244 lines
7.9 KiB
Python
244 lines
7.9 KiB
Python
import torch
|
||
import torch.nn as nn
|
||
import torch.nn.functional as F
|
||
from torchvision import transforms
|
||
from PIL import Image
|
||
from settings import settings
|
||
|
||
|
||
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）（32）改为（224）
|
||
self.preprocess = transforms.Compose([
|
||
transforms.Resize((224, 224),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.fc1 = nn.Linear(128 * 28 * 28, 512)
|
||
self.dropout4 = nn.Dropout(0.5)
|
||
self.fc2 = nn.Linear(512, settings.NUM_CLASSES)
|
||
|
||
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)
|
||
tensor = F.interpolate(tensor, size=(224, 224), 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)
|
||
x = F.interpolate(x, size=(224, 224), 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)
|
||
x = x.view(-1, 128 * 28 * 28)
|
||
# x = x.view(-1, 65536)
|
||
|
||
# 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}") |