胶囊网络实现手写数字分类

前言

胶囊网络的概念可以先行搜索。

一、完整代码

import torch
import torch.nn.functional as F
from torch import nn
from torchvision import transforms, datasets
from torch.optim import Adam
from torch.utils.data import DataLoader

# 定义胶囊网络中的胶囊层
class CapsuleLayer(nn.Module):
    def __init__(self, num_capsules, num_route_nodes, in_channels, out_channels, kernel_size=None, stride=None,
                 num_iterations=3):
        super(CapsuleLayer, self).__init__()

        self.num_route_nodes = num_route_nodes
        self.num_iterations = num_iterations
        self.num_capsules = num_capsules

        if num_route_nodes != -1:
            self.route_weights = nn.Parameter(torch.randn(num_capsules, num_route_nodes, in_channels, out_channels))
        else:
            self.capsules = nn.ModuleList([
                nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=0)
                for _ in range(num_capsules)])

    def squash(self, tensor, dim=-1):
        squared_norm = (tensor ** 2).sum(dim=dim, keepdim=True)
        scale = squared_norm / (1 + squared_norm)
        return scale * tensor / torch.sqrt(squared_norm)

    def forward(self, x):
        if self.num_route_nodes != -1:
            priors = x[None, :, :, None, :] @ self.route_weights[:, None, :, :, :]
            logits = torch.zeros(*priors.size()).to(x.device)
            for i in range(self.num_iterations):
                probs = F.softmax(logits, dim=2)
                outputs = self.squash((probs * priors).sum(dim=2, keepdim=True))
                if i != self.num_iterations - 1:
                    delta_logits = (priors * outputs).sum(dim=-1, keepdim=True)
                    logits = logits + delta_logits
        else:
            outputs = [capsule(x).view(x.size(0), -1, 1) for capsule in self.capsules]
            outputs = torch.cat(outputs, dim=-2)
            outputs = self.squash(outputs)

        return outputs

# 定义整个胶囊网络模型
class CapsuleNet(nn.Module):
    def __init__(self):
        super(CapsuleNet, self).__init__()
        
        self.conv1 = nn.Conv2d(in_channels=1, out_channels=256, kernel_size=9, stride=1)
        self.primary_capsules = CapsuleLayer(num_capsules=8, num_route_nodes=-1, in_channels=256, out_channels=32,
                                             kernel_size=9, stride=2)
        self.digit_capsules = CapsuleLayer(num_capsules=10, num_route_nodes=32 * 6 * 6, in_channels=8,
                                           out_channels=16)

    def forward(self, x):
        x = F.relu(self.conv1(x), inplace=True)
        x = self.primary_capsules(x)
        x = self.digit_capsules(x).squeeze().transpose(0, 1)
        x = (x ** 2).sum(dim=-1) ** 0.5
        return x

# 训练和评估
def train(model, train_loader, optimizer, epoch):
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):
        data, target = data.to(device), target.to(device)
        optimizer.zero_grad()
        output = model(data)
        loss = F.cross_entropy(output, target)
        loss.backward()
        optimizer.step()

        if batch_idx % 10 == 0:
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                epoch, batch_idx * len(data), len(train_loader.dataset),
                       100. * batch_idx / len(train_loader), loss.item()))

def test(model, test_loader):
    model.eval()
    test_loss = 0
    correct = 0
    with torch.no_grad():
        for data, target in test_loader:
            data, target = data.to(device), target.to(device)
            output = model(data)
            test_loss += F.cross_entropy(output, target, reduction='sum').item()
            pred = output.argmax(dim=1, keepdim=True)
            correct += pred.eq(target.view_as(pred)).sum().item()

    test_loss /= len(test_loader.dataset)
    print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
        test_loss, correct, len(test_loader.dataset),
        100. * correct / len(test_loader.dataset)))

# 数据加载和预处理
transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.1307,), (0.3081,))
])

train_dataset = datasets.MNIST(root='./data', train=True, download=True, transform=transform)
test_dataset = datasets.MNIST(root='./data', train=False, transform=transform)

train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=64, shuffle=True)

# 设置设备
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# 初始化模型和优化器
model = CapsuleNet().to(device)
optimizer = Adam(model.parameters())

# 训练和测试模型
num_epochs = 10
for epoch in range(num_epochs):
    train(model, train_loader, optimizer, epoch)
    test(model, test_loader)

二、修改成自己的数据集

以下几个位置是需要修改的。

# 数据加载和预处理
transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.1307,), (0.3081,))
])

train_dataset = datasets.MNIST(root='./data', train=True, download=True, transform=transform)
test_dataset = datasets.MNIST(root='./data', train=False, transform=transform)

train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=64, shuffle=True)

这些位置要根据数据集实际情况修改。主要是如果分辨率修改了，那么下面的也要跟着修改。

self.conv1 = nn.Conv2d(in_channels=1, out_channels=256, kernel_size=9, stride=1)
self.primary_capsules = CapsuleLayer(num_capsules=8, num_route_nodes=-1, in_channels=256, out_channels=32, kernel_size=9, stride=2)
self.digit_capsules = CapsuleLayer(num_capsules=10, num_route_nodes=32 * 6 * 6, in_channels=8,out_channels=16)

修改这3行代码很容易报错。要理解了以后修改。

总结

多试试。