forked from SJTU-Lucy/CS238-VirtualReality
-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
182 lines (159 loc) · 7.79 KB
/
main.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
import os
import sys
import torch
import torch as T
import torch.optim as optim
from torchvision.transforms import functional as F
from neural_monitor import monitor as mon
from neural_monitor import logger
import argparse
from param_stroke import BrushStrokeRenderer
import feature
import utils
import losses
parser = argparse.ArgumentParser()
parser.add_argument('content_img_file', type=str, help='Content image file')
parser.add_argument('style_img_file', type=str, help='Style image file')
parser.add_argument('--img_size', '-s', type=int, default=512,
help='The smaller dimension of content image is resized into this size. Default: 512.')
parser.add_argument('--canvas_color', default='gray', type=str,
help='Canvas background color (`gray` (default), `white`, `black` or `noise`).')
parser.add_argument('--num_strokes', default=10000, type=int,
help='Number of strokes to draw. Default: 5000.')
parser.add_argument('--samples_per_curve', default=20, type=int,
help='Number of points to sample per parametrized curve. Default: 10.')
parser.add_argument('--brushes_per_pixel', default=20, type=int,
help='Number of brush strokes to be drawn per pixel. Default: 20.')
parser.add_argument('--output_path', '-o', type=str, default='results',
help='Storage for results. Default: `results`.')
parser.add_argument('--device', '-d', type=str, default='cpu',
help='Device to perform stylization. Default: `cuda`.')
try:
args = parser.parse_args()
except:
parser.print_help()
sys.exit(0)
# inputs
style_img_file = args.style_img_file
content_img_file = args.content_img_file
# setup logging
model_name = 'nst-stroke'
root = args.output_path
vgg_weight_file = 'vgg_weights/vgg19_weights_normalized.h5'
# vgg_weight_file = 'vgg_weights/vgg19_weights_tf_dim_ordering_tf_kernels_notop.h5'
print_freq = 10
mon.initialize(model_name=model_name, root=root, print_freq=print_freq)
mon.backup(('main.py', 'param_stroke.py', 'utils.py', 'losses.py', 'vgg.py'))
# device
device = torch.device(args.device)
# desired size of the output image
imsize = args.img_size
content_img = utils.image_loader(content_img_file, imsize, device)
style_img = utils.image_loader(style_img_file, imsize, device)
output_name = f'{os.path.basename(content_img_file).split(".")[0]}-{os.path.basename(style_img_file).split(".")[0]}'
# desired depth layers to compute style/content losses :
bs_content_layers = ['conv4_1', 'conv5_1']
bs_style_layers = ['conv1_1', 'conv2_1', 'conv3_1', 'conv4_1', 'conv5_1']
px_content_layers = ['conv1_1', 'conv2_1', 'conv3_1', 'conv4_1', 'conv5_1']
px_style_layers = ['conv1_1', 'conv2_1', 'conv3_1', 'conv4_1', 'conv5_1']
# brush strokes parameters
canvas_color = args.canvas_color
num_strokes = args.num_strokes
samples_per_curve = args.samples_per_curve
brushes_per_pixel = args.brushes_per_pixel
_, _, H, W = content_img.shape
canvas_height = H
canvas_width = W
length_scale = 1.1
width_scale = 0.1
# 笔画的风格化
def run_stroke_style_transfer(num_steps=100, style_weight=3., content_weight=1., tv_weight=0.008, curv_weight=4):
# 用于计算content loss和style loss
vgg_loss = losses.StyleTransferLosses(vgg_weight_file, content_img, style_img,
bs_content_layers, bs_style_layers, scale_by_y=True)
# 用于计算feature loss
feature_loss = feature.feature_loss(content_img)
vgg_loss.to(device).eval()
# brush stroke init
bs_renderer = BrushStrokeRenderer(canvas_height, canvas_width, num_strokes, samples_per_curve, brushes_per_pixel,
canvas_color, length_scale, width_scale,
content_img=content_img[0].permute(1, 2, 0).cpu().numpy())
bs_renderer.to(device)
optimizer = optim.Adam([bs_renderer.location, bs_renderer.curve_s, bs_renderer.curve_e,
bs_renderer.curve_c, bs_renderer.width], lr=1e-1)
optimizer_color = optim.Adam([bs_renderer.color], lr=1e-2)
logger.info('Optimizing brushstroke-styled canvas..')
