o h@s.dZddlZddlmZGdddejZdS)a A layer which implements maxout from the "Maxout Networks" paper https://arxiv.org/pdf/1302.4389v4.pdf Goodfellow, Warde-Farley, Mirza, Courville, Bengio or a simpler explanation here: https://stats.stackexchange.com/questions/129698/what-is-maxout-in-neural-network/298705#298705 The implementation here: for k layers of maxout, in -> out channels, we make a single linear map of size in -> out*k then we reshape the end to be (..., k, out) and return the max over the k layers Ncs$eZdZfddZddZZS) MaxoutLinearcs2t||_||_||_t||||_dS)N)super__init__ in_channels out_channelsmaxout_knnLinearlinear)selfrrr __class__]/var/www/html/env_mimamsha/lib/python3.10/site-packages/stanza/models/common/maxout_linear.pyrs zMaxoutLinear.__init__cCsF||}|jg|jdd|j|jR}tj|ddd}|S)z Use the oversized linear as the repeated linear, then take the max One large linear map makes the implementation simpler and easier for pytorch to make parallel N)dimr)r viewshaperrtorchmax)r inputsoutputsrrrforward s &zMaxoutLinear.forward)__name__ __module__ __qualname__rr __classcell__rrr rrs  r)__doc__rtorch.nnrModulerrrrrs