deformable attention

Atención deformable

Implementación de Atención Deformable de este artículo en Pytorch, que parece ser una mejora de lo propuesto en DETR. La incrustación posicional relativa también se ha modificado para una mejor extrapolación, utilizando la incrustación posicional continua propuesta en SwinV2.

Instalar

$ pip install deformable-attention

Uso

 import torch
from deformable_attention import DeformableAttention

attn = DeformableAttention (
    dim = 512 ,                   # feature dimensions
    dim_head = 64 ,               # dimension per head
    heads = 8 ,                   # attention heads
    dropout = 0. ,                # dropout
    downsample_factor = 4 ,       # downsample factor (r in paper)
    offset_scale = 4 ,            # scale of offset, maximum offset
    offset_groups = None ,        # number of offset groups, should be multiple of heads
    offset_kernel_size = 6 ,      # offset kernel size
)

x = torch . randn ( 1 , 512 , 64 , 64 )
attn ( x ) # (1, 512, 64, 64)

atención deformable 3d

 import torch
from deformable_attention import DeformableAttention3D

attn = DeformableAttention3D (
    dim = 512 ,                          # feature dimensions
    dim_head = 64 ,                      # dimension per head
    heads = 8 ,                          # attention heads
    dropout = 0. ,                       # dropout
    downsample_factor = ( 2 , 8 , 8 ),      # downsample factor (r in paper)
    offset_scale = ( 2 , 8 , 8 ),           # scale of offset, maximum offset
    offset_kernel_size = ( 4 , 10 , 10 ),   # offset kernel size
)

x = torch . randn ( 1 , 512 , 10 , 32 , 32 ) # (batch, dimension, frames, height, width)
attn ( x ) # (1, 512, 10, 32, 32)

1d atención deformable por si acaso

 import torch
from deformable_attention import DeformableAttention1D

attn = DeformableAttention1D (
    dim = 128 ,
    downsample_factor = 4 ,
    offset_scale = 2 ,
    offset_kernel_size = 6
)

x = torch . randn ( 1 , 128 , 512 )
attn ( x ) # (1, 128, 512) 

Citación

 @misc { xia2022vision ,
    title   = { Vision Transformer with Deformable Attention } , 
    author  = { Zhuofan Xia and Xuran Pan and Shiji Song and Li Erran Li and Gao Huang } ,
    year    = { 2022 } ,
    eprint  = { 2201.00520 } ,
    archivePrefix = { arXiv } ,
    primaryClass = { cs.CV }
}

 @misc { liu2021swin ,
    title   = { Swin Transformer V2: Scaling Up Capacity and Resolution } ,
    author  = { Ze Liu and Han Hu and Yutong Lin and Zhuliang Yao and Zhenda Xie and Yixuan Wei and Jia Ning and Yue Cao and Zheng Zhang and Li Dong and Furu Wei and Baining Guo } ,
    year    = { 2021 } ,
    eprint  = { 2111.09883 } ,
    archivePrefix = { arXiv } ,
    primaryClass = { cs.CV }
}