用Python做有趣的AI项目5：AI 画画机器人（图像风格迁移）

这个项目将使用 PyTorch 实现图像风格迁移（Neural Style Transfer），让一张图片看起来具有另一张图片的“艺术风格”。

开发环境建议

Python 3.8+

PyTorch（pip install torch torchvision）

PIL（pip install pillow）

CUDA（可选，但建议有 GPU）

️ 项目结构示例

bash

style_transfer/
├── style.jpg         # 风格图像（如：星夜）
├── content.jpg       # 内容图像（如：你的自拍）
├── style_transfer.py # 主程序

✅ Step-by-step：图像风格迁移项目详解（含代码）

✅ Step 1：导入库

python

import torch
import torch.nn as nn
import torch.optim as optim
import torchvision.transforms as transforms
import torchvision.models as models
from PIL import Image
import copy
import matplotlib.pyplot as plt

✅ Step 2：设备设置（使用 GPU 优先）

python

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

✅ Step 3：图像加载与预处理函数

python

def load_image(img_path, max_size=400):
    image = Image.open(img_path).convert('RGB')
    
    # 限制最大尺寸
    if max(image.size) > max_size:
        size = max_size
    else:
        size = max(image.size)

    in_transform = transforms.Compose([
        transforms.Resize(size),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406],
                             std=[0.229, 0.224, 0.225])
    ])

    image = in_transform(image).unsqueeze(0)
    return image.to(device)

✅ Step 4：定义函数来展示图像

python

def im_convert(tensor):
    image = tensor.to("cpu").clone().detach()
    image = image.squeeze(0)
    image = transforms.ToPILImage()(image)
    return image

✅ Step 5：加载内容图像和风格图像

python

content = load_image("content.jpg").to(device)
style = load_image("style.jpg").to(device)

✅ Step 6：加载预训练的 VGG 模型（VGG19）

python

vgg = models.vgg19(pretrained=True).features.to(device).eval()

我们只取模型中的某些层用于提取风格与内容特征。

✅ Step 7：定义提取内容和风格特征的函数

python

def get_features(image, model, layers=None):
    if layers is None:
        layers = {
            '0': 'conv1_1',
            '5': 'conv2_1',
            '10': 'conv3_1',
            '19': 'conv4_1',
            '21': 'conv4_2',  # content representation
            '28': 'conv5_1'
        }
        
    features = {}
    x = image
    for name, layer in model._modules.items():
        x = layer(x)
        if name in layers:
            features[layers[name]] = x
            
    return features

✅ Step 8：定义 Gram 矩阵函数（风格提取核心）

python

def gram_matrix(tensor):
    b, d, h, w = tensor.size()
    tensor = tensor.view(d, h * w)
    gram = torch.mm(tensor, tensor.t())
    return gram 

✅ Step 9：提取内容和风格特征

python

content_features = get_features(content, vgg)
style_features = get_features(style, vgg)


# 为每个风格层计算 Gram 矩阵
style_grams = {layer: gram_matrix(style_features[layer]) for layer in style_features}

✅ Step 10：初始化目标图像（从内容图像复制）

python

target = content.clone().requires_grad_(True).to(device)

✅ Step 11：设置超参数

python

style_weights = {
    'conv1_1': 1.0,
    'conv2_1': 0.75,
    'conv3_1': 0.2,
    'conv4_1': 0.2,
    'conv5_1': 0.2
}
content_weight = 1e4  # α
style_weight = 1e2    # β

✅ Step 12：设置优化器

python

optimizer = optim.Adam([target], lr=0.003)

✅ Step 13：训练模型

python

steps = 300

for i in range(1, steps+1):
    target_features = get_features(target, vgg)

    content_loss = torch.mean((target_features['conv4_2'] - content_features['conv4_2']) ** 2)

    style_loss = 0
    for layer in style_weights:
        target_feature = target_features[layer]
        target_gram = gram_matrix(target_feature)
        style_gram = style_grams[layer]
        layer_style_loss = style_weights[layer] * torch.mean((target_gram - style_gram) ** 2)
        style_loss += layer_style_loss / (target_feature.shape[1] ** 2)

    total_loss = content_weight * content_loss + style_weight * style_loss

    optimizer.zero_grad()
    total_loss.backward()
    optimizer.step()

    if i % 50 == 0:
        print(f"Step {i}, Total loss: {total_loss.item():.4f}")

