FilesLast commitLast update
4 months ago
3 months ago
1 month ago
2 months ago
2 months ago
1 month ago
1 month ago
16 days ago
README

CosineSimilarity

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 cosine similarity of two input tensors along the specified dimension. The cosine similarity measures the cosine of the angle between two vectors, falls within the range [-1, 1], and is commonly used to measure how similar two vectors are.

  • Computation formula:

cosine_similarity(x1,x2,dim,eps)=∑(x1⋅x2,dim)max⁡(∑(x12,dim),eps)⋅max⁡(∑(x22,dim),eps)\text{cosine\_similarity}(x_1, x_2, \text{dim}, \text{eps}) = \frac{\sum(x_1 \cdot x_2, \text{dim})}{\max(\sqrt{\sum(x_1^2, \text{dim})}, \text{eps}) \cdot \max(\sqrt{\sum(x_2^2, \text{dim})}, \text{eps})}

Parameter Description

Parameter Name Input/Output/Attribute Description Data Type Data Format
input_x1 Input The first input tensor, whose rank ranges from 1 to 8. FLOAT ND
input_x2 Input The second input tensor, which must have exactly the same shape as input_x1. FLOAT ND
output_y Output The output tensor after reduction along the dim dimension. Its shape is the input shape with the dimension corresponding to dim removed. FLOAT ND
dim Attribute The dimension along which the cosine similarity is computed. The value range is [-rank, rank-1], and the default value is 1. INT32 -
eps Attribute A non-negative numerical stability parameter that prevents division by zero. The default value is 1e-8. FLOAT -

Constraint Description

  • Only the FLOAT (float32) data type is supported for the inputs and output.
  • The two inputs must have the same rank, and the rank ranges from 1 to 8. 0D tensors are not supported.
  • Each dimension of the two inputs must have the same length, that is, input_x1.shape == input_x2.shape.
  • No form of broadcasting is supported, including expanding dimensions of length 1 and left-padding 1s for inputs of different ranks.
  • A shape that contains a dimension of length 0 is not supported.
  • dim must be within the [-rank, rank-1] range for the input rank. A negative index is processed as the corresponding positive dimension, for example, dim=-1 indicates the last dimension.
  • The output shape is the common shape of the two inputs with the dimension corresponding to dim removed.
  • eps must be a non-negative value.
  • If both inputs have a rank of 1, dim must be explicitly specified within the [-1, 0] range.

Invocation Description

Invocation Method Sample Code Description
Graph-mode invocation test_geir_cosine_similarity Invokes the CosineSimilarity operator by building a graph through the operator IR