PlumedSerpent / Maxout-PyTorch

Here is the implementation of Maxout Layer from paper: https://arxiv.org/pdf/1302.4389.pdf in PyTorch. In forward pass, maxout for the feature maps of CNN is calculated. For backward pass, derivative of error w.r.t input is only propagated through the cells of feature maps array which were activated in forward pass(which is derivative of Maxout). To be sure this works fine, results were compared with Maxout in Theano(which is with the name featurepool layer) and were confirmed to be working fine. To make it work according to your desire, just change max_out parameter in forward function. The code is very efficient and fast.

Maxout-PyTorch

Here is the implementation of Maxout Layer from paper: https://arxiv.org/pdf/1302.4389.pdf in PyTorch. In forward pass, maxout for the feature maps of CNN is calculated. For backward pass, derivative of error w.r.t input is only propagated through the cells of feature maps array which were activated in forward pass(which is derivative of Maxout). To be sure this works fine, results were compared with Maxout in Theano(which is with the name featurepool layer) and were confirmed to be working fine. To make it work according to your desire, just change max_out parameter in forward function. The code is very efficient is fast.