DropBlock is a type of regularisation that tries to replace dropout. The original paper describes it as best used with a linear scaling rate across blocks in a model, as has been implemented in #174. However, timm adds several experimental features, including:
DropBlockis a type of regularisation that tries to replace dropout. The original paper describes it as best used with a linear scaling rate across blocks in a model, as has been implemented in #174. However, timm adds several experimental features, including:gammacalculated (already included in Overhaul of ResNet API #174).DropBlockrates picked by experimentation to give better results on ResNets - these however extend only to four-stage ResNets i.e. those similar to the models in the original ResNet paper.