-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathPULSE.py
184 lines (153 loc) · 7.19 KB
/
PULSE.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
from SphericalOptimizer import SphericalOptimizer
from pathlib import Path
import numpy as np
import time
import torch
from loss import LossBuilder
from functools import partial
from stylegan2 import Generator
from drive import open_url
class PULSE(torch.nn.Module):
def __init__(self, cache_dir,verbose=True):
super(PULSE, self).__init__()
self.verbose = verbose
self.generator = Generator(z_dim=512, w_dim=512, c_dim=0, resolution=1024, img_channels=3,
fused_modconv_default='inference_only', embed_res=64).cuda()#ADDED
cache_dir = Path(cache_dir)
cache_dir.mkdir(parents=True, exist_ok = True)
if self.verbose: print("Loading StyleGAN2 Network")
gen_pretrained = torch.load("generator.pth", map_location="cpu")
self.generator.load_state_dict(gen_pretrained, strict=True)
print("StyleGAN2 succesfully loaded")
###############################################################
'''
with open_url("https://drive.google.com/uc?id=1GvqxkJ9BdRj_RNEvj2sY2B8V_qyaY6E7", cache_dir=cache_dir, verbose=verbose) as f:
gen_pretrained = torch.load(f)
self.generator.load_state_dict(gen_pretrained,strict=True)
for param in self.generator.parameters():
param.requires_grad = False
print("StyleGAN2 succesfully loaded")
'''
self.lrelu = torch.nn.LeakyReLU(negative_slope=0.2)
###############################################################
if Path("gaussian_fit.pt").exists():
self.gaussian_fit = torch.load("gaussian_fit.pt")
else:
if self.verbose: print("\tLoading Mapping Network")
if self.verbose: print("\tRunning Mapping Network")
with torch.no_grad():
torch.manual_seed(0)
latent = torch.randn((100000,512),dtype=torch.float32, device="cuda")
latent_out = torch.nn.LeakyReLU(5)(self.generator(latent,0, mode='mapping'))
self.gaussian_fit = {"mean": latent_out.mean(0), "std": latent_out.std(0)}
torch.save(self.gaussian_fit,"gaussian_fit.pt")
if self.verbose: print("\tSaved \"gaussian_fit.pt\"")
def forward(self, ref_im,
seed,
loss_str,
eps,
noise_type,
num_trainable_noise_layers,
tile_latent,
bad_noise_layers,
opt_name,
learning_rate,
steps,
lr_schedule,
save_intermediate,
**kwargs):
if seed:
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
batch_size = ref_im.shape[0]
# Generate latent tensor
if(tile_latent):
latent = torch.randn(
(batch_size, 1, 512), dtype=torch.float, requires_grad=True, device='cuda')
else:
latent = torch.randn(
(batch_size, 18, 512), dtype=torch.float, requires_grad=True, device='cuda')
# Generate list of noise tensors
noise = [] # stores all of the noise tensors
noise_vars = [] # stores the noise tensors that we want to optimize on
for i in range(18):
# dimension of the ith noise tensor
res = (batch_size, 1, 2**(i//2+2), 2**(i//2+2))
if(noise_type == 'zero' or i in [int(layer) for layer in bad_noise_layers.split('.')]):
new_noise = torch.zeros(res, dtype=torch.float, device='cuda')
new_noise.requires_grad = False
elif(noise_type == 'fixed'):
new_noise = torch.randn(res, dtype=torch.float, device='cuda')
new_noise.requires_grad = False
elif (noise_type == 'trainable'):
new_noise = torch.randn(res, dtype=torch.float, device='cuda')
if (i < num_trainable_noise_layers):
new_noise.requires_grad = True
noise_vars.append(new_noise)
else:
new_noise.requires_grad = False
else:
raise Exception("unknown noise type")
noise.append(new_noise)
var_list = [latent]+noise_vars
opt_dict = {
'sgd': torch.optim.SGD,
'adam': torch.optim.Adam,
'sgdm': partial(torch.optim.SGD, momentum=0.9),
'adamax': torch.optim.Adamax
}
opt_func = opt_dict[opt_name]
opt = SphericalOptimizer(opt_func, var_list, lr=learning_rate)
schedule_dict = {
'fixed': lambda x: 1,
'linear1cycle': lambda x: (9*(1-np.abs(x/steps-1/2)*2)+1)/10,
'linear1cycledrop': lambda x: (9*(1-np.abs(x/(0.9*steps)-1/2)*2)+1)/10 if x < 0.9*steps else 1/10 + (x-0.9*steps)/(0.1*steps)*(1/1000-1/10),
}
schedule_func = schedule_dict[lr_schedule]
scheduler = torch.optim.lr_scheduler.LambdaLR(opt.opt, schedule_func)
loss_builder = LossBuilder(ref_im, loss_str, eps).cuda()
min_loss = np.inf
min_l2 = np.inf
best_summary = ""
start_t = time.time()
gen_im = None
if self.verbose: print("Optimizing")
for j in range(steps):
opt.opt.zero_grad()
# Duplicate latent in case tile_latent = True
if (tile_latent):
latent_in = latent.expand(-1, 18, -1)
else:
latent_in = latent
# Apply learned linear mapping to match latent distribution to that of the mapping network
latent_in = self.lrelu(latent_in*self.gaussian_fit["std"] + self.gaussian_fit["mean"])
# Normalize image to [0,1] instead of [-1,1]
gen_im = (self.generator(latent_in, None)+1)/2
# Calculate Losses
loss, loss_dict = loss_builder(latent_in, gen_im)
loss_dict['TOTAL'] = loss
# Save best summary for log
if(loss < min_loss):
min_loss = loss
best_summary = f'BEST ({j+1}) | '+' | '.join(
[f'{x}: {y:.4f}' for x, y in loss_dict.items()])
best_im = gen_im.clone()
loss_l2 = loss_dict['L2']
if(loss_l2 < min_l2):
min_l2 = loss_l2
# Save intermediate HR and LR images
if(save_intermediate):
yield (best_im.cpu().detach().clamp(0, 1),loss_builder.D(best_im).cpu().detach().clamp(0, 1))
loss.backward()
opt.step()
scheduler.step()
print(gen_im.shape)
print(latent_in.shape)
total_t = time.time()-start_t
current_info = f' | time: {total_t:.1f} | it/s: {(j+1)/total_t:.2f} | batchsize: {batch_size}'
if self.verbose: print(best_summary+current_info)
if(min_l2 <= eps):
yield (gen_im.clone().cpu().detach().clamp(0, 1),loss_builder.D(best_im).cpu().detach().clamp(0, 1))
else:
print("Could not find a face that downscales correctly within epsilon")