MoGA

MoGA"/>

MoGA

class MoGaA(nn.Module):def __init__(self, n_class=1000, input_size=224):super(MoGaA, self).__init__()assert input_size % 32 == 0  # 输入数据要是32的整数倍# 参数配置mb_config = [# expansion, out_channel, kernel_size, stride, act(0 RE 1 Hs), se# 扩张倍数，输出通道数，卷积大小，步长，激活函数（relu 或 Hswish），se模块[6, 24, 5, 2, 0, 0],  # 5x5卷积[6, 24, 7, 1, 0, 0],  # 7x7卷积[6, 40, 3, 2, 0, 0],  # 3x3卷积[6, 40, 3, 1, 0, 1],  # 3x3卷积 se[3, 40, 3, 1, 0, 1],  # 3x3卷积 se[6, 80, 3, 2, 1, 1],  # 3x3卷积 Hs se[6, 80, 3, 1, 1, 0],  # 3x3卷积 Hs[6, 80, 7, 1, 1, 0],  # 7x7卷积 Hs[3, 80, 7, 1, 1, 1],  # 7x7卷积 Hs se[6, 112, 7, 1, 1, 0],  # 7x7卷积 Hs[6, 112, 3, 1, 1, 0],  # 3x3卷积 Hs[6, 160, 3, 2, 1, 0],  # 3x3卷积 Hs[6, 160, 5, 1, 1, 1],  # 5x5卷积 Hs se[6, 160, 5, 1, 1, 1],  # 5x5卷积 Hs se]first_filter = 16  # 第一个通道数second_filter = 16  # 第二个通道数second_last_filter = 960  # 倒数第二个通道数last_channel = 1280  # 最后一个通道数self.last_channel = last_channel  # 1280self.stem = stem(3, first_filter, 2)  #  第一个模块self.separable_conv = separable_conv(first_filter, second_filter)  # 第二个模块 可分离卷积self.mb_module = list()  # mb模块listinput_channel = second_filter  # 输入通道数16 for t, c, k, s, a, se in mb_config:  # 遍历配置参数# 通道数扩张率t，输出通道数c，卷积核大小k，步长s，激活函数选择a，se模块选择seoutput_channel = c  # 输出通道数24act = nn.ReLU(inplace=True) if a == 0 else Hswish(inplace=True)  # 激活函数选择self.mb_module.append(InvertedResidual(input_channel, output_channel, k, s, expand_ratio=t, act=act, se=se != 0))input_channel = output_channel  # 更新输入通道数self.mb_module = nn.Sequential(*self.mb_module)  # 类型转换为nn.Sequentialself.conv_before_pooling = conv_before_pooling(input_channel, second_last_filter)  # 在pool前的卷积操作self.global_pooling = nn.AdaptiveAvgPool2d(1)  # 自适应均值池化，输出大小为（1，1）self.conv_head = conv_head(second_last_filter, last_channel)  #  在pool后的卷积操作self.classifier = classifier(last_channel, n_class)  # 分类器self._initialize_weights()  # 初始化def forward(self, x):x = self.stem(x)x = self.separable_conv(x)x = self.mb_module(x)x = self.conv_before_pooling(x)x = self.global_pooling(x)x = self.conv_head(x)x = x.view(x.size(0), -1)x = self.classifier(x)return xdef _initialize_weights(self):# 初始化参数for m in self.modules():if isinstance(m, nn.Conv2d):n = m.kernel_size[0] * m.kernel_size[1] * m.out_channelsm.weight.data.normal_(0, math.sqrt(2. / n))if m.bias is not None:m.bias.data.zero_()elif isinstance(m, nn.BatchNorm2d):m.weight.data.fill_(1)m.bias.data.zero_()elif isinstance(m, nn.Linear):n = m.weight.size(0)  # fan-outinit_range = 1.0 / math.sqrt(n)m.weight.data.uniform_(-init_range, init_range)m.bias.data.zero_()

MoGaB参数设置mb_config = [# expansion, out_channel, kernel_size, stride, act(0 RE 1 Hs), se[3, 24, 3, 2, 0, 0],[3, 24, 3, 1, 0, 0],[6, 40, 7, 2, 0, 0],[3, 40, 3, 1, 0, 0],[6, 40, 5, 1, 0, 0],[6, 80, 3, 2, 1, 1],[6, 80, 5, 1, 1, 1],[3, 80, 3, 1, 1, 0],[6, 80, 7, 1, 1, 1],[6, 112, 7, 1, 1, 0],[3, 112, 5, 1, 1, 0],[6, 160, 7, 2, 1, 1],[6, 160, 7, 1, 1, 1],[6, 160, 3, 1, 1, 1],]
MoGaC参数设置
mb_config = [# expansion, out_channel, kernel_size, stride, act(0 RE 1 Hs), se[3, 24, 5, 2, 0, 0],[3, 24, 3, 1, 0, 0],[3, 40, 5, 2, 0, 0],[3, 40, 3, 1, 0, 0],[3, 40, 5, 1, 0, 0],[3, 80, 5, 2, 1, 0],[6, 80, 5, 1, 1, 1],[3, 80, 5, 1, 1, 0],[3, 80, 5, 1, 1, 0],[6, 112, 3, 1, 1, 0],[6, 112, 3, 1, 1, 1],[6, 160, 3, 2, 1, 1],[6, 160, 3, 1, 1, 1],[6, 160, 3, 1, 1, 1],]

