Style Transfer using Pytorch (Part 4)
Goal¶
This post aims to explain the concept of style transfer step-by-step. Part 4 is about executing the neural transfer.
Reference
- Ste-by-step Data Science - Style Transfer using Pytorch (Part 1)
- Ste-by-step Data Science - Style Transfer using Pytorch (Part 2)
- Ste-by-step Data Science - Style Transfer using Pytorch (Part 3)
- Original paper in arxiv - A Neural Algorithm of Artistic Style
- Colab - Neural style transfer using tesnorslow
- Towards Data Science - An Intuitive Understanding to Neural Style Transfer
Libraries¶
import pandas as pd
import copy
# Torch & Tensorflow
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
import tensorflow as tf
# Visualization
from torchviz import make_dot
from PIL import Image
import matplotlib.pyplot as plt
%matplotlib inline
import warnings
Configuration¶
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
Functions from Part 1 - image loader¶
# desired size of the output image
imsize = (512, 512) if torch.cuda.is_available() else (128, 128) # use small size if no gpu
loader = torchvision.transforms.Compose([
torchvision.transforms.Resize(imsize), # scale imported image
torchvision.transforms.ToTensor()]) # transform it into a torch tensor
def image_loader(image_name):
image = Image.open(image_name)
# fake batch dimension required to fit network's input dimensions
image = loader(image).unsqueeze(0)
return image.to(device, torch.float)
unloader = torchvision.transforms.ToPILImage()
def imshow_tensor(tensor, ax=None):
image = tensor.cpu().clone() # we clone the tensor to not do changes on it
image = image.squeeze(0) # remove the fake batch dimension
image = unloader(image)
if ax:
ax.imshow(image)
else:
plt.imshow(image)
Functions from Part 2 - loss functions¶
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
def gram_matrix(input):
# Get the size of tensor
# a: batch size
# b: number of feature maps
# c, d: the dimension of a feature map
a, b, c, d = input.size()
# Reshape the feature
features = input.view(a * b, c * d)
# Multiplication
G = torch.mm(features, features.t())
# Normalize
G_norm = G.div(a * b * c * d)
return G_norm
class StyleLoss(nn.Module):
def __init__(self, target_feature):
super(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
Functions from Part 3 - modeling¶
Normalization¶
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)
# create a module to normalize input image so we can easily put it in a
# nn.Sequential
class Normalization(nn.Module):
def __init__(self, mean, std):
super(Normalization, self).__init__()
# .view the mean and std to make them [C x 1 x 1] so that they can
# directly work with image Tensor of shape [B x C x H x W].
# B is batch size. C is number of channels. H is height and W is width.
self.mean = torch.tensor(mean).view(-1, 1, 1)
self.std = torch.tensor(std).view(-1, 1, 1)
def forward(self, img):
# normalize img
return (img - self.mean) / self.std
Create a sequential model for style transfer¶
# desired depth layers to compute style/content losses :
content_layers_default = ['conv_4']
style_layers_default = ['conv_1', 'conv_2', 'conv_3', 'conv_4', 'conv_5']
def get_style_model_and_losses(cnn, normalization_mean, normalization_std,
style_img, content_img,
content_layers=content_layers_default,
style_layers=style_layers_default):
cnn = copy.deepcopy(cnn)
# normalization module
normalization = Normalization(normalization_mean, normalization_std).to(device)
# just in order to have an iterable access to or list of content/syle
# losses
content_losses = []
style_losses = []
# assuming that cnn is a nn.Sequential, so we make a new nn.Sequential
# to put in modules that are supposed to be activated sequentially
model = nn.Sequential(normalization)
i = 0 # increment every time we see a conv
for n_child, layer in enumerate(cnn.children()):
# print()
# print(f"n_child: {n_child}")
if isinstance(layer, nn.Conv2d):
i += 1
name = 'conv_{}'.format(i)
elif isinstance(layer, nn.ReLU):
name = 'relu_{}'.format(i)
