search

BVLC / caffe

Caffe: a fast open framework for deep learning.
http://caffe.berkeleyvision.org/
Other
34.13k stars 18.68k forks source link

Max Pooling Layers Requires 4D Input #4618

Open RoyiAvital opened 8 years ago

RoyiAvital commented 8 years ago

Hi, I'm using Caffe for 1D Signals.

Though the Convolution Layer can accept 3D Input (numSignalBatch * numChannels * numSamples) the Pooling layer input must be in 4D.

Could you make 1D signals first class citizens in Caffe? Moreover, could you add support for Pooling?

Thank You.

williford commented 8 years ago

Pooling implements spatial pooling. If there is no spatial component, spatial pooling does not make sense. You might want to ask for help with 1D signal processing on the Caffe user group or on StackOverflow.

abhimanyudubey commented 8 years ago

Any signal in a lower dimension is a signal in a higher dimension. N 1D signals (of shape NxD) need to just be kept in a 4D container (of shape Nx1x1xD) to be operational with the Caffe format, and this can be done with the caffe Reshape layer. Remember to set the pooling parameters to have only one non-unitary pooling dimension, otherwise you will face issues.

Again, I request you to post all operation issues and queries on the caffe-users group.