-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathcoach_test.py
201 lines (167 loc) · 8.91 KB
/
coach_test.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
import os
import torch
import torchvision
import torch.nn.functional as F
import torchvision.transforms as transforms
#sys path dependency
import sys
sys.path.append("/hd3/lidongze/animation/RiDDLE")
sys.path.append("/hd3/lidongze/animation/RiDDLE/mapper")
from torch.utils.data import TensorDataset
from models.encoders.model_irse import Backbone
from models.stylegan2.model import Generator
from mapper import latent_id_mappers
from models.psp import get_keys,pSp
from models import seg_model_2
import logging
from argparse import Namespace
from mapper.datasets.simple_dataset import SimpleDataset
#just for testing
from mapper.options.test_options_pwd import TestOptions
from criteria.id_loss import IDLossExtractor
from os.path import join
def reorder(ts,batch_size,range_num):
index=torch.arange(ts.shape[0]).view(range_num,batch_size)
index=index.T.reshape(batch_size,range_num).flatten()
return ts[index]
#test engine
class CoachTest:
def __init__(self,opts):
#model related
self.opts=opts
self.device=self.opts.device
self.e4e=self.load_e4e()
self.decoder=Generator(self.opts.stylegan_size,512,8).to(self.device)
self.decoder.load_state_dict(torch.load(self.opts.stylegan_weights,map_location='cpu')['g_ema'],strict=False)
if self.opts.mapper_type=='simple':
#self.mapper= latent_id_mappers.SimpleMapper().to(self.device)
self.mapper= latent_id_mappers.SimpleMapper_morefc(1024).to(self.device)
elif self.opts.mapper_type=='transformer':
self.mapper = latent_id_mappers.TransformerMapperSplit(split_list=self.opts.transformer_split_list,normalize_type=self.opts.transformer_normalize_type,\
add_linear=self.opts.transformer_add_linear,add_pos_embedding=self.opts.transformer_add_pos_embedding).to(self.device)
self.mapper.load_state_dict(torch.load(opts.mapper_weight,map_location='cpu')['mapper_state_dict'])
#load segmentation model
self.segmentation_model = seg_model_2.BiSeNet(19).eval().cuda().requires_grad_(False)
self.segmentation_model.load_state_dict(torch.load("/hd3/lidongze/animation/STIT/pretrained_models/79999_iter.pth"))
#load id extractor
self.arcface = Backbone(input_size=112, num_layers=50, drop_ratio=0.6, mode='ir_se').to(self.device)
self.arcface.load_state_dict(torch.load(opts.ir_se50_weights,map_location='cpu'))
self.id_loss=IDLossExtractor(self.opts,self.arcface).to(self.device).eval()
#tensor dataset related
self.latent=torch.load(opts.latent_path,map_location='cpu')
self.latent_dataset = TensorDataset(self.latent)
self.latent_loader = torch.utils.data.DataLoader(
self.latent_dataset,
batch_size=self.opts.batch_size,
shuffle=False,
num_workers=4,
pin_memory=True,
drop_last=False
)
#image dataset related
trans=transforms.Compose([
transforms.Resize((self.opts.image_size, self.opts.image_size)),
transforms.ToTensor(),
transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])
self.image_dataset = SimpleDataset(self.opts.image_path,transform=trans)
self.image_loader = torch.utils.data.DataLoader(
self.image_dataset,
batch_size=self.opts.batch_size,
shuffle=False,
num_workers=4,
pin_memory=True,
drop_last=False
)
#working dir
self.exp_dir=self.opts.exp_dir
os.makedirs(self.exp_dir,exist_ok=True)
#pwd list
self.w_pwd_list=[self.sample_w(1) for i in range(self.opts.pwd_num)]
self.w_wrong_pwd_list=[self.sample_w(1) for i in range(self.opts.pwd_num)]
def load_e4e(self):
print('loading e4e')
print(self.opts.e4e_model_weights)
ckpt = torch.load(self.opts.e4e_model_weights, map_location='cpu')
opts = ckpt['opts'] #may need to see the opt
opts['checkpoint_path'] = self.opts.e4e_model_weights
opts['stylegan_weights'] = self.opts.stylegan_weights
