In image retrieval tasks, the sum pooling is a common operator.

But I can’t find it in pytorch.

Could anyone tell me how to perform sum pooling in CNN?

You can view the feature map as a 1d tensor, and call sum on it.

Thanks! It’s the key to handle the CNN tensor!

For anyone who is also interested on how to do it exactly. Suppose the feature map is of size `N*C*H*W`

, after sum-pooling, it will become tensor of size `N*C`

(N image, each has a feature vector of dimension C). Here is a code snippet to do it,

```
# suppose x is the feature map after some layer
x = torch.sum(x.view(x.size(0), x.size(1), -1), dim=2)
```

The above code flattens the 2nd and 3rd dimension of original tensor to a vector and calculate sum on the newly created tensor on dimension 2, which is exactly what sum-pooling does.

1 Like

How to perform Sum pooling with view if one needs to use a particular kernel size and stride just like nn.Maxpool2d ? In this case input is **N*C*W_in*H_in** and output should be **N*C*W_out*H_out**

.