/
ica_single_img_demo.py
183 lines (155 loc) · 4.86 KB
/
ica_single_img_demo.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
import argparse
import torch
import random
from torchvision import utils
from sklearn.decomposition import FastICA
import numpy as np
import random
def independent_components_decomposition(W, n_components):
fast_ica = FastICA(n_components=n_components)
fast_ica.fit(W)
W_ = fast_ica.components_
norm = np.linalg.norm(W_, axis = 1).reshape(-1, n_components)
W_nomralize = W_ / norm.T
independent_components = torch.from_numpy(W_nomralize.T).float()
return independent_components
from model import Generator
if __name__ == "__main__":
torch.manual_seed(1)
np.random.seed(1)
random.seed(1)
torch.set_grad_enabled(False)
parser = argparse.ArgumentParser(description="Apply closed form factorization")
parser.add_argument(
"-d",
"--degree",
type=float,
default=4,
help="scalar factors for moving latent vectors along eigenvector",
)
parser.add_argument(
"-n", "--number_of_component", type=int, default=8, help="index of eigenvector"
)
parser.add_argument(
"--channel_multiplier",
type=int,
default=2,
help='channel multiplier factor. config-f = 2, else = 1',
)
parser.add_argument(
"--latent",
type=int,
default=512,
help="demension of the latent",
)
parser.add_argument(
"--n_mlp",
type=int,
default=8,
help="n_mlp",
)
parser.add_argument(
"--max_channel_size",
type=int,
default=512,
help="max channel size",
)
parser.add_argument("--ckpt", type=str, required=True, help="stylegan2 checkpoints")
parser.add_argument(
"--size", type=int, default=256, help="output image size of the generator"
)
parser.add_argument(
"--truncation", type=float, default=0.7, help="truncation factor"
)
parser.add_argument(
"--device", type=str, default="cuda", help="device to run the model"
)
parser.add_argument(
"--out_prefix",
type=str,
default="factor",
help="filename prefix to result samples",
)
parser.add_argument(
"factor",
type=str,
help="name of the closed form factorization result factor file",
)
parser.add_argument('--full_model', default=False, action='store_true')
parser.add_argument('--transpose', default=False, action='store_true')
args = parser.parse_args()
eigvec = torch.load(args.factor)["eigvec"].to(args.device)
ckpt = torch.load(args.ckpt)
if args.full_model:
state_dict = ckpt.state_dict()
g = ckpt.to(args.device)
else:
state_dict = ckpt["g_ema"]
g = Generator(
args.size, args.latent, args.n_mlp, channel_multiplier=args.channel_multiplier, max_channel_size=args.max_channel_size
).to(args.device)
g.load_state_dict(state_dict)
modulate = {
k: v
for k, v in state_dict.items()
if "modulation" in k and "to_rgbs" not in k and "weight" in k
}
weight_mat = []
for k, v in modulate.items():
weight_mat.append(v)
W = torch.cat(weight_mat, 0)
num_of_components = args.number_of_component
np.random.seed(0)
random.seed(0)
components = independent_components_decomposition(W, num_of_components).to(args.device)
trunc = g.mean_latent(128)
latent = trunc.tile((num_of_components,1))
direction = args.degree * components[:, 0:num_of_components].T
img, _ = g(
[latent],
truncation=args.truncation,
truncation_latent=trunc,
input_is_latent=True,
)
img1, _ = g(
[latent + direction],
truncation=args.truncation,
truncation_latent=trunc,
input_is_latent=True,
)
img2, _ = g(
[latent - direction],
truncation=args.truncation,
truncation_latent=trunc,
input_is_latent=True,
)
img3, _ = g(
[latent + 2 * direction],
truncation=args.truncation,
truncation_latent=trunc,
input_is_latent=True,
)
img4, _ = g(
[latent - 2 * direction],
truncation=args.truncation,
truncation_latent=trunc,
input_is_latent=True,
)
print(img.shape)
final_image = torch.cat([img3.unsqueeze(0), img1.unsqueeze(0), img.unsqueeze(0), img2.unsqueeze(0), img4.unsqueeze(0)], 0)
print(final_image.shape)
if args.transpose:
final_image = torch.transpose(final_image, 0, 1)
nrow = 5
else:
nrow = num_of_components
print(final_image.shape)
final_image = final_image.reshape(final_image.shape[0] * final_image.shape[1], final_image.shape[2], final_image.shape[3], final_image.shape[4])
print(final_image.shape)
grid = utils.save_image(
final_image,
f"demo.png",
normalize=True,
range=(-1, 1),
nrow=nrow,
)