You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
pytorch-image-models/timm/models/layers/squeeze_excite.py

75 lines
2.9 KiB

""" Squeeze-and-Excitation Channel Attention
An SE implementation originally based on PyTorch SE-Net impl.
Has since evolved with additional functionality / configuration.
Paper: `Squeeze-and-Excitation Networks` - https://arxiv.org/abs/1709.01507
Also included is Effective Squeeze-Excitation (ESE).
Paper: `CenterMask : Real-Time Anchor-Free Instance Segmentation` - https://arxiv.org/abs/1911.06667
Hacked together by / Copyright 2021 Ross Wightman
"""
from torch import nn as nn
from .create_act import create_act_layer
from .helpers import make_divisible
class SEModule(nn.Module):
""" SE Module as defined in original SE-Nets with a few additions
Additions include:
* divisor can be specified to keep channels % div == 0 (default: 8)
* reduction channels can be specified directly by arg (if rd_channels is set)
* reduction channels can be specified by float rd_ratio (default: 1/16)
* global max pooling can be added to the squeeze aggregation
* customizable activation, normalization, and gate layer
"""
def __init__(
self, channels, rd_ratio=1. / 16, rd_channels=None, rd_divisor=8, add_maxpool=False,
act_layer=nn.ReLU, norm_layer=None, gate_layer='sigmoid'):
super(SEModule, self).__init__()
self.add_maxpool = add_maxpool
if not rd_channels:
rd_channels = make_divisible(channels * rd_ratio, rd_divisor, round_limit=0.)
self.fc1 = nn.Conv2d(channels, rd_channels, kernel_size=1, bias=True)
self.bn = norm_layer(rd_channels) if norm_layer else nn.Identity()
self.act = create_act_layer(act_layer, inplace=True)
self.fc2 = nn.Conv2d(rd_channels, channels, kernel_size=1, bias=True)
self.gate = create_act_layer(gate_layer)
def forward(self, x):
x_se = x.mean((2, 3), keepdim=True)
if self.add_maxpool:
# experimental codepath, may remove or change
x_se = 0.5 * x_se + 0.5 * x.amax((2, 3), keepdim=True)
x_se = self.fc1(x_se)
x_se = self.act(self.bn(x_se))
x_se = self.fc2(x_se)
return x * self.gate(x_se)
SqueezeExcite = SEModule # alias
class EffectiveSEModule(nn.Module):
""" 'Effective Squeeze-Excitation
From `CenterMask : Real-Time Anchor-Free Instance Segmentation` - https://arxiv.org/abs/1911.06667
"""
def __init__(self, channels, add_maxpool=False, gate_layer='hard_sigmoid'):
super(EffectiveSEModule, self).__init__()
self.add_maxpool = add_maxpool
self.fc = nn.Conv2d(channels, channels, kernel_size=1, padding=0)
self.gate = create_act_layer(gate_layer)
def forward(self, x):
x_se = x.mean((2, 3), keepdim=True)
if self.add_maxpool:
# experimental codepath, may remove or change
x_se = 0.5 * x_se + 0.5 * x.amax((2, 3), keepdim=True)
x_se = self.fc(x_se)
return x * self.gate(x_se)
EffectiveSqueezeExcite = EffectiveSEModule # alias