学习参考来自：

• Image Style Transform–关于图像风格迁移的介绍
• github：https://github.com/wmn7/ML_Practice/tree/master/2019_06_03

---

---

风格迁移

风格迁移出处：

《A Neural Algorithm of Artistic Style》（arXiv-2015）

风格迁移的实现

让 Random Image 在内容上可以接近 Content Image，在风格上可以接近 Style Image，当然， Random Image 可以初始化为 Content Image

导入基本库，数据读取

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optimfrom PIL import Image
import matplotlib.pyplot as pltimport torchvision.transforms as transforms
import torchvision.models as modelsimport numpy as np
import copy
import osdevice = torch.device("cuda" if torch.cuda.is_available() else "cpu")def image_loader(image_name, imsize):loader = transforms.Compose([transforms.Resize(imsize),  # scale imagestransforms.ToTensor()])image = Image.open(image_name).convert("RGB")image = loader(image).unsqueeze(0)return image.to(device, torch.float)def image_util(img_size=512, style_img="./1.jpg", content_img="./2.jpg"):"the size of style_img and contend_img should be same"imsize = img_size if torch.cuda.is_available() else 128  # use small size if no gpustyle_img = image_loader(style_img, imsize)content_img = image_loader(content_img, imsize)print("Style Image Size:{}".format(style_img.size()))print("Content Image Size:{}".format(content_img.size()))assert style_img.size() == content_img.size(), "we need to import style and content images of the same size"return style_img, content_img

定义内容损失

"content loss"
class ContentLoss(nn.Module):def __init__(self, target):super(ContentLoss, self).__init__()self.target = target.detach()def forward(self, input):self.loss = F.mse_loss(input, self.target)return input

定义风格损失

def gram_matrix(input):a, b, c, d = input.size()  # N, C,features = input.view(a * b, c * d)G = torch.mm(features, features.t())return G.div(a * b * c * d)

Gram Matrix 最后输出大小只和 filter 的个数有关（channels），上面的例子输出为 3x3

Gram Matrix 可以表示出特征出现的关系（特征 f1、f2、f3 之间的关系）。

我们可以通过计算 Gram Matrix 的差，来计算两张图片风格上的差距

class StyleLoss(nn.Module):def __init__(self, target_feature):# we "detach" the target content from the tree used to dynamically# compute the gradient: this is stated value, not a variable .# Otherwise the forward method of the criterion will throw an errorsuper(StyleLoss, self).__init__()self.target = gram_matrix(target_feature).detach()def forward(self, input):G = gram_matrix(input)self.loss = F.mse_loss(G, self.target)return input

写好前处理减均值，除方差

"based on VGG-16"
"put the normalization to the first layer"
class  Normalization(nn.Module):def __init__(self, mean, std):super(Normalization, self).__init__()# view the mean and std to make them [C,1,1] so that they can directly work with image Tensor of shape [B,C,H,W]self.mean = mean.view(-1, 1, 1)  # [3] -> [3, 1, 1]self.std = std.view(-1, 1, 1)def forward(self, img):return (img - self.mean) / self.std

定义网络，引入 loss

"modify to a style network"
def get_style_model_and_losses(cnn, normalization_mean, normalization_std,style_img, content_img,content_layers,style_layers):cnn = copy.deepcopy(cnn)# normalization modulenormalization = Normalization(normalization_mean, normalization_std).to(device)# just in order to have an iterable acess to or list of content / style# lossescontent_losses = []style_losses = []# assuming that cnn is a nn.Sequantial, so we make a new nn.Sequential to put# in modules that are supposed to be activated sequantiallymodel = nn.Sequential(normalization)i = 0  # increment every time we see a convfor layer in cnn.children():if isinstance(layer, nn.Conv2d):i += 1name = "conv_{}".format(i)elif isinstance(layer, nn.ReLU):name = "relu_{}".format(i)layer = nn.ReLU(inplace=False)elif isinstance(layer, nn.MaxPool2d):name = "pool_{}".format(i)elif isinstance(layer, nn.BatchNorm2d):name = "bn_{}".format(i)else:raise RuntimeError("Unrecognized layer: {}".format(layer.__class__.__name__))model.add_module(name, layer)if name in content_layers:# add content losstarget = model(content_img).detach()content_loss = ContentLoss(target)model.add_module("content_loss_{}".format(i), content_loss)content_losses.append(content_loss)if name in style_layers:# add style losstarget_feature = model(style_img).detach()style_loss = StyleLoss(target_feature)model.add_module("style_loss_{}".format(i), style_loss)style_losses.append(style_loss)# now we trim off the layers afater the last content and style lossesfor i in range(len(model)-1, -1, -1):if isinstance(model[i], ContentLoss) or isinstance(model[i], StyleLoss):breakmodel = model[:(i+1)]return model, style_losses, content_lossesdef get_input_optimizer(input_img):optimizer = optim.LBFGS([input_img.requires_grad_()])return optimizerdef run_style_transfer(cnn, normalization_mean, normalization_std, content_img, style_img, input_img, content_layers,style_layers, num_steps=50, style_weight=1000000, content_weight=1):print('Building the style transfer model..')model, style_losses, content_losses = get_style_model_and_losses(cnn, normalization_mean, normalization_std,style_img, content_img, content_layers,style_layers)optimizer = get_input_optimizer(input_img) # 网络不变，反向传播优化的是输入图片print('Optimizing..')run = [0]while run[0] <= num_steps:def closure():# correct the values of updated input imageinput_img.data.clamp_(0, 1)optimizer.zero_grad()model(input_img)  # 前向传播style_score = 0content_score = 0for sl in style_losses:style_score += sl.lossfor cl in content_losses:content_score += cl.lossstyle_score *= style_weightcontent_score *= content_weight# loss为style loss 和 content loss的和loss = style_score + content_scoreloss.backward()  # 反向传播# 打印loss的变化情况run[0] += 1if run[0] % 50 == 0:print("run {}:".format(run))print('Style Loss : {:4f} Content Loss: {:4f}'.format(style_score.item(), content_score.item()))print()return style_score + content_score# 进行参数优化optimizer.step(closure)# a last correction...# 数值范围的纠正, 使其范围在0-1之间input_img.data.clamp_(0, 1)return input_img