def stem(inp, oup, stride):# 输入通道数，输出通道数，步长return nn.Sequential(nn.Conv2d(inp, oup, 3, stride, 1, bias=False),  # 3x3卷积nn.BatchNorm2d(oup),# nn.ReLU6(inplace=True)Hswish()  # Hs激活函数)

def separable_conv(inp, oup):# 可分离卷积return nn.Sequential(nn.Conv2d(inp, inp, 3, 1, 1, groups=inp, bias=False),  # 3x3分离卷积nn.BatchNorm2d(inp),nn.ReLU(inplace=True),nn.Conv2d(inp, oup, 1, 1, 0, bias=False),  # 1x1卷积nn.BatchNorm2d(oup),)

class InvertedResidual(nn.Module):def __init__(self, inp, oup, kernel_size, stride, expand_ratio, act, se):super(InvertedResidual, self).__init__()assert stride in [1, 2]self.stride = strideself.act = actself.se = sepadding = kernel_size // 2  # padding补零hidden_dim = round(inp * expand_ratio)  # 中间隐藏层的通道数，扩增expand_ratio倍self.use_res_connect = self.stride == 1 and inp == oup    # 步长为1且输入与输出通道数相等，则用残差self.conv1 = nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False)  # 1x1卷积self.bn1 = nn.BatchNorm2d(hidden_dim)self.conv2 = nn.Conv2d(hidden_dim, hidden_dim, kernel_size, stride, padding, groups=hidden_dim, bias=False)  # kernel_size x kernel_size 分离卷积self.bn2 = nn.BatchNorm2d(hidden_dim)if self.se:  # se模块self.mid_se = SEModule(hidden_dim, act)  self.conv3 = nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False)self.bn3 = nn.BatchNorm2d(oup)def forward(self, x):inputs = x  # 原始输入x = self.conv1(x)x = self.bn1(x)x = self.act(x)x = self.conv2(x)x = self.bn2(x)x = self.act(x)if self.se:  # se模块计算x = self.mid_se(x)x = self.conv3(x)x = self.bn3(x)if self.use_res_connect:  # 残差连接return inputs + xelse:return x

class SEModule(nn.Module):# se模块def __init__(self, channel, act, reduction=4):# reduction = 4 中间层数除以4super(SEModule, self).__init__()self.avg_pool = nn.AdaptiveAvgPool2d(1)  # 自适应均值池化，输出大小为（1，1）self.conv = nn.Sequential(nn.Conv2d(channel, channel // reduction, 1, 1, 0, bias=True),  # 1x1卷积act)self.fc = nn.Sequential(nn.Conv2d(channel // reduction, channel, 1, 1, 0, bias=True),  # 1x1卷积Hsigmoid()  #激活函数)def forward(self, x):y = self.avg_pool(x)y = self.conv(y)y = self.fc(y)return torch.mul(x, y)  # torch.mul是矩阵x和y对应位相乘

class Hswish(nn.Module):# Hswish激活函数def __init__(self, inplace=True):super(Hswish, self).__init__()self.inplace = inplacedef forward(self, x):return x * F.relu6(x + 3., inplace=self.inplace) / 6.  # 与Hsigmoid相比多乘了个xclass Hsigmoid(nn.Module):# Hsigmoid激活函数def __init__(self, inplace=True):super(Hsigmoid, self).__init__()self.inplace = inplacedef forward(self, x):return F.relu6(x + 3., inplace=self.inplace) / 6.

def conv_before_pooling(inp, oup):return nn.Sequential(nn.Conv2d(inp, oup, 1, 1, 0, bias=False),  # 1x1卷积nn.BatchNorm2d(oup),# nn.ReLU6(inplace=True)Hswish())def conv_head(inp, oup):return nn.Sequential(nn.Conv2d(inp, oup, 1, bias=False),  # 1x1卷积Hswish(inplace=True),nn.Dropout2d(0.2)  # dropout层)def classifier(inp, nclass):# 分类器return nn.Linear(inp, nclass)

论文链接
代码链接