当前位置：网站首页>Developers share ｜ handwriting operator is not so difficult. I'll teach you to use mindspire to realize adaptive average pooling operator!

Developers share ｜ handwriting operator is not so difficult. I'll teach you to use mindspire to realize adaptive average pooling operator!

2022-07-18 22:23:00 【Shengsi mindspire】

Recently, with MindSpore Reappear Fast SCNN When it comes to the Internet , It uses an adaptive average pooling operator “nn.AdaptiveAvgPool2d”, however MindSpore The current version does not provide the corresponding operator for developers , So the author consulted some materials , Understand its calculation principle . Experimental results show that , The following methods can replace AdaptiveAvgPool2d operator , Based on this, you can migrate to other AI In the frame .

01

AdaptiveAvgPool2d

AdaptiveAvgPool2d The function of is simply , Developers just need to pass in “ Data to be processed ” and “ Target size ”, This operator will automatically calculate the kernel_size and stride Data such as , Make the output result shape by “ Target size ”.

But in fact, the above understanding is not completely correct , Even wrong . Based on this understanding, we can slightly understand its function , But if you want to reproduce it , Will fall into the misunderstanding completely .

At present, a common way of recurrence is , Since we know that the calculation process of common pooling operation is , The of the pool layer is known kernel_size、padding、stride And the size of the input tensor input_size, Then the output tensor size output_size by ：

output_size =（input_size+2*padding-kernel_size）/stride +1

（ The calculation is simplified here , If you input tensor column Values and row It's not worth it , Then calculate respectively ）

Then we'll find a way to pass input_size and output_size Roll back kernel_size、stride Just wait for the value , The data we need can be obtained by reverse calculation .

But in fact, this method only fits its shape , Not to his liking . We can only make the output tensor what we need “ Target size ”, Its internal value is the same as “nn.AdaptiveAvgPool2d” There is no small difference in the calculation results of , The reason is , Lies in the wrong starting point .

02

AdaptiveAvgPool2d Calculation principle

The author has read a lot of information , In the end in https://discuss.pytorch.org/t/what-is-adaptiveavgpool2d/26897 Found what you need . among Thomas Yes AdaptiveAvgPool2d Made a fairly accurate explanation , I improved the code they shared to NCHW The pattern of ：

import torch.nn as nn
import torch

def torch_pool(inputs, target_size):
    #NCHW
    H = target_size[0]
    W = target_size[1]
    s_p1 = (torch.arange(W, dtype=torch.float32) * (inputs.size(-1) / W)).long()
    e_p1 = ((torch.arange(W, dtype=torch.float32)+1) * (inputs.size(-1) / W)).ceil().long()

    s_p2 = (torch.arange(H, dtype=torch.float32) * (inputs.size(-2) / H)).long()
    e_p2 = ((torch.arange(H, dtype=torch.float32)+1) * (inputs.size(-2) / H)).ceil().long()

    pooled2 = []
    for i_H in range(H):
      pooled = []
      for i_W in range(W):
          res = torch.mean(inputs[:, :, s_p2[i_H]:e_p2[i_H],s_p1[i_W]:e_p1[i_W]], dim=(-2,-1), keepdim=True)
          pooled.append(res)
      pooled = torch.cat(pooled, -1)
      pooled2.append(pooled)
    pooled2 = torch.cat(pooled2,-2)
    return pooled2

if __name__ == '__main__':

    data = [[[[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]]
              ,
            [[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]]
              ]]

    inputs = torch.tensor(data,dtype=torch.float32)
    print(inputs)
    print(inputs.size())
    print("*********************************")

    avgpool1 = torch_pool(inputs, (1,3))
    avgpool2 = torch_pool(inputs, (2,3))
    avgpool3 = torch_pool(inputs, (3,3))
    avgpool6 = torch_pool(inputs, (6,5))

    print(avgpool1)
    print("*********************************")
    print(avgpool2)
    print("*********************************")
    print(avgpool3)
    print("*********************************")
    print(avgpool6)

The result of the calculation is ：

tensor([[[[2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.]],

         [[2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.]]]])
torch.Size([1, 2, 6, 8])
*********************************
tensor([[[[3., 6., 8.]],

