import torch import torch.nn as nn import torch.nn.functional as F from torchvision import transforms from PIL import Image class FoodCNN(nn.Module): """ 食物分类CNN模型 基于CIFAR10结构,适配3分类任务 use_internal_preprocess 是否在模型内部预处理 训练的时候, """ def __init__(self, use_internal_preprocess=False): super(FoodCNN, self).__init__() self.use_internal_preprocess = use_internal_preprocess # 图片预处理变换(仅在推理时使用) self.preprocess = transforms.Compose([ transforms.Resize((32, 32),interpolation=transforms.InterpolationMode.BILINEAR), transforms.ToTensor(), transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)) ]) # 内部预处理变换(用于已经是tensor但未归一化的数据) self.internal_preprocess = transforms.Compose([ # 不包含Resize,因为训练时已经在DataLoader中处理了 transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)) ]) # 第一个卷积块 self.conv1 = nn.Conv2d(3, 32, 3, padding=1) self.conv2 = nn.Conv2d(32, 32, 3, padding=1) self.pool1 = nn.MaxPool2d(2, 2) self.dropout1 = nn.Dropout2d(0.25) # 第二个卷积块 self.conv3 = nn.Conv2d(32, 64, 3, padding=1) self.conv4 = nn.Conv2d(64, 64, 3, padding=1) self.pool2 = nn.MaxPool2d(2, 2) self.dropout2 = nn.Dropout2d(0.25) # 第三个卷积块 self.conv5 = nn.Conv2d(64, 128, 3, padding=1) self.conv6 = nn.Conv2d(128, 128, 3, padding=1) self.pool3 = nn.MaxPool2d(2, 2) self.dropout3 = nn.Dropout2d(0.25) # 全连接层 self.fc1 = nn.Linear(128 * 4 * 4, 512) self.dropout4 = nn.Dropout(0.5) self.fc2 = nn.Linear(512, 3) # 3分类 def preprocess_image(self, image): """ 预处理单张图片(使用与移动端相同的插值方法) Args: image: PIL Image 或 numpy array Returns: torch.Tensor: 预处理后的张量,形状为 (1, 3, 32, 32) """ if not isinstance(image, Image.Image): # 如果是numpy array,转换为PIL Image if hasattr(image, 'shape'): image = Image.fromarray(image) else: raise ValueError("输入必须是PIL Image或numpy array") # 转换为tensor(不做resize) tensor = transforms.ToTensor()(image) # 添加batch维度 tensor = tensor.unsqueeze(0) # 使用与移动端相同的插值方法缩放到32x32 tensor = F.interpolate(tensor, size=(32, 32), mode='bilinear', align_corners=False) # ImageNet标准化 mean = torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1) std = torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1) tensor = (tensor - mean) / std return tensor def mobile_preprocess(self, x): """ 移动端预处理(TorchScript兼容) Args: x: 输入tensor,形状为 [batch, 3, height, width],值范围 0-255 Returns: torch.Tensor: 预处理后的tensor,形状为 [batch, 3, 32, 32] """ # 归一化到 [0, 1] # 好像安卓已经做了归一化了。 # x = x.float() / 255.0 # 缩放到 32x32 x = F.interpolate(x, size=(32, 32), mode='bilinear', align_corners=False) # 缩放到 32x32 - 使用 align_corners=True 来匹配 PIL 的默认行为 # x = F.interpolate(x, size=(32, 32), mode='bilinear', align_corners=True) # ImageNet标准化 mean = torch.tensor([0.485, 0.456, 0.406], device=x.device).view(1, 3, 1, 1) std = torch.tensor([0.229, 0.224, 0.225], device=x.device).view(1, 3, 1, 1) x = (x - mean) / std return x def _forward_network(self, x): """ 网络的核心前向传播(不包含预处理) Args: x: 预处理后的tensor,形状为 [batch, 3, 32, 32] Returns: torch.Tensor: 模型输出 """ # 第一个卷积块 x = F.relu(self.conv1(x)) x = F.relu(self.conv2(x)) x = self.pool1(x) x = self.dropout1(x) # 第二个卷积块 x = F.relu(self.conv3(x)) x = F.relu(self.conv4(x)) x = self.pool2(x) x = self.dropout2(x) # 第三个卷积块 x = F.relu(self.conv5(x)) x = F.relu(self.conv6(x)) x = self.pool3(x) x = self.dropout3(x) # 展平 x = x.view(-1, 128 * 4 * 4) # 全连接层 x = F.relu(self.fc1(x)) x = self.dropout4(x) x = self.fc2(x) return x def forward(self, x): # 如果启用内部预处理且输入是tensor if self.use_internal_preprocess and isinstance(x, torch.Tensor): x = self.internal_preprocess(x) return self._forward_network(x) def forward_mobile(self, x): """ 移动端前向传播(包含预处理) Args: x: 输入tensor,形状为 [batch, 3, height, width],值范围 0-255 Returns: torch.Tensor: 模型输出 """ # 移动端预处理 x = self.mobile_preprocess(x) # 执行网络前向传播 return self._forward_network(x) def create_food_cnn(use_internal_preprocess=False): """ 创建食物分类CNN模型 Args: use_internal_preprocess: 是否在forward中进行预处理 Returns: FoodCNN: 网络模型实例 """ return FoodCNN(use_internal_preprocess=use_internal_preprocess) def create_mobile_food_cnn(): """ 创建移动端食物分类CNN模型 Returns: FoodCNN: 配置为移动端使用的网络模型实例 """ class MobileFoodCNN(FoodCNN): """ 移动端食物分类模型 重写forward方法以包含预处理 """ def __init__(self): super().__init__(use_internal_preprocess=False) def forward(self, x): """ 移动端前向传播(自动包含预处理) Args: x: 输入tensor,形状为 [batch, 3, height, width],值范围 0-255 Returns: torch.Tensor: 模型输出 """ return self.forward_mobile(x) return MobileFoodCNN() if __name__ == "__main__": # 测试网络 model = create_food_cnn() print(f"模型参数数量: {sum(p.numel() for p in model.parameters() if p.requires_grad)}") # 测试前向传播 dummy_input = torch.randn(1, 3, 32, 32) output = model(dummy_input) print(f"输出形状: {output.shape}") # 测试图片预处理 try: import numpy as np # 创建一个测试图片 (RGB格式) test_image = Image.fromarray(np.random.randint(0, 255, (224, 224, 3), dtype=np.uint8)) processed = model.preprocess_image(test_image) print(f"预处理后图片形状: {processed.shape}") print(f"预处理后数值范围: [{processed.min():.3f}, {processed.max():.3f}]") except Exception as e: print(f"预处理测试失败: {e}")