Thanks for the nice article. Actually, there is a 2d dropout for conv layers, for example nn.Dropout2d in pytorch or SpatialDropout2D in tensorflow, which behaves differently than described in the article :)