for _ in mon.iter_batch(range(num_steps)):
optimizer.zero_grad()
optimizer_color.zero_grad()
# 用当前的渲染器渲染图像
input_img = bs_renderer()
input_img = input_img[None].permute(0, 3, 1, 2).contiguous()
# input of [1, 3, height, width]，评价人脸特征的丢失程度
feature_weight = 5
feature_score = feature_weight * feature_loss.compute(input_img)
# style_core: 风格化评价 content_score: 还原度评价 tv_score: 笔画分布评价 curv_score: 笔画弯曲程度评价
content_score, style_score = vgg_loss(input_img)
style_score *= style_weight
content_score *= content_weight
tv_score = tv_weight * losses.total_variation_loss(bs_renderer.location, bs_renderer.curve_s,
bs_renderer.curve_e, K=10)
curv_score = curv_weight * losses.curvature_loss(bs_renderer.curve_s, bs_renderer.curve_e, bs_renderer.curve_c)
loss = style_score + content_score + tv_score + curv_score + feature_score
loss.backward(inputs=[bs_renderer.location, bs_renderer.curve_s, bs_renderer.curve_e,
bs_renderer.curve_c, bs_renderer.width], retain_graph=True)
optimizer.step()
style_score.backward(inputs=[bs_renderer.color])
optimizer_color.step()
# plot some stuffs
mon.plot('stroke feature loss', feature_score)
mon.plot('stroke style loss', style_score.item())
mon.plot('stroke content loss', content_score.item())
mon.plot('stroke tv loss', tv_score.item())
mon.plot('stroke curvature loss', curv_score.item())
if mon.iter % mon.print_freq == 0:
mon.imwrite('stroke stylized', input_img)
with T.no_grad():
return bs_renderer()
# 像素级优化
def run_style_transfer(input_img: T.Tensor, num_steps=100, style_weight=10000., content_weight=1., tv_weight=0):
# input size of [1, 3, 1364, 1024]
input_img = input_img.detach()[None].permute(0, 3, 1, 2).contiguous()
input_img = F.resize(input_img, imsize)
vgg_loss = losses.StyleTransferLosses(vgg_weight_file, input_img, style_img,
px_content_layers, px_style_layers)
vgg_loss.to(device).eval()
input_img = T.nn.Parameter(input_img, requires_grad=True)
feature_loss = feature.feature_loss(input_img)
optimizer = optim.Adam([input_img], lr=1e-3)
logger.info('Optimizing pixel-wise canvas..')
for _ in mon.iter_batch(range(num_steps)):
optimizer.zero_grad()
input = T.clamp(input_img, 0., 1.)
content_score, style_score = vgg_loss(input)
style_score *= style_weight
content_score *= content_weight
feature_weight = 100
feature_score = feature_weight * feature_loss.compute(input_img)
tv_score = 0. if not tv_weight else tv_weight * losses.tv_loss(input_img)
loss = style_score + content_score + tv_score + feature_score
loss.backward(inputs=[input_img])
optimizer.step()
# plot some stuffs
mon.plot('feature loss', feature_score)
mon.plot('pixel style loss', style_score)
mon.plot('pixel content loss', content_score)
if tv_weight:
mon.plot('pixel tv loss', tv_score)
if mon.iter % mon.print_freq == 0:
mon.imwrite('pixel stylized', input)
return T.clamp(input_img, 0., 1.)
if __name__ == '__main__':
# optimize brush style transfer model
canvas = run_stroke_style_transfer()
# optimize the canvas pixel-wise
mon.iter = 0
mon.print_freq = 10
output = run_style_transfer(canvas)
mon.imwrite(output_name, output)
logger.info('Finished!')