- 前置知识：MobileNet, ResNet
- 作者：Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, Liang-Chieh Chen
- 文章链接
- 代码链接
在原始MobileNet中，使用Depthwise Separable Convolution来降低计算量。但是这个架构在实际训练中常常会使得大多数卷积核参数为0，而在ReLU进一步影响下，会使得神经元最终的输出为0。输出为0会导致梯度为0，使得整体网络无法进行优化训练。因此，Andrew等人在MobileNet基础上进行改进，提出了MobileNetV2。
MobileNetV2改进点主要在于两个部分，Linear Bottlenecks和Inverted residuals。
这部分可能是文中最难以理解的部分，论文不讲人话，读起来总是玄之又玄。论文中提到了一个概念，manifold of interest，它是指多层神经网络以及激活函数输出张量的一个集合。例如在ResNet34中，总共分为了4层，那么每一层的输出都是一个manifold of interest。
长久以来，研究人员认为每个像素特征是被编码在这个manifold of interest之中的，而我们将manifold of interest送入神经网络，实际上是将manifold of interest进行一个空间变换。可以膨胀到更高维的空间，也可以嵌入到低维的子空间。但是，论文认为神经网络使用非线性激活函数会造成空间映射的坍塌。
图1：Manifold of Interest
如图所示，输入是一个2D数据，其中manifold of interest是其中的蓝色螺旋线。通过矩阵$T$将输入嵌入到高维，再经过ReLU，随后用$T^{-1}$反投影到2D空间。当嵌入的维度较高时，2D数据可以较好的恢复，但是维度较低时，反投影回来的特征空间就发生了坍塌。这就意味着，如果我们既想要提升效果，又想要降低维度，采用ReLU是不现实的。因为ReLU函数可能会滤除很多有用信息。
因此，论文采用Linear Bottleneck进行替代，实际上也就是一个linear layer加上一个batch normalization。
在ResNet原有的残差连接中，是中间小两头大的架构。也就是两头的通道数大，中间的通道数小。而在MobileNetV2中，则是中间大两头小的架构，所以文章作者将其称之为Inverted residuals。这个主要还是因为采用了Depthwise Convolution，从而降低计算量。
图2：Residuals架构对比
MobileNetV2整体架构如图3所示。其中，每一个bottleneck都是图4所示的结构。
图3：MobileNetV2架构
图4：Bottleneck架构
class MobileNetV2(nn.Module):
def __init__(
self,
num_classes: int = 1000,
width_mult: float = 1.0,
inverted_residual_setting: Optional[List[List[int]]] = None,
round_nearest: int = 8,
block: Optional[Callable[..., nn.Module]] = None,
norm_layer: Optional[Callable[..., nn.Module]] = None,
dropout: float = 0.2,
) -> None:
"""
MobileNet V2 main class
Args:
num_classes (int): Number of classes
width_mult (float): Width multiplier - adjusts number of channels in each layer by this amount
inverted_residual_setting: Network structure
round_nearest (int): Round the number of channels in each layer to be a multiple of this number
Set to 1 to turn off rounding
block: Module specifying inverted residual building block for mobilenet
norm_layer: Module specifying the normalization layer to use
dropout (float): The droupout probability
"""
super().__init__()
_log_api_usage_once(self)
if block is None:
block = InvertedResidual
if norm_layer is None:
norm_layer = nn.BatchNorm2d
input_channel = 32
last_channel = 1280
if inverted_residual_setting is None:
inverted_residual_setting = [
# t, c, n, s
[1, 16, 1, 1],
[6, 24, 2, 2],
[6, 32, 3, 2],
[6, 64, 4, 2],
[6, 96, 3, 1],
[6, 160, 3, 2],
[6, 320, 1, 1],
]
# only check the first element, assuming user knows t,c,n,s are required
if len(inverted_residual_setting) == 0 or len(inverted_residual_setting[0]) != 4:
raise ValueError(
f"inverted_residual_setting should be non-empty or a 4-element list, got {inverted_residual_setting}"
)
# building first layer
input_channel = _make_divisible(input_channel * width_mult, round_nearest)
self.last_channel = _make_divisible(last_channel * max(1.0, width_mult), round_nearest)
features: List[nn.Module] = [
Conv2dNormActivation(3, input_channel, stride=2, norm_layer=norm_layer, activation_layer=nn.ReLU6)
]
# building inverted residual blocks
for t, c, n, s in inverted_residual_setting:
output_channel = _make_divisible(c * width_mult, round_nearest)
for i in range(n):
stride = s if i == 0 else 1
features.append(block(input_channel, output_channel, stride, expand_ratio=t, norm_layer=norm_layer))
input_channel = output_channel
# building last several layers
features.append(
Conv2dNormActivation(
input_channel, self.last_channel, kernel_size=1, norm_layer=norm_layer, activation_layer=nn.ReLU6
)
)
# make it nn.Sequential
self.features = nn.Sequential(*features)
# building classifier
self.classifier = nn.Sequential(
nn.Dropout(p=dropout),
nn.Linear(self.last_channel, num_classes),
)
# weight initialization
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode="fan_out")
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.ones_(m.weight)
nn.init.zeros_(m.bias)
elif isinstance(m, nn.Linear):
nn.init.normal_(m.weight, 0, 0.01)
nn.init.zeros_(m.bias)
def _forward_impl(self, x: Tensor) -> Tensor:
# This exists since TorchScript doesn't support inheritance, so the superclass method
# (this one) needs to have a name other than `forward` that can be accessed in a subclass
x = self.features(x)
# Cannot use "squeeze" as batch-size can be 1
x = nn.functional.adaptive_avg_pool2d(x, (1, 1))
x = torch.flatten(x, 1)
x = self.classifier(x)
return x
def forward(self, x: Tensor) -> Tensor:
return self._forward_impl(x)上述代码是PyTorch的官方实现，整体架构比较简单。
class InvertedResidual(nn.Module):
def __init__(
self, inp: int, oup: int, stride: int, expand_ratio: int, norm_layer: Optional[Callable[..., nn.Module]] = None
) -> None:
super().__init__()
self.stride = stride
if stride not in [1, 2]:
raise ValueError(f"stride should be 1 or 2 insted of {stride}")
if norm_layer is None:
norm_layer = nn.BatchNorm2d
hidden_dim = int(round(inp * expand_ratio))
self.use_res_connect = self.stride == 1 and inp == oup
layers: List[nn.Module] = []
if expand_ratio != 1:
# pw
layers.append(
Conv2dNormActivation(inp, hidden_dim, kernel_size=1, norm_layer=norm_layer, activation_layer=nn.ReLU6)
)
layers.extend(
[
# dw
Conv2dNormActivation(
hidden_dim,
hidden_dim,
stride=stride,
groups=hidden_dim,
norm_layer=norm_layer,
activation_layer=nn.ReLU6,
),
# pw-linear
nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
norm_layer(oup),
]
)
self.conv = nn.Sequential(*layers)
self.out_channels = oup
self._is_cn = stride > 1
def forward(self, x: Tensor) -> Tensor:
if self.use_res_connect:
return x + self.conv(x)
else:
return self.conv(x)这个也就是Bottleneck架构。
采用linear layer替换ReLU，并引入残差连接，在降低参数计算量的同时提高性能。