Custom Pooling Layers
This repo contains implementations for the following papers in tensorflow or pytorch.
Tensorflow
- Convolutional Bottleneck Attention Module (arxiv)(Not pooling I know)
- Stochastic Spatial Sampling (arxiv)
- Detail Preserving Pooling (arxiv)
Pytorch
- Lossless Pooling (arxiv)
Notice
- I just picked these layers because they seem simple but the papers claim decent improvments using them.
- The S3Pool and DPP implementations are hacky approximations which use softmax instead of selecting indices. If you can think of a more efficient but also clean way to do it in Tensorflow please submit a PR.