图片风格迁移(Neural style)

图片风格迁移训练时使用预处理的模型，不进行模型的训练，将图片像素值设置为可训练参数，进行图片的训练。

结合一张图片的内容和另一张图片的风格，生成一张新图片

论文地址：https://arxiv.org/pdf/1508.06576.pdf

损失函数：

 

以下代码通过将图片x输入到预训练好的vgg19模型中，取模型features模块的0, 5, 10, 19, 28五层的输出值计算contentLoss与StyleLoss之和，对该loss进行反向传播，从而修改图片的像素值。

"""
图片风格迁移，使用预训练的vgg，不训练vgg模型，而是进行图片的训练，训练我们想要的风格图片
"""


import torch
from PIL import Image
import argparse
import torch.nn as nn
import matplotlib.pyplot as plt
import numpy as np
import torchvision


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


def load_image(image_path, transform=None, max_size=None, shape=None):
    image = Image.open(image_path)
    if max_size:
        scale = max_size/max(image.size)
        size = np.array(image.size)*scale
        image = image.resize(size.astype(int), Image.ANTIALIAS)

    if shape:
        image = image.resize(shape, Image.LANCZOS)
    if transform:
        image = transform(image).unsqueeze(0)
    return image.to(device)


transform = torchvision.transforms.Compose([
    torchvision.transforms.ToTensor(),
    # mean and std based on trainDatasets
    torchvision.transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

content = load_image("data/content.jpg", transform, max_size=400)
style = load_image("data/style.jpg", transform, shape=[content.size(2), content.size(3)])


print(content.shape)  # torch.Size([1, 3, 400, 300])
print(style.shape)   # torch.Size([1, 3, 300, 400])


def imshow(tensor, title=None):
    image = tensor.cpu().clone()
    image = image.squeeze(0)
    # ToPILImage为类对象，因此需要ToPILImage()(image)
    image = torchvision.transforms.ToPILImage()(image)
    plt.imshow(image)
    if title is not None:
        plt.title(title)
    plt.pause(2)  # pause a bit so that plots are updated


class VGGNet(nn.Module):
    def __init__(self):
        super(VGGNet, self).__init__()
        self.select = ['0', '5', '10', '19', '28']
        self.vgg = torchvision.models.vgg19(pretrained=True).features

    def forward(self, x):
        features = []
        for name, layer in self.vgg._modules.items():
            x = layer(x)
            if name in self.select:
                features.append(x)
        return features

# 不训练模型，eval()模式
vgg = VGGNet().to(device).eval()

for feat in vgg(content):
    print(feat.shape)



# 训练图片
target = content.clone().requires_grad_(True)
print("target shape: ", target.shape)
optimizer = torch.optim.Adam([target], lr=0.003, betas=[0.5, 0.999])


num_steps = 2000

for step in range(num_steps):
    target_features = vgg(target)
    content_features = vgg(content)
    style_features = vgg(style)

    content_loss = style_loss = 0.
    for f1, f2, f3 in zip(target_features, content_features, style_features):
        content_loss += torch.mean((f1-f2)**2)
        _, c, h, w = f1.size()
        f1 = f1.view(c, h*w)
        f3 = f3.view(c, h*w)

        f1 = torch.mm(f1, f1.t())
        f3 = torch.mm(f3, f3.t())
        style_loss += torch.mean((f1-f3)**2)/(c*h*w)

    loss = content_loss + style_loss*100.

    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

    if step % 100 == 0:
        print("Step [{}/{}, Content Loss: {:.4f}, Style Loss: {:.4f}"
              .format(step, num_steps, content_loss.item(), style_loss.item()))


denorm = torchvision.transforms.Normalize([-2.12, -2.04, -1.8], [4.37, 4.46, 4.44])
img = target.clone().squeeze()
img = denorm(img).clamp_(0, 1)
imshow(img, title="Target Image")
plt.pause(100)

 

来源：oschina

链接：https://my.oschina.net/u/4228078/blog/4467847