Closed psicalculus closed 8 years ago
you actually misunderstand. Torch does and always supported "sliding window CNNs". If you have your entire neural network consiting of Spatial* modules, for example SpatialConvolution, SpatialMaxPooling etc., then if you give a larger input than what the network experts, the network is slid over the larger input to give a larger output than 1x1. Hope this helps.
Could you please add support for 2D/3D sliding window CNNs in Torch nn?
The following papers use sliding window CNNs: 1) http://people.idsia.ch/~juergen/nips2012.pdf 2) https://papers.nips.cc/paper/5636-recursive-training-of-2d-3d-convolutional-networks-for-neuronal-boundary-prediction.pdf 3) http://arxiv.org/pdf/1606.02382.pdf 4) http://arxiv.org/pdf/1606.07372.pdf
The only software framework available for doing this is ZNN but it only runs on CPUs: http://znn-release.readthedocs.io/en/latest/index.html
We would appreciate a GPU-based implementation of sliding window CNNs (similar to what the people in paper 1) used).