takato3000 / nystrom-attention

Implementation of Nystrom Self-attention, from the paper Nystromformer

Nystrom Attention

Implementation of Nystrom Self-attention, from the paper Nystromformer.

Yannic Kilcher video

Install

$ pip install nystrom-attention

Usage

import torch
from nystrom_attention import NystromAttention

attn = NystromAttention(
    dim = 512,
    dim_head = 64,
    heads = 8,
    m = 256,                # number of landmarks
    pinv_iterations = 6,    # number of moore-penrose iterations for approximating pinverse. 6 was recommended by the paper
    residual = True         # whether to do an extra residual with the value or not. supposedly faster convergence if turned on
)

x = torch.randn(1, 16384, 512)
mask = torch.ones(1, 16384).bool()

attn(x, mask = mask) # (1, 1024, 512)

Citations

@misc{xiong2021nystromformer,
    title   = {Nystr\"omformer: A Nystr\"om-Based Algorithm for Approximating Self-Attention},
    author  = {Yunyang Xiong and Zhanpeng Zeng and Rudrasis Chakraborty and Mingxing Tan and Glenn Fung and Yin Li and Vikas Singh},
    year    = {2021},
    eprint  = {2102.03902},
    archivePrefix = {arXiv},
    primaryClass = {cs.CL}
}