-
Notifications
You must be signed in to change notification settings - Fork 1
/
stylizer.py
219 lines (178 loc) · 7.89 KB
/
stylizer.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
from dataclasses import dataclass
from typing import List, Tuple, Optional
import numpy as np
import torch
from torch import Tensor
# noinspection PyPep8Naming
from torch.nn import functional as F
from tqdm import tqdm
import image_utils
from vgg import Vgg19
@dataclass
class StylizerConfig:
# the weight of the content term in the total loss.
lambda_content: float = 1
# the weight of the style term in the total loss.
# empirically good range: 10 - 100_000
lambda_style: float = 100
# the weight of the generated image's total variation
# in the total loss. empirically good range: 0 - 1_000.
lambda_tv: float = 10
# the size of each step of the optimization process.
step_size: float = 0.1
# number of optimization iterations.
iterations: int = 500
# the weight of each convolutional block in the content loss.
# These five numbers refer to the following five activations of
# the VGG19 model: conv1_1, conv2_1, conv3_1, conv4_1, conv5_1.
content_block_weights: Tuple[float] = (0.0, 0.0, 0.0, 1.0, 0.0)
# the weight of each convolutional block in the style loss.
# These five numbers refer to the following five activations of
# the VGG19 model: conv1_1, conv2_1, conv3_1, conv4_1, conv5_1.
style_block_weights: Tuple[float] = (1/5, 1/5, 1/5, 1/5, 1/5)
# whether or not the optimization process should start with a
# random initial image (True), or the input content image (False).
random_initial_image: bool = False
# the maximal allowed input image dimension. input images of
# which max(H,W) is larger than this number will be downscaled appropriately.
# this also defines the dimension of the generated stylized image.
# Raising this value will allow creating larger stylized images, but
# will also require more time and memory.
max_input_dim: int = 512
# the interval (number of iterations) after which an intermediate
# result of the stylized image will be saved to the disk.
save_interval: int = 50
def update(self, **kwargs) -> 'StylizerConfig':
for key, value in kwargs.items():
if key in self.__dict__:
setattr(self, key, value)
else:
raise KeyError(f'Unknown configuration value: "{key}"')
return self
class Stylizer:
"""
A class that generates stylized images using the method presented in
"A Neural Algorithm of Artistic Style" by Gatys et. al (2015)
Paper: https://arxiv.org/abs/1508.06576.
"""
def __init__(self, use_gpu: bool = True):
gpu_available = torch.cuda.is_available()
self._device = 'cuda' if use_gpu and gpu_available else 'cpu'
self._vgg = Vgg19(use_avg_pooling=True).to(self._device)
self._opt = None
def stylize(
self,
content: np.ndarray,
style: np.ndarray,
config: Optional[StylizerConfig] = None,
) -> np.ndarray:
"""
Creates a stylized image in which the content is taken from the input
content image, and the style is taken from the input style image.
:param content: The content image: np.ndarray of shape (h, w, 3) in range [0, 1].
:param style: The style image: np.ndarray of shape (h, w, 3) in range [0, 1].
:param config: (optional) an instance of `StyleTransferConfig`.
:return: The generated stylized image: np.ndarray of shape (h, w, 3) in range [0, 1].
"""
config = config or StylizerConfig()
print(config)
content_t = self._preprocess(content, config.max_input_dim)
style_t = self._preprocess(style, config.max_input_dim)
if config.random_initial_image:
opt_t = torch.rand(
size=content_t.shape,
dtype=torch.float32,
device=self._device,
requires_grad=True,
)
else:
opt_t = content_t.clone().requires_grad_(True)
with torch.no_grad():
content_features = self._vgg(content_t)
style_features = self._vgg(style_t)
self._opt = torch.optim.Adam([opt_t], lr=config.step_size)
prog_bar = tqdm(range(1, config.iterations + 1))
for i in prog_bar:
loss = self._step(content_features, style_features, opt_t, config)
mean_grad = opt_t.grad.abs().mean().item()
prog_bar.set_description(f'Loss: {loss:.2f}, mean grad: {mean_grad:.7f}')
if i % config.save_interval == 0:
opt = self._postprocess(opt_t)
image_utils.save(opt, f'images/progress/opt_{i}.jpg')
opt = self._postprocess(opt_t)
return opt
def _preprocess(self, image: np.ndarray, max_dim: int) -> Tensor:
h, w = image.shape[:-1]
if max(h, w) > max_dim:
resize_factor = max_dim / max(h, w)
size = int(w * resize_factor), int(h * resize_factor)
image = image_utils.resize(image, size)
image_t = torch.tensor(image, device=self._device)
image_t = image_t.unsqueeze(0).permute(0, 3, 1, 2)
return image_t
@staticmethod
def _postprocess(image_t: Tensor) -> np.ndarray:
image_t = image_t.permute(0, 2, 3, 1).squeeze(0)
image = image_t.detach().cpu().numpy()
return image
def _step(
self,
content_features: List[Tensor],
style_features: List[Tensor],
opt_t: Tensor,
config: StylizerConfig,
) -> float:
opt_features = self._vgg(opt_t)
content_loss = self._content_loss(opt_features, content_features, config.content_block_weights)
style_loss = self._style_loss(opt_features, style_features, config.style_block_weights)
tv_loss = self._tv_loss(opt_t)
loss = content_loss * config.lambda_content + style_loss * config.lambda_style + config.lambda_tv * tv_loss
self._opt.zero_grad()
loss.backward()
self._opt.step()
loss_f = loss.item()
with torch.no_grad():
opt_t.clamp_(0.0, 1.0)
return loss_f
@staticmethod
def _content_loss(features_input: List[Tensor],
features_target: List[Tensor],
weights: Tuple[float]) -> Tensor:
assert len(features_input) == len(features_target) == len(weights)
device = features_input[0].device
total = torch.zeros(1, dtype=torch.float32, device=device)
num_features = len(features_input)
for i in range(num_features):
if weights[i] > 0:
block_loss = F.mse_loss(features_input[i], features_target[i])
block_loss = block_loss
total = total + block_loss * weights[i]
return total
@staticmethod
def _style_loss(features_input: List[Tensor],
features_target: List[Tensor],
weights: Tuple[float]) -> Tensor:
assert len(features_input) == len(features_target) == len(weights)
device = features_input[0].device
total = torch.zeros(1, dtype=torch.float32, device=device)
num_features = len(features_input)
for i in range(num_features):
if weights[i] > 0:
gram_input = Stylizer._gram_matrix(features_input[i])
gram_target = Stylizer._gram_matrix(features_target[i])
block_loss = F.mse_loss(gram_input, gram_target)
total = total + block_loss * weights[i]
return total
@staticmethod
def _tv_loss(image: Tensor) -> Tensor:
tv_loss = (image[:, :, :, :-1] - image[:, :, :, 1:]).abs().mean() + \
(image[:, :, :-1, :] - image[:, :, 1:, :]).abs().mean()
return tv_loss
@staticmethod
def _gram_matrix(features: Tensor) -> Tensor:
n, c, h, w = features.shape
x = features.view(n, c, h * w)
y = features.view(n, c, h * w).permute(0, 2, 1)
gram = torch.bmm(x, y)
gram = gram / (h * w)
return gram