         [[3., 6., 8.]]]])
*********************************
tensor([[[[3., 6., 8.],
          [3., 6., 8.]],

         [[3., 6., 8.],
          [3., 6., 8.]]]])
*********************************
tensor([[[[3., 6., 8.],
          [3., 6., 8.],
          [3., 6., 8.]],

         [[3., 6., 8.],
          [3., 6., 8.],
          [3., 6., 8.]]]])
*********************************
tensor([[[[2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000]],

         [[2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000]]]])

And pytorch Of nn.AdaptiveAvgPool2d Operators for comparison and verification ：

import torch.nn as nn
import torch

if __name__ == '__main__':

    data = [[[[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]]
              ,
            [[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]
              ,[2,3,4,5,6,9,7,8]]
              ]]

    x = torch.tensor(data,dtype=torch.float32)
    print(x)
    print(x.size())
    print("*********************************")

    avgpool1 = nn.AdaptiveAvgPool2d((1,3))
    avgpool2 = nn.AdaptiveAvgPool2d((2,3))
    avgpool3 = nn.AdaptiveAvgPool2d((3,3))
    avgpool6 = nn.AdaptiveAvgPool2d((6,5))
    print(avgpool1(x))
    print("*********************************")
    print(avgpool2(x))
    print("*********************************")
    print(avgpool3(x))
    print("*********************************")
    print(avgpool6(x))

The result of the calculation is ：

tensor([[[[2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.]],

         [[2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.],
          [2., 3., 4., 5., 6., 9., 7., 8.]]]])
torch.Size([1, 2, 6, 8])
*********************************
tensor([[[[3., 6., 8.]],

         [[3., 6., 8.]]]])
*********************************
tensor([[[[3., 6., 8.],
          [3., 6., 8.]],

         [[3., 6., 8.],
          [3., 6., 8.]]]])
*********************************
tensor([[[[3., 6., 8.],
          [3., 6., 8.],
          [3., 6., 8.]],

         [[3., 6., 8.],
          [3., 6., 8.],
          [3., 6., 8.]]]])
*********************************
tensor([[[[2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000]],

         [[2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000],
          [2.5000, 4.0000, 5.5000, 7.3333, 7.5000]]]])

You can find , Whether it's output shape still Tensor Internal values , Both are the same . And no matter what changes NCHW Which part of , expand N、C、H、W Any one dimension of , The final calculation results are consistent .

03

MindSpore edition AdaptiveAvgPool2d

Look at the code above , If you want to rewrite MindSpore Version of the code , We just need to replace ‘torch.arange’、‘torch.mean’、‘torch.cat’ These three main operators , And add a rounding operation , stay MindSpore The middle is ops.ReduceMean(keep_dims=True)、P.Concat(axis=-1) Equal operator , Just do the corresponding replacement .
But I was rewriting res = torch.mean(inputs[:, :, s_p2[i_H]:e_p2[i_H],s_p1[i_W]:e_p1[i_W]], dim=(-2,-1), keepdim=True) In this sentence , Find out MindSpore Yes ‘ Variable subscript ’ Exceptions will occur when slicing , I'm not sure if there is something wrong with my usage .

But I wrote a provisional version , For example, if you want to NCx32x64 Data pooling of NCx6x6 size , We can calculate the subscript of the slice in advance , You can get such a version of available code ：

def _AvgPool2d6x6(self,x):
        s_p1 = [ 0, 10, 21, 32, 42, 53]
        e_p1 = [11, 22, 32, 43, 54, 64]
        s_p2 = [ 0,  5, 10, 16, 21, 26]
        e_p2 = [ 6, 11, 16, 22, 27, 32]       
        pooled2 = []
        for i_H in range(6):
          pooled = []
          for i_W in range(6):
              res = self.reduceMean(x[:, :, s_p2[i_H]:e_p2[i_H],s_p1[i_W]:e_p1[i_W]], (-2,-1))
              pooled.append(res)
          pooled = self.concat1((pooled[0],pooled[1],pooled[2],pooled[3],pooled[4],pooled[5]))
          pooled2.append(pooled)
        pooled2 = self.concat2((pooled2[0],pooled2[1],pooled2[2],pooled2[3],pooled2[4],pooled2[5]))       
        return pooled2