-
Notifications
You must be signed in to change notification settings - Fork 2
/
stylize.py
346 lines (273 loc) · 11.2 KB
/
stylize.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
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# pylint: disable=missing-docstring
import argparse
import os
import sys
import timeit
import typing
from copy import deepcopy
import matplotlib
import matplotlib.pyplot as plt
import torch
torch.backends.cudnn.benchmark = True
from torchvision import transforms, utils
from model import *
from util import *
device = "cuda"
class StyleNet(typing.NamedTuple):
name: str
preserve_color: bool
ckpt_path: str
reference_image_path: str
class Styles:
models_dir = "models"
reference_image_dir = "style_images_aligned"
default = ["arcane_multi"]
pretrained = [
#"art",
#"arcane_multi",
#"supergirl",
"arcane_jinx",
"arcane_caitlyn",
"jojo_yasuho",
"jojo",
"disney",
]
@classmethod
def pretrained_net(cls, style: str, preserve_color: bool) -> StyleNet:
ckpt_path = os.path.join(cls.models_dir, f"{style}.pt")
if preserve_color:
ckpt_path_preserve_color = \
os.path.join(cls.models_dir, f"{style}_preserve_color.pt")
# load base version if preserve_color version not available
if os.path.isfile(ckpt_path_preserve_color):
ckpt_path = ckpt_path_preserve_color
if style == "arcane_multi":
reference_image_path = f"{cls.reference_image_dir}/arcane_jinx.png"
else:
reference_image_path = f"{cls.reference_image_dir}/{style}.png"
net = StyleNet(
name=style,
preserve_color=preserve_color,
ckpt_path=ckpt_path,
reference_image_path=reference_image_path,
)
cls.check_net(net)
return net
@classmethod
def check_net(cls, net: StyleNet) -> None:
if not net.name:
raise ValueError("style name not given")
if not os.path.isfile(net.ckpt_path):
raise FileNotFoundError(f"style ckpt not found: [{net.name}] {net.ckpt_path}")
if not os.path.isfile(net.reference_image_path):
raise FileNotFoundError(f"style reference image not found: [{net.name}] {net.reference_image_path}")
class StylizeOptions(typing.NamedTuple):
net: StyleNet
input: str
output_dir: str
do_show_all: bool
do_save_all: bool
class StylizeResult(typing.NamedTuple):
save_path: str
save_size: typing.Tuple[int, int]
class StylizeDisplay:
def __init__(self, options: StylizeOptions) -> None:
self._options = options
self._output_styles = []
self._output_images = []
def add(self, style: str, images: typing.List[torch.Tensor]) -> None:
self._output_styles.append(style)
self._output_images.append(images)
def run(self) -> None:
row = max(len(imgs) for imgs in self._output_images)
col = len(self._output_images)
grid_zeros = None
for img in self._output_images[0]:
if img is not None:
grid_zeros = torch.zeros_like(img)
assert grid_zeros is not None
grid_images = []
for i in range(row):
for imgs in self._output_images:
if i < len(imgs):
grid_images.append(
imgs[i] if imgs[i] is not None else grid_zeros)
else:
grid_images.append(grid_zeros)
grid_batch = torch.cat(grid_images, 0)
grid = utils.make_grid(grid_batch, nrow=col, normalize=True, value_range=(-1, 1))
save_fig_path = None
if self._options.do_save_all:
root, ext = os.path.splitext(os.path.basename(self._options.input))
save_root = os.path.join(self._options.output_dir,
f"{root}-all{'_preserve_color' if self._options.net.preserve_color else ''}")
save_path = f"{save_root}{ext}"
utils.save_image(grid, save_path)
print(f"Save all to: {save_path}, size={grid.shape}")
save_fig_path = f"{save_root}_fig{ext}"
if self._options.do_show_all:
matplotlib.use("Agg")
plt.rcParams["figure.dpi"] = 150
display_image(grid, title=os.path.basename(self._options.input))
plt.xlabel(" ".join(s if s else "☐" for s in self._output_styles))
plt.axis("on")
if save_fig_path:
plt.savefig(save_fig_path)
print(f" fig to: {save_fig_path}")
plt.show()
class Stylize:
def __init__(self) -> None:
self._latent_dim = 512
self._init_generator()
self._init_projection()
self._display = None
def _init_generator(self):
# Load original generator
original_generator = Generator(1024, self._latent_dim, 8, 2).to(device)
ckpt = torch.load(f"{Styles.models_dir}/stylegan2-ffhq-config-f.pt",
map_location=lambda storage, loc: storage)
original_generator.load_state_dict(ckpt["g_ema"], strict=False)
self._original_generator = original_generator
self._mean_latent = original_generator.mean_latent(10000)
# to be finetuned generator
self._generator = deepcopy(original_generator)
self._transform = transforms.Compose(
[
transforms.Resize((1024, 1024)),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
]
)
def _init_projection(self):
from e4e_projection import projection as e4e_projection
self._projection = e4e_projection
# from stylize_projection import StylizeProjection
# self._projection = StylizeProjection(device=device)
def __call__(self, *args, **kwds):
