Bidirectional Long Expressive Memory

This repository contains the implementation of Bidirectional Long Expressive Memory, which is the bidirectional form of the paper Long Expressive Memory for Sequence Modeling.

Note:

• Cuda implementation required for faster implementation.

Usage:

import torch
from lem import BidirectionalSeqLEM


input = torch.ones([2, 1, 784])    # Size : [B, dim, seq_len]

model = BidirectionalSeqLEM( ninp = 1, nhid = 128)

ys = model(input.permute(2, 0, 1)) # ys shape : [seq_len, B, 2 * dim]

References:

• https://github.com/tk-rusch/LEM