-
Notifications
You must be signed in to change notification settings - Fork 5
/
transfer.py
216 lines (170 loc) · 8.89 KB
/
transfer.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
import os
import torch
import argparse
from PIL import Image
import torchvision.utils as vutils
from utils import Reformer
import time
import numpy as np
import torch.nn.functional as function
import torchvision.transforms as transforms
parser = argparse.ArgumentParser(description='Texture Reformer Pytorch')
# Specify inputs and outputs
parser.add_argument('-content', type=str, default='inputs/doodles/Seth.jpg', help="File path to the content image, valid for style transfer and invalid for texture transfer")
parser.add_argument('-style', type=str, default='inputs/doodles/Gogh.jpg', help="File path to the style/source image")
parser.add_argument('-content_sem', type=str, default='inputs/doodles/Seth_sem.png', help="File path to the semantic map of content/target image")
parser.add_argument('-style_sem', type=str, default='inputs/doodles/Gogh_sem.png', help="File path to the semantic map of style/source image")
parser.add_argument('-outf', type=str, default='outputs', help="Folder to save output images")
parser.add_argument('-content_size', type=int, default=0, help="Resize content/target, leave it to 0 if not resize")
parser.add_argument('-style_size', type=int, default=0, help="Resize style/source, leave it to 0 if not resize")
parser.add_argument('-style_transfer', action="store_true", help="Activate it if you want style transfer rather than texture transfer")
# Runtime controls
parser.add_argument('-coarse_alpha', type=float, default=1, help="Hyperparameter to blend transformed feature with content feature in coarse level (level 5)")
parser.add_argument('-fine_alpha', type=float, default=1, help="Hyperparameter to blend transformed feature with content feature in fine level (level 4)")
parser.add_argument('-semantic', type=str, default='concat', choices=['add', 'concat', 'concat_ds'], help="Choose different modes to embed semantic maps, 'add': our addition, 'concat': our concatenation, 'concat_ds': concat downsampled semantic maps")
parser.add_argument('-concat_weight', type=float, default=50, help="Hyperparameter to control the semantic guidance/awareness weight for '-semantic concat' mode and '-semantic concat_ds' mode, range 0-inf")
parser.add_argument('-add_weight', type=float, default=0.6, help="Hyperparameter to control the semantic guidance/awareness weight for '-semantic add' mode, range 0-1")
parser.add_argument('-coarse_psize', type=int, default=0, help="Patch size in coarse level (level 5), 0 means using global view")
parser.add_argument('-fine_psize', type=int, default=3, help="Patch size in fine level (level 4)")
parser.add_argument('-enhance', type=str, default='adain', choices=['adain', 'wct'], help="Choose different enhancement modes in level 3, level 2, and level 1. 'adain': first-order statistics enhancement, 'wct': second-order statistics enhancement.")
parser.add_argument('-enhance_alpha', type=float, default=1, help="Hyperparameter to control the enhancement degree in level 3, level 2, and level 1")
# Compress models
parser.add_argument('-compress', action="store_true", help="Use the compressed models for faster inference")
args = parser.parse_args()
args.cuda = torch.cuda.is_available()
if args.compress:
args.e5 = 'small_models/SE5.pth'
args.e4 = 'small_models/SE4.pth'
args.e3 = 'small_models/SE3.pth'
args.e2 = 'small_models/SE2.pth'
args.e1 = 'small_models/SE1.pth'
args.d5 = 'small_models/SD5.pth'
args.d4 = 'small_models/SD4.pth'
args.d3 = 'small_models/SD3.pth'
args.d2 = 'small_models/SD2.pth'
args.d1 = 'small_models/SD1.pth'
else:
args.e5 = 'models/E5.pth'
args.e4 = 'models/E4.pth'
args.e3 = 'models/E3.pth'
args.e2 = 'models/E2.pth'
args.e1 = 'models/E1.pth'
args.d5 = 'models/D5.pth'
args.d4 = 'models/D4.pth'
args.d3 = 'models/D3.pth'
args.d2 = 'models/D2.pth'
args.d1 = 'models/D1.pth'
print(args._get_kwargs())
# Set up data
# content/target data
# if the task is texture transfer, the content/target semantic map must be provided
if args.style_transfer:
content = Image.open(args.content).convert('RGB')
if args.content_size:
content = transforms.Resize(args.content_size)(content)
content = transforms.ToTensor()(content).unsqueeze(0)
if args.content_sem:
