ARMANets.jl Documentation

This package provides a Julia, Flux-based implementation of the paper ARMA Nets: Expanding Receptive Field for Dense Prediction. (The official implementation from the authors is in PyTorch.)

The layer ARConv is an implementation of the autoregressive component[1] of the ARMA layer (the moving average component is just an ordinary convolution). For more details, see v1 and v2 of the paper.

  • 1This is the reparameterized variant, as the general ARMA layer is not BIBO stable and is very difficult to train on its own.