# The in-place version doesn't play very nicely with the ContentLoss
# and StyleLoss we insert below. So we replace with out-of-place
# ones here.
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)
# print(f'Name: {name}')
if name in content_layers:
# print(f'Add content loss {i}')
# add content loss:
target = 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:
# print(f'Add style loss {i}')
# add style loss:
target_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 after the last content and style losses
for i in range(len(model) - 1, -1, -1):
if isinstance(model[i], ContentLoss) or isinstance(model[i], StyleLoss):
break
model = model[:(i + 1)]
return model, style_losses, content_losses
Load images¶
d_path = {}
d_path['content'] = tf.keras.utils.get_file('turtle.jpg','https://storage.googleapis.com/download.tensorflow.org/example_images/Green_Sea_Turtle_grazing_seagrass.jpg')
d_path['style'] = tf.keras.utils.get_file('kandinsky.jpg','https://storage.googleapis.com/download.tensorflow.org/example_images/Vassily_Kandinsky%2C_1913_-_Composition_7.jpg')
style_img = image_loader(d_path['style'])[:, :, :, :170]
content_img = image_loader(d_path['content'])[:, :, :, :170]
input_img = content_img.clone()
assert style_img.size() == content_img.size(), \
"we need to import style and content images of the same size"
Modeling¶
# Obtain the model for style transfer
# with warnings.catch_warnings():
warnings.filterwarnings("ignore")
cnn = torchvision.models.vgg19(pretrained=True).features.to(device).eval()
model, style_losses, content_losses = get_style_model_and_losses(cnn, cnn_normalization_mean, cnn_normalization_std, style_img, content_img)
Executing a neural transfer¶
Gradient Decent¶
L-BFGS stands for Limited-memory Broyden–Fletcher–Goldfarb–Shanno according to wiki - Limited-memory_BFGS, which is one of the optimization algorithm using limited amount of memory.
def get_input_optimizer(input_img):
# this line to show that input is a parameter that requires a gradient
optimizer = torch.optim.LBFGS([input_img.requires_grad_()])
return optimizer
optimizer = get_input_optimizer(input_img)
Execution¶
The execution steps in the function get_style_model_and_losses
in NEURAL TRANSFER USING PYTORCH are as follows:
- Initialization
- Parameter
- Define
closure
function to re-evaluate the model to execute the followings:- masking images between 0 and 1 by
.clamp
method - reset gradient by
zero_grad
method - reset the error score for style and content
- compute the style and content loss in each inserted layer
- compute the sum of the losses for style and content
- multiply the weight for style and content to manipulate the style transfer balance by input argument
- execute error back propagation
- masking images between 0 and 1 by
- Execute the steps by gradient descent optimizer
# Parameters
num_steps = 10
style_weight=5000
content_weight=1
input_img = content_img[:, :, :, :170].clone()
d_images = {}
print('Building the style transfer model..')
model, style_losses, content_losses = get_style_model_and_losses(cnn,
cnn_normalization_mean, cnn_normalization_std, style_img, content_img)
optimizer = get_input_optimizer(input_img)
# Execution
run = [0]
while run[0] <= num_steps:
def closure():
# correct the values of updated input image
input_img.data.clamp_(0, 1)
optimizer.zero_grad()
model(input_img)
style_score = 0
content_score = 0
for sl in style_losses:
style_score += sl.loss
for cl in content_losses:
content_score += cl.loss
style_score *= style_weight
content_score *= content_weight
loss = style_score + content_score
loss.backward()
run[0] += 1
if run[0] % 2 == 0:
print("run {}:".format(run))
print('Style Loss : {:4f} Content Loss: {:4f}'.format(
style_score.item(), content_score.item()))
input_img.data.clamp_(0, 1)
d_images[run[0]] = input_img
print()
return style_score + content_score
optimizer.step(closure)
# a last correction...
input_img.data.clamp_(0, 1)
fig, axes = plt.subplots(1, 3, figsize=(16, 8))
d_img = {"Content": content_img,
"Style": style_img,
"Output": input_img}
for i, key in enumerate(d_img.keys()):
imshow_tensor(d_img[key], ax=axes[i])
axes[i].set_title(f"{key} Image")
axes[i].axis('off')
Visualize the process of style transfer¶
This is not yet obvious to see the processes of style transfer. It seems run 2 already finish most of the transfer processes. This needs to be investigated later.
fig, axes = plt.subplots(int(len(d_images)/2), 2, figsize=(16, 20))
for i, key in enumerate(d_images.keys()):
imshow_tensor(d_images[key], ax=axes[i//2][i%2])
axes[i//2][i%2].set_title("run {}:".format(key))
axes[i//2][i%2].axis('off')