Adaptive version? about antialiased-cnns HOT 2 CLOSED

Sorry, not quite sure what you mean exactly. But if you want to replace a stride=4 layer, rather than stride=2, which we experimented with, yes, a wider blur kernel can be used.

from antialiased-cnns.

@Vlad0922 I think Adaptive pooling layers are typically used on the final convolutional feature map, before being fed into the hidden layer. Are those the layers you were referring to? (Though in this implementation of MobileNet-V3, they use adaptive pooling inside the inverted residual blocks)

@richzhang's Adaptive pooling takes as input only the output size and internally adjusts the kernel size to produce the desired output size (docs). I think an adaptive BlurPool version would require computing the filter_size on the fly?

On a related note, would you recommend adding a BlurPool layer for the final feature map downsampling?

from antialiased-cnns.