content_sem = Image.open(args.content_sem).convert('RGB')
if args.content_size:
content_sem = transforms.Resize(args.content_size)(content_sem)
content_sem = transforms.ToTensor()(content_sem).unsqueeze(0)
else:
content_sem = torch.ones(content.shape)
else:
content_sem = Image.open(args.content_sem).convert('RGB')
if args.content_size:
content_sem = transforms.Resize(args.content_size)(content_sem)
content_sem = transforms.ToTensor()(content_sem).unsqueeze(0)
# for texture transfer, use the content/target semantic map as the content
content = content_sem
# style/source data
style = Image.open(args.style).convert('RGB')
if args.style_size:
style = transforms.Resize(args.style_size)(style)
style = transforms.ToTensor()(style).unsqueeze(0)
if args.style_sem:
style_sem = Image.open(args.style_sem).convert('RGB')
if args.style_size:
style_sem = transforms.Resize(args.style_size)(style_sem)
style_sem = transforms.ToTensor()(style_sem).unsqueeze(0)
else:
style_sem = torch.ones(style.shape)
# Set up model and Texture Reformer
if args.cuda:
TR = Reformer(args).cuda()
else:
TR = Reformer(args)
# View-Specific Texture Reformation (VSTR) operation
@torch.no_grad()
def VSTR(encoder, decoder, content, style, content_sem, style_sem, patch_size, alpha, semantic):
# make the width and height of the temporary content/target image the same as the content/target semantic map
if content_sem.shape[2] != content.shape[2] or content_sem.shape[3] != content.shape[3]:
content = content.squeeze(0).cpu().clone()
content = transforms.ToPILImage()(content)
content = transforms.Resize([content_sem.shape[2],content_sem.shape[3]])(content)
if args.cuda:
content = transforms.ToTensor()(content).unsqueeze(0).cuda()
else:
content = transforms.ToTensor()(content).unsqueeze(0)
sF = encoder(style)
cF = encoder(content)
# add the features of the semantic maps
if semantic == "add":
sF_sem = encoder(style_sem)
cF_sem = encoder(content_sem)
csF = TR.VSTR_add(cF, sF, cF_sem, sF_sem, patch_size, alpha, args.add_weight)
# concatenate the features of the semantic maps
elif semantic == "concat":
sF_sem = encoder(style_sem)
cF_sem = encoder(content_sem)
csF = TR.VSTR_concat(cF, sF, cF_sem, sF_sem, patch_size, alpha, args.concat_weight)
# directly concatenate with the downsampled semantic maps
elif semantic == "concat_ds":
sF_sem = function.interpolate(style_sem, [sF.size(2), sF.size(3)], mode="nearest")
cF_sem = function.interpolate(content_sem, [cF.size(2), cF.size(3)], mode="nearest")
csF = TR.VSTR_concat(cF, sF, cF_sem, sF_sem, patch_size, alpha, args.concat_weight)
#csF = cF
Img = decoder(csF)
return Img
# Statistic-based Enhancement (SE) operation
@torch.no_grad()
def SE(encoder, decoder, content, style, enhance, enhance_alpha):
sF = encoder(style)
cF = encoder(content)
# match the first-order statistics, i.e., channel-wise mean and standard deviation
if enhance == "adain":
csF = TR.adain(cF, sF, enhance_alpha)
# match the second-order statistics, i.e., covariance
elif enhance == "wct":
sF = sF.data.cpu().squeeze(0) # note: svd runs on CPU
cF = cF.data.cpu().squeeze(0)
csF = TR.wct(cF, sF, enhance_alpha)
if args.cuda:
csF = csF.cuda()
#csF = cF
Img = decoder(csF)
return Img
# Run
if args.cuda:
content = content.cuda()
style = style.cuda()
content_sem = content_sem.cuda()
style_sem = style_sem.cuda()
# Start texture/style transfer
if args.cuda:
torch.cuda.synchronize()
start_time = time.time()
# Global view structure alignment stage
print("Processing level 5"); content = VSTR(TR.e5, TR.d5, content, style, content_sem, style_sem, args.coarse_psize, args.coarse_alpha, args.semantic)
# Local view texture refinement stage
print("Processing level 4"); content = VSTR(TR.e4, TR.d4, content, style, content_sem, style_sem, args.fine_psize, args.fine_alpha, args.semantic)
# Holistic effect enhancement stage
print("Processing level 3"); content = SE(TR.e3, TR.d3, content, style, args.enhance, args.enhance_alpha)
print("Processing level 2"); content = SE(TR.e2, TR.d2, content, style, args.enhance, args.enhance_alpha)
print("Processing level 1"); content = SE(TR.e1, TR.d1, content, style, args.enhance, args.enhance_alpha)
if args.cuda:
torch.cuda.synchronize()
end_time = time.time()
print('Elapsed time is: %.4f seconds' % (end_time - start_time))
out_path = os.path.join(args.outf, "output.jpg")
vutils.save_image(content.data.cpu(), out_path)