CumulativeLogsumexp

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: computes the cumulative log-sum-exp of the input tensor x along the dimension specified by axis and saves the result to the output tensor y.
  • Computation formula: let ii be the subscript on the axis dimension, and SiS_i be the set of prefix or suffix subscripts involved in computing the ii-th output element, then:

yi=log⁡∑j∈Siexjy_i = \log \sum_{j \in S_i} e^{x_j}

When exclusive is false, SiS_i includes the current position; when exclusive is true, SiS_i does not include the current position. When reverse is false, accumulation proceeds in the direction of increasing subscripts; when reverse is true, accumulation proceeds in the direction of decreasing subscripts.

Parameter Description

Parameter Name Input/Output/Attribute Description Data Type Data Format
x Input Input Tensor. The shape supports 1 to 6 dimensions, and the length of the dimension specified by axis must be greater than 0. FLOAT, FLOAT16 ND
axis Input The dimension along which the cumulative log-sum-exp is computed. It must be a compile-time constant scalar, and the value range is [-rank(x), rank(x)-1]. INT32, INT64, INT16 ND
exclusive Attribute The default value is false. false indicates that the current position is included; true indicates that the current position is not included. BOOL -
reverse Attribute The default value is false. false indicates forward accumulation; true indicates backward accumulation. BOOL -
y Output Output Tensor, whose shape is the same as that of x. FLOAT, FLOAT16 ND

Constraint Description

  • axis must be a constant scalar.
  • x does not support scalar input. A parameter error is returned when the length of the dimension specified by axis is 0.
  • The shape and data type of y must be the same as those of x.

Invocation Description

Invocation Method Sample Code Description
Graph-mode invocation test_geir_cumulative_logsumexp.cpp Invokes the CumulativeLogsumexp operator by building a graph through the operator IR.