Detailed explanation of pytoch's scatter function

Preface

  Looking at FCOS Algorithm source code , It is found that... Is used to obtain positive sample points scatter This function , So record it .

1、 Interpretation of official documents

  Post the link first ：scatter Official interpretation

Tensor.scatter_(dim, index, src, reduce=None) → Tensor

  Receive three parameters ： dim, index and src. This function is used to stay dim Dimensionally , according to index Index provided , from src Extract the corresponding element from the to assign to Tensor. The following is an official example of a three-dimensional tensor .

  Two points need to be noted ：index and src Of dim The dimensions must be the same ！ In the official 3-D tensor For example , namely self、src and index The dimensions of are 3; if 2D-tensor be self、src and index The dimensions of are 2. Because I need to use index As an index , so index The size of the element in must be <self.size(d) And src.size(d).

2、 for instance

Attach code ：

src = torch.Tensor([[0,1,2,3,4],[5,6,7,8,9]])
self = torch.zeros((3,5))
index = torch.tensor([[0,1,2,0,0],[2,0,0,1,2]])
self.scatter_(dim=0,index=index,src=src)
print(self)

Output ：

summary

  In actual programming ,src It is often scalar , Is a constant . According to the definition , On the right side of the equation src[i][j] Identical to scalar . Is this time scatter The function is based on index take self The corresponding position in becomes a constant .