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.)
ARConv is an implementation of the autoregressive component 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.