Shulou(Shulou.com)06/01 Report--

Editor to share with you what is the difference between F.avg_pool1d () and F.avg_pool2d () in pytorch, I believe most people don't know much about it, so share this article for your reference. I hope you will gain a lot after reading this article. Let's learn about it together.

F.avg_pool1d () data is 3D input

Input dimension: (batch_size,channels,width) channel can be regarded as height

Kenerl dimension: (one dimension: represents the span of the width) channel and the input channel can be regarded as the height of the matrix

If kernel_size=2 is assumed, the average of every two columns is calculated. Stride is consistent with kernel_size by default, and is discarded if it crosses the boundary. (the following shows that 1 and 3 columns add up to average)

Input = torch.tensor (input = torch.tensor). Float () print (input) m = F.avg_pool1d (input,kernel_size=2) mtensor ([[1.1,1.1.1.1.1.1.1], [1.1.1.1.1.1.1.1.1], [0.000,0.00. 0, 1, 1.], [1, 1, 1, 1.], [1, 1, 1, 1.]]) tensor ([[1.0000, 1.0000], [1.0000, 1.0000], [0.0000, 0.5000], [1.0000, 1.0000], [1.0000, 1.0000]])

Suppose kenerl_size=3, which means that the first three columns are added to average, and the next three columns are discarded.

Input = torch.tensor (input = torch.tensor). Float () print (input) m = F.avg_pool1d (input,kernel_size=3) mtensor ([[1.1,1.1.1.1.1.1.1], [1.1.1.1.1.1.1.1.1], [0.000,0.00. 0, 1, 1.], [1, 1, 1, 1.], [1, 1, 1, 1.]]) tensor ([1.], [1.], [0.], [1.], [1.]]) input = torch.tensor Unsqueeze (0). Float () print (input) m = F.avg_pool1d (input,kernel_size=4) mtensor ([1, 1, 1, 1, 1.], 1.], [0, 0, 0, 1.], [1. 1, 1.], [1, 1., 1.]]) tensor ([1.0000], [1.0000], [0.2500], [1.0000], [1.0000])

Suppose stride=1 moves one step at a time

Input = torch.tensor (input = torch.tensor). Float () print (input) m = F.avg_pool1d (input,kernel_size=2,stride=1) mtensor ([[1.1,1.1.1.1.1.1.1], [1.1.1.1.1.1.1.1.1], [0.000,0.00. 0, 1, 1.], [1, 1, 1, 1.], [1, 1, 1, 1.]) tensor ([[1.0000, 1.0000, 1.0000, 1.0000], [1.0000, 1.0000, 1.0000, 1.0000], [0.0000, 0.0000, 0.5000, 1.0000] [1.0000, 1.0000, 1.0000, 1.0000], [1.0000, 1.0000, 1.0000, 1.0000]]) input = torch.tensor ([1 float () print (input) m = F.avg_pool1d (input,kernel_size=4)) Stride=1) mtensor ([[1, 1, 1, 1, 1.], [1, 1, 1, 1.], [0, 0, 0, 1, 1.], [1, 1, 1, 1.], [1, 1, 1, 1.]]) tensor ([[1.0000, 1.0000]]) [1.0000, 1.0000], [0.2500, 0.5000], [1.0000, 1.0000], [1.0000, 1.0000]]) F.avg_pool2d () data is a four-dimensional input

Input Dimension: (batch_size,channels,height,width)

Kenerl dimension: (2D: represents the span of width) channel is the same as the input channle, if the data is 3D, then channel is 1. (if you write only one number, nfocus kenerl = (nmenn))

Stride is consistent with kenerl by default, which is two-dimensional, so it is consistent with kenerl on both height and width, and is also discarded when crossing the boundary.

Consistent with cnn convolution

Input = torch.tensor ([1) print (input.size ()) print (input) m = F.avg_pool2d (input,kernel_size= (4)) mtorch.Size ([1,5,5]) tensor ([[1,5,5]) tensor 1, 1.], [0, 0, 0, 1, 1.], [1, 1, 1, 1.], [1, 1, 1, 1.], [1, 1, 1, 1, 1.]]) tensor ([0.8125]) input = torch.tensor ([[1mel]]) input = torch.tensor. Float () print (input.size ()) print (input) m = F.avg_pool2d (input,kernel_size= (4), stride=1) mtorch.Size ([1,5,5]) tensor ([[1.1,1.1.1.1.1.], [1.1.1.1.1.1.1.], [0.0.0.0.1.1.1.1.], [1. 1, 1, 1.], [1, 1, 1, 1.]]) tensor ([0.8125, 0.8750], [0.8125, 0.8750])

If you find the average kenerl= of the column (1), the default is stride= (1).

Input = torch.tensor ([1 print () input.size () print (input) m = F.avg_pool2d (input,kernel_size= (1) mtorch.Size ([1J 5])) tensor ([[1.1,1.1.1.1.1.1], [1.1.1]. 1, 1, 1.], [0, 0, 0, 1, 1.], [1, 1, 1, 1.], [1, 1, 1, 1.]]) tensor ([1.0000], [1.0000], [0.4000], [1.0000]) [1.0000])

If you find the average kenerl= of a row (5d1), then the default stride= (5d1), think with the concept of convolution

Input = torch.tensor ([1) print (input.size ()) print (input) m = F.avg_pool2d (input,kernel_size= (5) mtorch.Size ([1,5,5]) tensor ([[1,5,5]) tensor 1, 1.], [0, 0, 0, 1, 1.], [1, 1, 1, 1, 1.], [1, 1, 1, 1.]]) tensor ([0.8000, 0.8000, 0.8000, 1.0000, 1.0000])

For four-dimensional data, channel defaults to the same input

Input=torch.randn m=F.avg_pool2d (input, (4pr 4)) print (m.size ()) torch.Size ([10Jing 3,1,1])

Supplement: parsing of AdaptiveAvgPool function in PyTorch

Adaptive pooling (AdaptiveAvgPool1d):

For the input signal, provide 1-dimensional adaptive average pooling operation for any input size, the output size can be specified as Hauw, but the number of input and output features will not change.

Torch.nn.AdaptiveAvgPool1d (output_size) # output_size: output Siz

For the input signal, provide 1-dimensional adaptive average pooling operation for any input size, the output size can be specified as Hauw, but the number of input and output features will not change.

# target output size of 5m = nn.AdaptiveAvgPool1d (5) input = autograd.Variable (torch.randn (1,64,8)) output = m (input) adaptive pooling (AdaptiveAvgPool2d): class torch.nn.AdaptiveAvgPool2d (output_size)

For input signals, provide two-dimensional adaptive average pooling operation for any input size, the output size can be specified as Hauw, but the number of input and output features will not change.

Parameters:

Output_size: the size of the output signal, which can be represented by (HMagneW) or the delay digit H, which can be used to denote the output of Hidden

# target output size of 5x7m = nn.AdaptiveAvgPool2d ((5p7)) input = autograd.Variable (torch.randn (1,64,8,9)) # target output size of 7x7 (square) m = nn.AdaptiveAvgPool2d (7) input = autograd.Variable (torch.randn (1,64,10,9)) output = m (input) above is all the contents of this article entitled "what's the difference between F.avg_pool1d () and F.avg_pool2d () in pytorch". Thank you for reading! I believe we all have a certain understanding, hope to share the content to help you, if you want to learn more knowledge, welcome to follow the industry information channel!

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.