torch.nn.functional.avg_pool1d is not working correctly

[yueyinqiu]

Hello! I've found that torch.nn.functional.avg_pool1d is not working as expected.

Here is my test code:

using TorchSharp;

var x = torch.zeros(5, 7, 128);
Console.WriteLine(x.metastr());
// [5x7x128], type = Float32, device = cpu

var y1 = torch.nn.functional.avg_pool1d(x, 2);
Console.WriteLine(y1.metastr());
// [5x7x127], type = Float32, device = cpu

var y2 = torch.nn.AvgPool1d(2).call(x);
Console.WriteLine(y2.metastr());
// [5x7x64], type = Float32, device = cpu

And in PyTorch:

import torch

x = torch.zeros(5, 7, 128)
print(x.size())
# torch.Size([5, 7, 128])

y1 = torch.nn.functional.avg_pool1d(x, 2)
print(y1.size())
# torch.Size([5, 7, 64])

y2 = torch.nn.AvgPool1d(2)(x)
print(y2.size())
# torch.Size([5, 7, 64])

There seems to be something going wrong with torch.nn.functional.avg_pool1d, while torch.nn.AvgPool1d still works well.

[shaltielshmid]

Thank you for reporting! The discrepancy between PyTorch and TorchSharp is the default value assigned to the stride parameter. I submitted a fix for the next version, but until then you can just specify the stride parameter to be the same as the kernelSize parameter to get the desired results.

E.g.,

var y1 = torch.nn.functional.avg_pool1d(x, 2, 2);

[NiklasGustafsson]

Fixed in v0.102.1.