113 lines
3.6 KiB
Python
113 lines
3.6 KiB
Python
""" Split Attention Conv2d (for ResNeSt Models)
|
|
|
|
Paper: `ResNeSt: Split-Attention Networks` - /https://arxiv.org/abs/2004.08955
|
|
|
|
Adapted from original PyTorch impl at https://github.com/zhanghang1989/ResNeSt
|
|
|
|
Modified for torchscript compat, performance, and consistency with timm by Ross Wightman
|
|
"""
|
|
from typing import Optional, Type, Union
|
|
|
|
import torch
|
|
import torch.nn.functional as F
|
|
from torch import nn
|
|
|
|
from .helpers import make_divisible
|
|
|
|
|
|
class RadixSoftmax(nn.Module):
|
|
def __init__(self, radix: int, cardinality: int):
|
|
super().__init__()
|
|
self.radix = radix
|
|
self.cardinality = cardinality
|
|
|
|
def forward(self, x):
|
|
batch = x.size(0)
|
|
if self.radix > 1:
|
|
x = x.view(batch, self.cardinality, self.radix, -1).transpose(1, 2)
|
|
x = F.softmax(x, dim=1)
|
|
x = x.reshape(batch, -1)
|
|
else:
|
|
x = torch.sigmoid(x)
|
|
return x
|
|
|
|
|
|
class SplitAttn(nn.Module):
|
|
"""Split-Attention (aka Splat)
|
|
"""
|
|
def __init__(
|
|
self,
|
|
in_channels: int,
|
|
out_channels: Optional[int] = None,
|
|
kernel_size: int = 3,
|
|
stride: int = 1,
|
|
padding: Optional[int] = None,
|
|
dilation: int = 1,
|
|
groups: int = 1,
|
|
bias: bool = False,
|
|
radix: int = 2,
|
|
rd_ratio: float = 0.25,
|
|
rd_channels: Optional[int] = None,
|
|
rd_divisor: int = 8,
|
|
act_layer: Type[nn.Module] = nn.ReLU,
|
|
norm_layer: Optional[Type[nn.Module]] = None,
|
|
drop_layer: Optional[Type[nn.Module]] = None,
|
|
**kwargs,
|
|
):
|
|
dd = {'device': kwargs.pop('device', None), 'dtype': kwargs.pop('dtype', None)}
|
|
super().__init__()
|
|
out_channels = out_channels or in_channels
|
|
self.radix = radix
|
|
mid_chs = out_channels * radix
|
|
if rd_channels is None:
|
|
attn_chs = make_divisible(in_channels * radix * rd_ratio, min_value=32, divisor=rd_divisor)
|
|
else:
|
|
attn_chs = rd_channels * radix
|
|
|
|
padding = kernel_size // 2 if padding is None else padding
|
|
self.conv = nn.Conv2d(
|
|
in_channels,
|
|
mid_chs,
|
|
kernel_size,
|
|
stride,
|
|
padding,
|
|
dilation,
|
|
groups=groups * radix,
|
|
bias=bias,
|
|
**kwargs,
|
|
**dd,
|
|
)
|
|
self.bn0 = norm_layer(mid_chs, **dd) if norm_layer else nn.Identity()
|
|
self.drop = drop_layer() if drop_layer is not None else nn.Identity()
|
|
self.act0 = act_layer(inplace=True)
|
|
self.fc1 = nn.Conv2d(out_channels, attn_chs, 1, groups=groups, **dd)
|
|
self.bn1 = norm_layer(attn_chs, **dd) if norm_layer else nn.Identity()
|
|
self.act1 = act_layer(inplace=True)
|
|
self.fc2 = nn.Conv2d(attn_chs, mid_chs, 1, groups=groups, **dd)
|
|
self.rsoftmax = RadixSoftmax(radix, groups)
|
|
|
|
def forward(self, x):
|
|
x = self.conv(x)
|
|
x = self.bn0(x)
|
|
x = self.drop(x)
|
|
x = self.act0(x)
|
|
|
|
B, RC, H, W = x.shape
|
|
if self.radix > 1:
|
|
x = x.reshape((B, self.radix, RC // self.radix, H, W))
|
|
x_gap = x.sum(dim=1)
|
|
else:
|
|
x_gap = x
|
|
x_gap = x_gap.mean((2, 3), keepdim=True)
|
|
x_gap = self.fc1(x_gap)
|
|
x_gap = self.bn1(x_gap)
|
|
x_gap = self.act1(x_gap)
|
|
x_attn = self.fc2(x_gap)
|
|
|
|
x_attn = self.rsoftmax(x_attn).view(B, -1, 1, 1)
|
|
if self.radix > 1:
|
|
out = (x * x_attn.reshape((B, self.radix, RC // self.radix, 1, 1))).sum(dim=1)
|
|
else:
|
|
out = x * x_attn
|
|
return out.contiguous()
|