return self.run(*args, **kwds)
def run(self, options: StylizeOptions) -> typing.Optional[StylizeResult]:
# aligns and crops face
aligned_face = align_face(options.input)
restyle_path = strip_path_extension(options.input) + ".pt"
my_w = self._projection(aligned_face, restyle_path, device).unsqueeze(0)
ckpt = torch.load(options.net.ckpt_path, map_location=lambda storage, loc: storage)
generator = self._generator
generator.load_state_dict(ckpt["g"], strict=False)
with torch.no_grad():
generator.eval()
my_sample = generator(my_w, input_is_latent=True)
result = self._save_result(options, my_sample)
if options.do_show_all or options.do_save_all:
transform = self._transform
if self._display is None:
face = transform(aligned_face).unsqueeze(0).to(device)
self._display = StylizeDisplay(options)
self._display.add("", [None, face])
# style reference image
style_path = options.net.reference_image_path
style_image = transform(Image.open(style_path)).unsqueeze(0).to(device)
self._display.add(options.net.name, [style_image, my_sample])
os.remove(restyle_path)
return result
def done(self):
if self._display:
self._display.run()
def _save_result(self, options: StylizeOptions, result: torch.Tensor) \
-> typing.Optional[StylizeResult]:
if not options.output_dir:
return None
root, ext = os.path.splitext(os.path.basename(options.input))
save_path = os.path.join(options.output_dir,
f"{root}-{options.net.name}{'_preserve_color' if options.net.preserve_color else ''}{ext}")
def norm_ip(img, low, high):
img.clamp_(min=low, max=high)
img.sub_(low).div_(max(high - low, 1e-5))
return img
img = norm_ip(result.squeeze(0).clone(), -1, 1)
utils.save_image(img, save_path)
return StylizeResult(save_path=save_path, save_size=img.shape)
def _parse_args():
parser = argparse.ArgumentParser()
parser.add_argument("-d", "--device", default="cuda",
choices=["cuda", "cpu"],
help="the device name: %(default)s")
parser.add_argument("-i", "--input", default="test_input/iu.jpeg",
help="the input path of face: %(default)s")
parser.add_argument("-s", "--style", default=[],
action="extend", nargs="+", choices=["all"].extend(Styles.pretrained),
help=f"the output style: {Styles.default}")
parser.add_argument("-p", "--preserve_color", action="store_true",
help="use preserve color version: %(default)s")
parser.add_argument("-o", "--output_dir", default="output",
help="the output directory: %(default)s")
parser.add_argument("--show-all", action="store_true",
help="show all (face, style, result): %(default)s")
parser.add_argument("--save-all", action="store_true",
help="save all (face, style, result): %(default)s")
parser.add_argument("--test_style", type=str,
help="the test style name: %(default)s")
parser.add_argument("--test_preserve_color", action="store_true",
help="the test style whether preserve color or not: %(default)s")
parser.add_argument("--test_ckpt", type=str,
help="the test ckpt path: %(default)s")
parser.add_argument("--test_ref", type=str,
help="the test reference image path: %(default)s")
args = parser.parse_args()
global device
device = args.device
if not os.path.isfile(args.input):
sys.exit(f"input path not existed: {args.input}")
if not args.style:
args.style = Styles.default
elif "all" in args.style:
args.style = Styles.pretrained
args.style = sorted(list(set(args.style)))
if args.output_dir:
os.makedirs(args.output_dir, mode=0o774, exist_ok=True)
print("Args")
print(f" device: {args.device}")
print(f" input: {args.input}")
print(f" style: {args.style}")
print(f" preserve_color: {args.preserve_color}")
print(f" output_dir: {args.output_dir}")
print(f" show_all: {args.show_all}")
print(f" save_all: {args.save_all}")
print(f" test_style: {args.test_style}")
print(f" test_preserve_color: {args.test_preserve_color}")
print(f" test_ckpt: {args.test_ckpt}")
print(f" test_ref: {args.test_ref}")
return args
def _main():
args = _parse_args()
stylize = Stylize()
def stylize_run(net: StyleNet):
t_beg = timeit.default_timer()
save_path, save_size = stylize(StylizeOptions(
net=net,
input=args.input,
output_dir=args.output_dir,
do_show_all=args.show_all,
do_save_all=args.save_all,
))
t_end = timeit.default_timer()
print(f" [{net.name}] cost {t_end-t_beg:.2f} s")
print(f" > {save_path}, size={save_size}")
if args.test_ckpt is not None:
print(f"{args.input} stylizing (test) ...")
test_net = StyleNet(
name=args.test_style,
preserve_color=args.test_preserve_color,
ckpt_path=args.test_ckpt,
reference_image_path=args.test_ref,
)
Styles.check_net(test_net)
stylize_run(test_net)
else:
print(f"{args.input} stylizing ...")
for s in args.style:
stylize_run(Styles.pretrained_net(s, args.preserve_color))
stylize.done()
if __name__ == "__main__":
_main()