Accepted by WACV 2021，收录到PytorchNetHub
简体中文 | English
- 基于注意力机制的特征融合
- 特征融合的统一方式，以下均适用
(a)Same Layer (b)Short Skip (c)Long Skip
- 非官方
- 2020.11.20 支持AFFResNet, AFFResNeXt
- 2020.11.19 新增MS_CAM, AFF, iAFF
from fusion import MS_CAM
# x[B,C,H,W] like SE Module
fusion_mode = MS_CAM(channels=C)
x = fusion_mode(x)from fusion import AFF, iAFF
# x,residual [B,C,H,W]
fusion_mode = AFF(channels=C)
x = fusion_mode(x, residual)- resnet 18/34/50/101/152
- resnext50_32x4d / resnext101_32x8d
- wide_resnet50_2 / wide_resnet101_2
| Argument | Description |
|---|---|
fuse_type (str,default: DAF) |
特征融合类型，支持AFF,iAFF,DAF |
small_input (bool,default: False) |
WH<=112为True |
import resnet50
net = resnet50(fuse_type='DAF',small_input=False)
pred = net(imgs)