搭建完成，开始训练，仅优化更新 input image（get_input_optimizer），网络不更新

# 加载content image和style image
style_img,content_img = image_util(img_size=270, style_img="./style9.jpg", content_img="./content.jpg")  # [1, 3, 270, 270]
# input image使用content image
input_img = content_img.clone()
# 加载预训练好的模型
cnn = models.vgg19(pretrained=True).features.to(device).eval()
# 模型标准化的值
cnn_normalization_mean = torch.tensor([0.485, 0.456, 0.406]).to(device)
cnn_normalization_std = torch.tensor([0.229, 0.224, 0.225]).to(device)
# 定义要计算loss的层
content_layers_default = ['conv_4']
style_layers_default = ['conv_1', 'conv_2', 'conv_3', 'conv_4', 'conv_5']
# 模型进行计算
output = run_style_transfer(cnn, cnn_normalization_mean, cnn_normalization_std,content_img, style_img, input_img,content_layers=content_layers_default,style_layers=style_layers_default,num_steps=300, style_weight=100000, content_weight=1)image = output.cpu().clone()
image = image.squeeze(0)  # ([1, 3, 270, 270] -> [3, 270, 270])
unloader = transforms.ToPILImage()
image = unloader(image)
import cv2
image = cv2.cvtColor(np.asarray(image), cv2.COLOR_RGB2BGR)
cv2.imwrite("t9.jpg", image)
torch.cuda.empty_cache()"""VGG-19
Sequential((0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(1): ReLU(inplace=True)(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(3): ReLU(inplace=True)(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(6): ReLU(inplace=True)(7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(8): ReLU(inplace=True)(9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(11): ReLU(inplace=True)(12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(13): ReLU(inplace=True)(14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(15): ReLU(inplace=True)(16): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(17): ReLU(inplace=True)(18): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(19): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(20): ReLU(inplace=True)(21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(22): ReLU(inplace=True)(23): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(24): ReLU(inplace=True)(25): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(26): ReLU(inplace=True)(27): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(29): ReLU(inplace=True)(30): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(31): ReLU(inplace=True)(32): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(33): ReLU(inplace=True)(34): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(35): ReLU(inplace=True)(36): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
""""""modify name, add loss layer
Sequential((0): Normalization()(conv_1): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(style_loss_1): StyleLoss()(relu_1): ReLU()(conv_2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(style_loss_2): StyleLoss()(relu_2): ReLU()(pool_2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(conv_3): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(style_loss_3): StyleLoss()(relu_3): ReLU()(conv_4): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(content_loss_4): ContentLoss()(style_loss_4): StyleLoss()(relu_4): ReLU()(pool_4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(conv_5): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(style_loss_5): StyleLoss()(relu_5): ReLU()(conv_6): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(relu_6): ReLU()(conv_7): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(relu_7): ReLU()(conv_8): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(relu_8): ReLU()(pool_8): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(conv_9): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(relu_9): ReLU()(conv_10): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(relu_10): ReLU()(conv_11): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(relu_11): ReLU()(conv_12): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(relu_12): ReLU()(pool_12): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(conv_13): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(relu_13): ReLU()(conv_14): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(relu_14): ReLU()(conv_15): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(relu_15): ReLU()(conv_16): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(relu_16): ReLU()(pool_16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
""""""after trim
Sequential((0): Normalization()(conv_1): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(style_loss_1): StyleLoss()(relu_1): ReLU()(conv_2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(style_loss_2): StyleLoss()(relu_2): ReLU()(pool_2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(conv_3): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(style_loss_3): StyleLoss()(relu_3): ReLU()(conv_4): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(content_loss_4): ContentLoss()(style_loss_4): StyleLoss()(relu_4): ReLU()(pool_4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(conv_5): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(style_loss_5): StyleLoss()
)
"""

原图，花宝叽

不同风格

产生的结果

更直观的展示