SparseBincount

English | 简体中文

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: counts the occurrences or weighted accumulated sum of each value (bin index) in a sparse tensor. Both the 1D and multi-dimensional modes are supported, and the binary_output mode is supported.

• Computation formula:

$$\text{1D: }output[v]+=w_i(\text{if }v<size)$$

$$\text{Multi-dim: }output[batch][v]+=w_i(\text{if }v<size\text{ and }batch<dense\_shape[0])$$

Parameter Description

Parameter Name	Input/Output/Attribute	Description	Data Type	Data Format
indices	Input	Indices of the sparse tensor, a 2D tensor with shape (N, R).	INT64	ND
values	Input	Values of the sparse tensor (bin index), a 1D tensor with shape (N,).	INT32, INT64	ND
dense_shape	Input	Dense shape of the sparse tensor, a 1D tensor with shape (R,).	INT64	ND
size	Input	Number of bins (scalar tensor) with shape (1,).	INT32, INT64	ND
weights	Input	Weights, a 1D tensor with shape (N,) or (0,). If empty, each occurrence is counted as 1.	FLOAT	ND
binary_output	Optional attribute	Whether to output binary values. The default value is false. If True, bins that have appeared are set to 1. If False, the count or weighted accumulation is output.	BOOL	-
output	Output	Statistics result. In 1D mode, the shape is [size]. In multi-dimensional mode, the shape is [dense_shape[0], size].	FLOAT	ND

Constraint Description

• indices is a 2D tensor with shape (N, R); values, dense_shape, size, and weights are all 1D tensors
• The number of rows of indices (N) must be equal to the number of elements of values (N)
• The number of columns of indices (R) must be equal to the number of elements of dense_shape (R), and the number of elements of dense_shape (R) must be at least 1 (dense_shape is not empty)
• indices[:, j] must be in the range [0, dense_shape[j])
• values[i] must be in the range [0, size)
• If weights is not empty, the number of elements must be equal to the number of elements of values (N)
• size is a non-negative scalar with shape (1,)
• If weights is empty (the number of elements is 0), each occurrence is counted as 1.0
• If binary_output = True, bins that have appeared are set to 1.0 and weights are ignored

Invocation Description

Invocation Method	Sample Code	Description
GE graph mode	test_geir_sparse_bincount	Invokes the SparseBincount operator by building a graph through the operator IR.