特征金字塔网络python代码

特征金字塔网络（Feature Pyramid Network，FPN）是一种用于解决目标检测中物体尺度变化问题的网络结构。以下是使用Python和PyTorch实现FPN的示例代码：

import torch
import torch.nn as nn
import torch.nn.functional as F

class FPN(nn.Module):
    def __init__(self, num_channels):
        super(FPN, self).__init__()
        
        # Bottom-up layers
        self.conv1 = nn.Conv2d(3, num_channels, kernel_size=3, stride=1, padding=1)
        self.conv2 = nn.Conv2d(num_channels, num_channels, kernel_size=3, stride=1, padding=1)
        self.conv3 = nn.Conv2d(num_channels, num_channels, kernel_size=3, stride=1, padding=1)
        self.conv4 = nn.Conv2d(num_channels, num_channels, kernel_size=3, stride=1, padding=1)
        
        # Top-down layers
        self.conv5 = nn.Conv2d(num_channels, num_channels, kernel_size=3, stride=1, padding=1)
        self.conv6 = nn.Conv2d(num_channels, num_channels, kernel_size=3, stride=1, padding=1)
        self.conv7 = nn.Conv2d(num_channels, num_channels, kernel_size=3, stride=1, padding=1)
        
        # Lateral layers
        self.lateral_conv1 = nn.Conv2d(num_channels, num_channels, kernel_size=1, stride=1, padding=0)
        self.lateral_conv2 = nn.Conv2d(num_channels, num_channels, kernel_size=1, stride=1, padding=0)
        self.lateral_conv3 = nn.Conv2d(num_channels, num_channels, kernel_size=1, stride=1, padding=0)
        
        # Smooth layers
        self.smooth_conv1 = nn.Conv2d(num_channels, num_channels, kernel_size=3, stride=1, padding=1)
        self.smooth_conv2 = nn.Conv2d(num_channels, num_channels, kernel_size=3, stride=1, padding=1)
        self.smooth_conv3 = nn.Conv2d(num_channels, num_channels, kernel_size=3, stride=1, padding=1)
        
    def forward(self, x):
        # Bottom-up pathway
        c1 = F.relu(self.conv1(x))
        c2 = F.relu(self.conv2(c1))
        c3 = F.relu(self.conv3(c2))
        c4 = F.relu(self.conv4(c3))
        
        # Top-down pathway
        p5 = F.relu(self.conv5(c4))
        p4 = F.relu(self.conv6(p5) + self.lateral_conv1(c4))
        p3 = F.relu(self.conv7(p4) + self.lateral_conv2(c3))
        p2 = F.relu(self.lateral_conv3(c2))
        
        # Smooth
        p4 = F.relu(self.smooth_conv1(p4))
        p3 = F.relu(self.smooth_conv2(p3))
        p2 = F.relu(self.smooth_conv3(p2))
        
        return p2, p3, p4, p5

这是一个简单的FPN实现，其中包含两个路径：自下而上的路径和自上而下的路径。自下而上的路径生成一组具有不同尺度的特征图，自上而下的路径将这些特征图进行上采样和融合，生成一组具有更好的尺度不变性的特征图。这些特征图最终被送到目标检测网络中进行检测。