print(opts) #see the opts
opts= Namespace(**opts)
net = pSp(opts)
print(f'loaded e4e from {self.opts.e4e_model_weights}')
print(f'e4e stylegan is from { self.opts.stylegan_weights}')
return net.to(self.device)
def sample_w(self,b):
z = torch.randn(b, 512, device=self.device)
w = self.decoder.style(z).unsqueeze(1).repeat(1,self.opts.latents_num,1)
return w
def single_pwd_loop(self):
output_frame_path=join(self.exp_dir,'frames')
output_video_path=join(self.exp_dir,f'{self.exp_dir.split("/")[-1]}.mp4')
os.makedirs(output_frame_path,exist_ok=True)
w_pwd=self.w_pwd_list[0]
print(w_pwd.shape)
w_wrong_pwd=self.w_wrong_pwd_list[0]
with torch.no_grad():
for j,(latent_batch,image_batch) in enumerate(zip(self.latent_loader,self.image_loader)):
w_ori=latent_batch[0].cuda()
print(w_ori.shape)
x_ori_real=image_batch[0] #real image
x_ori, _ = self.decoder([w_ori], input_is_latent=True, randomize_noise=False, truncation=1)
b,c,h,w = x_ori.shape
img_name=[f"{(b_idx+j*b):0>4d}.jpg" for b_idx in range(b)]
print(img_name)
save_batch=torch.zeros((b,c,2*h,3*w))
save_batch[:,:,0:h,0:w]=x_ori_real
save_batch[:,:,0:h,w:2*w]=x_ori.cpu()
w_pwd=self.w_pwd_list[0].repeat(b,1,1)
w_wrong_pwd=self.w_wrong_pwd_list[0].repeat(b,1,1)
w_enc=self.mapper(torch.cat([w_ori,w_pwd],dim=-1))
x_enc, _ = self.decoder([w_enc], input_is_latent=True, randomize_noise=False, truncation=1)
w_cat_dec=torch.cat([w_enc,w_pwd],dim=-1)
w_dec=self.mapper(w_cat_dec)
x_recover, _ = self.decoder([w_dec], input_is_latent=True, randomize_noise=False, truncation=1)
w_cat_dec_wrong=torch.cat([w_enc,w_wrong_pwd],dim=-1)
w_dec_wrong=self.mapper(w_cat_dec_wrong)
x_wrong, _ = self.decoder([w_dec_wrong], input_is_latent=True, randomize_noise=False, truncation=1)
save_batch[:,:,0:h,2*w:3*w]=x_enc.cpu()
save_batch[:,:,h:2*h,w:2*w]=x_recover.cpu()
save_batch[:,:,h:2*h,2*w:3*w]=x_wrong.cpu()
for k in range(b):
torchvision.utils.save_image(save_batch[k],join(output_frame_path,img_name[k]),normalize=True)
def multi_pwd_loop(self):
output_frame_path=join(self.exp_dir,'frames')
output_video_path=join(self.exp_dir,f'{self.exp_dir.split("/")[-1]}.mp4')
os.makedirs(output_frame_path,exist_ok=True)
with torch.no_grad():
for j,(latent_batch,image_batch) in enumerate(zip(self.latent_loader,self.image_loader)):
w_ori=latent_batch[0].cuda()
x_ori_real=image_batch[0] #real image
x_ori, _ = self.decoder([w_ori], input_is_latent=True, randomize_noise=False, truncation=1)
b,c,h,w = x_ori.shape
img_name=[f"{(b_idx+j*b):0>4d}.jpg" for b_idx in range(b)]
print(img_name)
save_batch=torch.zeros((b,c,3*h,(self.opts.pwd_num+1)*w))
save_batch[:,:,0:h,0:w]=x_ori_real
save_batch[:,:,h:h+h,0:w]=x_ori.cpu()
for pwd_idx in range(len(self.w_pwd_list)):
w_pwd=self.w_pwd_list[pwd_idx].repeat(b,1,1)
w_wrong_pwd=self.w_wrong_pwd_list[pwd_idx].repeat(b,1,1)
w_enc=self.mapper(torch.cat([w_ori,w_pwd],dim=-1))
x_enc, _ = self.decoder([w_enc], input_is_latent=True, randomize_noise=False, truncation=1)
w_cat_dec=torch.cat([w_enc,w_pwd],dim=-1)
w_dec=self.mapper(w_cat_dec)
x_recover, _ = self.decoder([w_dec], input_is_latent=True, randomize_noise=False, truncation=1)
w_cat_dec_wrong=torch.cat([w_enc,w_wrong_pwd],dim=-1)
w_dec_wrong=self.mapper(w_cat_dec_wrong)
x_wrong, _ = self.decoder([w_dec_wrong], input_is_latent=True, randomize_noise=False, truncation=1)
save_batch[:,:,0:h,(1+pwd_idx)*w:(2+pwd_idx)*w]=x_enc.cpu()
save_batch[:,:,h:2*h,(1+pwd_idx)*w:(2+pwd_idx)*w]=x_recover.cpu()
save_batch[:,:,2*h:3*h,(1+pwd_idx)*w:(2+pwd_idx)*w]=x_wrong.cpu()
for k in range(b):
torchvision.utils.save_image(save_batch[k],join(output_frame_path,img_name[k]),normalize=True)
if __name__=='__main__':
opts=TestOptions().parse()
Coach=CoachTest(opts)
Coach.single_pwd_loop()
#Coach.multi_pwd_loop()