

output: output tensor. The rank is (q-1) + (r - indices_shape[-1]), that is, the shape of the output is [indices_shape[0:q-1], self_shape[indices_shape[-1]:r]].

Constraints

The double input is not supported when the atc tool parameter --precision_mode=must_keep_origin_dtype is used.

ONNX Opset Support

Opset v11/v12/v13/v14/v15/v16/v17/v18