local attention

当地关注

本地窗口注意力的实现，为语言建模设定了令人难以置信的强大基线。越来越明显的是，变压器需要底层的局部关注，而顶层则保留用于全局关注以整合先前层的发现。该存储库可以轻松地立即使用本地窗口注意力。

该代码已经在多个存储库中进行了战斗测试，同时还进行了稀疏远程注意力的不同实现。

安装

$ pip install local-attention

用法

 import torch
from local_attention import LocalAttention

q = torch . randn ( 2 , 8 , 2048 , 64 )
k = torch . randn ( 2 , 8 , 2048 , 64 )
v = torch . randn ( 2 , 8 , 2048 , 64 )

attn = LocalAttention (
    dim = 64 ,                # dimension of each head (you need to pass this in for relative positional encoding)
    window_size = 512 ,       # window size. 512 is optimal, but 256 or 128 yields good enough results
    causal = True ,           # auto-regressive or not
    look_backward = 1 ,       # each window looks at the window before
    look_forward = 0 ,        # for non-auto-regressive case, will default to 1, so each window looks at the window before and after it
    dropout = 0.1 ,           # post-attention dropout
    exact_windowsize = False # if this is set to true, in the causal setting, each query will see at maximum the number of keys equal to the window size
)

mask = torch . ones ( 2 , 2048 ). bool ()
out = attn ( q , k , v , mask = mask ) # (2, 8, 2048, 64)

该库还允许在共享查询/密钥空间（Reformer 架构）的设置中进行本地关注。密钥的规范化以及令牌自身的屏蔽都将得到处理。

 import torch
from local_attention import LocalAttention

qk = torch . randn ( 2 , 8 , 2048 , 64 )
v  = torch . randn ( 2 , 8 , 2048 , 64 )

attn = LocalAttention (
    dim = 64 ,
    window_size = 512 ,
    shared_qk = True ,
    causal = True
)

mask = torch . ones ( 2 , 2048 ). bool ()
out = attn ( qk , qk , v , mask = mask ) # (2, 8, 2048, 64)

如果您希望模块自动填充您的查询/键/值以及掩码，只需将autopad关键字设置为True

 import torch
from local_attention import LocalAttention

q = torch . randn ( 8 , 2057 , 64 )
k = torch . randn ( 8 , 2057 , 64 )
v = torch . randn ( 8 , 2057 , 64 )

attn = LocalAttention (
    window_size = 512 ,
    causal = True ,
    autopad = True      # auto pads both inputs and mask, then truncates output appropriately
)

mask = torch . ones ( 1 , 2057 ). bool ()
out = attn ( q , k , v , mask = mask ) # (8, 2057, 64)

局部注意力变压器

完整的局部注意力转换器

 import torch
from local_attention import LocalTransformer

model = LocalTransformer (
    num_tokens = 256 ,
    dim = 512 ,
    depth = 6 ,
    max_seq_len = 8192 ,
    causal = True ,
    local_attn_window_size = 256
). cuda ()

x = torch . randint ( 0 , 256 , ( 1 , 8192 )). cuda ()

logits = model ( x ) # (1, 8192, 256)

Enwik8 4096

窗口大小为 256，回顾为 1，总感受野为 512

$ python train.py

引文

 @inproceedings { rae-razavi-2020-transformers ,
    title   = " Do Transformers Need Deep Long-Range Memory? " ,
    author  = " Rae, Jack  and Razavi, Ali " ,
    booktitle = " Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics " ,
    month   = jul,
    year    = " 2020 " ,
    address = " Online " ,
    publisher = " Association for Computational Linguistics " ,
    url     = " https://www.aclweb.org/anthology/2020.acl-main.672 "
}

 @misc { roy*2020efficient ,
    title   = { Efficient Content-Based Sparse Attention with Routing Transformers } ,
    author  = { Aurko Roy* and Mohammad Taghi Saffar* and David Grangier and Ashish Vaswani } ,
    year    = { 2020 } ,
    url     = { https://arxiv.org/pdf/2003.05997.pdf }
}

 @misc { beltagy2020longformer ,
    title   = { Longformer: The Long-Document Transformer } ,
    author  = { Iz Beltagy and Matthew E. Peters and Arman Cohan } ,
    year    = { 2020 } ,
    eprint  = { 2004.05150 } ,
    archivePrefix = { arXiv } ,
    primaryClass = { cs.CL }
}

 @inproceedings { Sun2022ALT ,
    title     = { A Length-Extrapolatable Transformer } ,
    author    = { Yutao Sun and Li Dong and Barun Patra and Shuming Ma and Shaohan Huang and Alon Benhaim and Vishrav Chaudhary and Xia Song and Furu Wei } ,
    year      = { 2022 }
}

 @article { Bondarenko2023QuantizableTR ,
    title   = { Quantizable Transformers: Removing Outliers by Helping Attention Heads Do Nothing } ,
    author  = { Yelysei Bondarenko and Markus Nagel and Tijmen Blankevoort } ,
    journal = { ArXiv } ,
    year    = { 2023 } ,
    volume  = { abs/2306.12929 } ,
    url     = { https://api.semanticscholar.org/CorpusID:259224568 }
}