Wrong results of AdaptiveConcatPool2d in the docs.fast.ai

doc link: class AdaptiveConcatPool2d

In the doc, it says AdaptiveConcatPool2d should improve AdaptiveMaxPool2d or AdaptiveAvgPool2d; however, the examples given over there actually shows the MaxPool is the best (as of now).

I personally tried all three versions and AdaptiveConcatPool2d is the best. I guess the person who wrote the docs got it wrong somewhere? There may be negative examples but odd to see one in the doc, which stated otherwise.