def __init__(self,
in_channels: int,
out_channels: int,
kernel_size: int,
stride: int,
expansion: int,
activation: type = nn.Hardswish,
use_se: bool = True):
super(InvertedResidual, self).__init__()
# TODO: check if this is indeed correct
# this is the same as the implementation from
# https://github.com/d-li14/mobilenetv3.pytorch/
self.activation_after_se = use_se and expansion != 1
width = round_channels(expansion * in_channels)
self.conv1 = (ConvBnAct2d(in_channels, width, 1, activation=activation)
if expansion != 1 else nn.Identity())
self.conv2 = ConvBn2d(width,
width,
kernel_size,
padding=kernel_size // 2,
stride=stride,
groups=width)
self.act2 = activation()
self.se = (SqueezeExcitation(width, width)
if use_se else nn.Identity())
self.conv3 = ConvBn2d(width, out_channels, 1)
def __init__(self, in_channels, out_channels, reduction_ratio=4):
super(SEBlock, self).__init__()
reduced_channels = round_channels(in_channels / reduction_ratio, 4)
self.conv1 = nn.Conv2d(in_channels, reduced_channels, 1)
self.activation = nn.ReLU(inplace=True)
self.conv2 = nn.Conv2d(reduced_channels, out_channels, 1)
def c(channels):
return round_channels(channels * width_multiplier)
def c(channels):
return round_channels(width_multiplier * channels,
divisor=8 if width_multiplier >= 0.1 else 4)
def c(channels):
"channel number mapper"
return round_channels(width_multiplier * channels,
8 if width_multiplier > 0.1 else 4)
def __init__(self, in_channels, out_channels, reduction=4):
super().__init__()
mid_channels = round_channels(in_channels / reduction)
self.conv1 = nn.Conv2d(in_channels, mid_channels, 1)
self.relu = nn.ReLU(inplace=True)
self.conv2 = nn.Conv2d(mid_channels, out_channels, 1)