✅ Step 14：保存并显示结果图像

python

final_img = im_convert(target)
final_img.save("result.jpg")
final_img.show()

️ 示例效果
将自拍和《星夜.jpg》结合，输出一张油画风格的人像。

附上完整代码：

import torch
import torch.nn as nn
import torch.optim as optim
import torchvision.transforms as transforms
import torchvision.models as models
from PIL import Image
import copy


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


# 图像加载和预处理
def load_image(img_path, max_size=400):
    image = Image.open(img_path).convert('RGB')

    if max(image.size) > max_size:
        size = max_size
    else:
        size = max(image.size)

    in_transform = transforms.Compose([
        transforms.Resize(size),
        transforms.ToTensor(),
        transforms.Normalize(
            mean=[0.485, 0.456, 0.406],
            std=[0.229, 0.224, 0.225]
        )
    ])

    image = in_transform(image).unsqueeze(0)
    return image.to(device)


# 还原为可视图像
def im_convert(tensor):
    image = tensor.to("cpu").clone().detach()
    image = image.squeeze(0)
    image = transforms.ToPILImage()(image)
    return image


# 提取特征
def get_features(image, model, layers=None):
    if layers is None:
        layers = {
            '0': 'conv1_1',
            '5': 'conv2_1',
            '10': 'conv3_1',
            '19': 'conv4_1',
            '21': 'conv4_2',  # 内容层
            '28': 'conv5_1'
        }

    features = {}
    x = image
    for name, layer in model._modules.items():
        x = layer(x)
        if name in layers:
            features[layers[name]] = x
    return features


# Gram矩阵
def gram_matrix(tensor):
    b, d, h, w = tensor.size()
    tensor = tensor.view(d, h * w)
    gram = torch.mm(tensor, tensor.t())
    return gram


# 主程序入口
def main():
    # 加载图像
    content = load_image("content.jpg")
    style = load_image("style.jpg")

    # 加载预训练模型
    vgg = models.vgg19(pretrained=True).features.to(device).eval()

    content_features = get_features(content, vgg)
    style_features = get_features(style, vgg)

    style_grams = {layer: gram_matrix(style_features[layer]) for layer in style_features}

    target = content.clone().requires_grad_(True).to(device)

    # 权重设置
    style_weights = {
        'conv1_1': 1.0,
        'conv2_1': 0.75,
        'conv3_1': 0.2,
        'conv4_1': 0.2,
        'conv5_1': 0.2
    }

    content_weight = 1e4
    style_weight = 1e2

    optimizer = optim.Adam([target], lr=0.003)
    steps = 300

    print("开始风格迁移...")
    for i in range(1, steps + 1):
        target_features = get_features(target, vgg)

        content_loss = torch.mean((target_features['conv4_2'] - content_features['conv4_2']) ** 2)

        style_loss = 0
        for layer in style_weights:
            target_feature = target_features[layer]
            target_gram = gram_matrix(target_feature)
            style_gram = style_grams[layer]
            layer_style_loss = style_weights[layer] * torch.mean((target_gram - style_gram) ** 2)
            style_loss += layer_style_loss / (target_feature.shape[1] ** 2)

        total_loss = content_weight * content_loss + style_weight * style_loss

        optimizer.zero_grad()
        total_loss.backward()
        optimizer.step()

        if i % 50 == 0:
            print(f"Step {i}/{steps}, Total loss: {total_loss.item():.4f}")

    # 保存结果
    result = im_convert(target)
    result.save("result.jpg")
    print("风格迁移完成！结果保存在 result.jpg")


# 运行主函数
if __name__ == "__main__":
    main()

✅ 使用说明

️ 准备：

把你的内容图命名为 content.jpg

把你的风格图命名为 style.jpg

放在与 style_transfer.py 同一个目录下

▶️ 运行：

bash

python style_transfer.py

#️ 输出：
运行成功后，生成的图像将保存在：

result.jpg