forked from HideUnderBush/UI2I_via_StyleGAN2
-
Notifications
You must be signed in to change notification settings - Fork 1
/
gen_ref.py
160 lines (131 loc) · 4.9 KB
/
gen_ref.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
import os
import argparse
import math
import pdb
import torch
from torch import optim
from torch.nn import functional as F
from torchvision import transforms
from PIL import Image
from tqdm import tqdm
from utils import AddPepperNoise
import numpy as np
from scipy.linalg import solve
import lpips
from model import Generator
import random
import sys
parser = argparse.ArgumentParser()
parser.add_argument("--fact", type=str, required=True)
parser.add_argument("--fact_ref", type=str, required=True)
parser.add_argument("--model1", type=str, required=True)
parser.add_argument("--model2", type=str, required=True)
parser.add_argument("--size1", type=int, default=1024)
parser.add_argument("--size2", type=int, default=1024)
parser.add_argument("-o", "--output", type=str, default="output")
parser.add_argument("--device", type=str, default="cuda")
parser.add_argument('--truncation_mean', type=int, default=4096)
parser.add_argument("-r", "--randnum", type=int, default=10)
parser.add_argument("--fact_base1", type=str, required=True)
parser.add_argument("--fact_base2", type=str, required=True)
args = parser.parse_args()
device = args.device
## load eigvec
eigvec1 = torch.load(args.fact_base1)["eigvec"].to(args.device)
eigvec1.requires_grad = False
eigvec2 = torch.load(args.fact_base2)["eigvec"].to(args.device)
eigvec2.requires_grad = False
fact_path = args.fact
item = torch.load(fact_path)
vec = next(iter(item.values()))['weight'].to(device)
fact_path_ref = args.fact_ref
item_ref = torch.load(fact_path_ref)
vec_ref = next(iter(item_ref.values()))['weight'].to(device)
input_latent = torch.mm(vec, eigvec1)
style_latent = torch.mm(vec_ref, eigvec2)
def noise_normalize_(noises):
for noise in noises:
mean = noise.mean()
std = noise.std()
noise.data.add_(-mean).div_(std)
def make_image(tensor):
return (
tensor.detach()
.clamp_(min=-1, max=1)
.add(1)
.div_(2)
.mul(255)
.type(torch.uint8)
.permute(0, 2, 3, 1)
.to("cpu")
.numpy()
)
device = args.device
# generate images
## load model
g_ema1 = Generator(args.size1, 512, 8)
g_ema1.load_state_dict(torch.load(args.model1, map_location='cuda:0')["g_ema"], strict=False)
g_ema1.eval()
g_ema1 = g_ema1.to(device)
g_ema2 = Generator(args.size2, 512, 8)
g_ema2.load_state_dict(torch.load(args.model2, map_location='cuda:0')["g_ema"], strict=False)
g_ema2.eval()
g_ema2 = g_ema2.to(device)
## prepare input vector
sample_z_style = torch.randn(1, 512, device=args.device)
sample_z1 = torch.randn(1, 512, device=args.device)
num = 5
sample_z = []
sample_ref_z = []
for i in range(num):
sample_zi = torch.randn(1, 512, device=args.device)
sample_ref_zi = torch.randn(1, 512, device=args.device)
sample_z.append(sample_zi)
sample_ref_z.append(sample_ref_zi)
## noise
noises_single = g_ema2.make_noise()
noises = []
for noise in noises_single:
noises.append(noise.repeat(1, 1, 1, 1).normal_())
noise_normalize_(noises)
## gen images
with torch.no_grad():
mean_latent2 = g_ema2.mean_latent(args.truncation_mean)
# generate ref and identity
swap_res = []
swap_total = []
swap_ref_res = []
swap_ref_total = []
for i in range(num):
print("processing base figure [{}/{}]".format(i, num))
for j in range(1, 6, 2):
img1, swap_res_i = g_ema1([sample_z[i]], truncation=0.5, truncation_latent=mean_latent2, save_for_swap=True, swap_layer=j)
swap_res.append(swap_res_i)
img2, swap_ref_res_i = g_ema2([sample_ref_z[i]], truncation=0.5, truncation_latent=mean_latent2, save_for_swap=True, swap_layer=j)
swap_ref_res.append(swap_ref_res_i)
# identity
swap_total.append(swap_res)
swap_res = []
# ref
swap_ref_total.append(swap_ref_res)
swap_ref_res = []
for i in range(num):
print("processing I2I [{}/{}]".format(i, num))
# swap=5
img3, _ = g_ema2([sample_z[i]], truncation=0.5, truncation_latent=mean_latent2, swap=True, swap_layer=5, swap_tensor=swap_total[i][2], multi_style=True, multi_style_layers=3, multi_style_latent=[sample_ref_z[i]])
img3_name = args.output + str(i) + "_ls5_" + ".png"
img3 = make_image(img3)
out3 = Image.fromarray(img3[0])
out3.save(img3_name)
# swap=3
img4, _ = g_ema2([sample_z[i]], truncation=0.5, truncation_latent=mean_latent2, swap=True, swap_layer=3, swap_tensor=swap_total[i][1], multi_style=True, multi_style_layers=3, multi_style_latent=[sample_ref_z[i]])
img4_name = args.output + str(i) + "_ls3_" + ".png"
img4 = make_image(img4)
out4 = Image.fromarray(img4[0])
out4.save(img4_name)
# swap=1
img5, _ = g_ema2([sample_z[i]], truncation=0.5, truncation_latent=mean_latent2, swap=True, swap_layer=1, swap_tensor=swap_total[i][0], multi_style=True, multi_style_layers=3, multi_style_latent=[sample_ref_z[i]])
img5_name = args.output + str(i) + "_ls1_" + ".png"
img5 = make_image(img5)
out5 = Image.fromarray(img5[0])
out5.save(img5_name)