SparseReshape

English | 简体中文

Contribution Description

Contributor	Contributed Operator	Contribution Date	Contribution Content

Product Support Status

Product	Supported
Ascend 950PR&950DT products	√
Atlas A3 products	√
Atlas A2 products	√
Atlas 200I/500 A2 inference products	×
Atlas inference products	×
Atlas training products	×

Function Description

• Operator function: reshapes the indices of a sparse tensor (SparseTensor) from the input shape to the target shape. A sparse tensor consists of indices (coordinates of non-zero elements), values (values of non-zero elements), and shape (dense shape). SparseReshape reshapes only the indices and shape, and does not modify the values.

• Computation formula:

$$\text{flat\_id}=\sum _{j=0}^{\text{input\_rank}-1}\text{indices}[i,j]\times \text{input\_strides}[j]$$

$$\text{y\_indices}[i,j]=\lfloor \text{flat\_id}/\text{output\_strides}[j]\rfloor ,\text{flat\_id}=\text{flat\_id}\mathrm{m}\mathrm{o}\mathrm{d}\text{output\_strides}[j]$$

Parameter Description

Parameter Name	Input/Output/Attribute	Description	Data Type	Data Format
indices	Input	Multi-dimensional coordinate matrix of non-zero elements. This parameter is a 2D tensor with the shape (nnz, input_rank).	INT32, INT64	ND
shape	Input	Original dense shape. This parameter is a 1D tensor with the shape (input_rank,).	INT32, INT64	ND
new_shape	Input	Target dense shape. This parameter is a 1D tensor with the shape (output_rank,). One value can be -1 and is inferred automatically.	INT32, INT64	ND
y_indices	Output	Reshaped multi-dimensional coordinate matrix of non-zero elements. This parameter is a 2D tensor with the shape (nnz, output_rank).	INT32, INT64	ND
y_shape	Output	Reshaped dense shape. This parameter is a 1D tensor with the shape (output_rank,).	INT32, INT64	ND

Constraint Description

• The input indices must be a 2D matrix, and shape and new_shape must be 1D vectors.
• Each row of the input indices contains the multi-dimensional coordinates of a non-zero element. The value of each element must be greater than or equal to 0 and less than the size of the corresponding dimension in shape.
• Both input_rank and output_rank do not exceed 8 dimensions (MAX_RANK=8).
• At most one value in new_shape can be -1. This value is inferred automatically by dividing the product of the values of the shape parameter by the product of the other non--1 values of the new_shape parameter.
• The product of the values of the input shape must be the same as the product of the values of the input new_shape.
• The dtypes of all inputs and outputs must be the same (all int32 or all int64).

Invocation Description

Invocation Method	Invocation Sample	Description
Graph-mode invocation	test_geir_sparse_reshape	For operator compilation and verification, refer to Operator Invocation.