How to use Bilinear interpolation to reduce the number of feature maps?

1

I am a freshman and I meet an issue.

I have a batch of feature maps [640, 28, 28]. I want to use bilinear interpolation to convert to [3, 28, 28]. How can I do that?

The function torchvision.transforms.Resize looks like do not have this option.

2

Interpolation techniques, such as F.interpolate, work usually in the spatial dimensions, while it seems you would like to reduce the number of channels.
If that’s the case, you could use a conv layer with in_channels=640, out_channels=3 and kernel_size=1.

Alternatively you could also permute the input and use F.interpolate on the original channel dimension, which would be then in a spatial dim.

3

Thanks, I got it!!! :muscle: :muscle: :muscle: :muscle: