-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_personalized.py
270 lines (222 loc) · 10.4 KB
/
train_personalized.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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
import os
import pickle
import random
import argparse
import matplotlib.pyplot as plt
import numpy as np
import torch
from typing_extensions import final
from PIL import Image
import torch.nn.functional as F
from torch import optim
from torch.autograd import Variable, grad
from torch.nn.functional import binary_cross_entropy_with_logits
from torchvision.utils import save_image
from torchvision import transforms
from tqdm import tqdm
from lpips import LPIPS
import pytorch_msssim
from dataset import PersonalizedSyntheticDataset
from models.emo_mapping import EmoMappingW, EmoMappingWplus
from models.landmark import FaceAlignment
from models.emonet import EmoNet
from models.vggface2 import VGGFace2
from models.stylegan2_interface import StyleGAN2
from losses import *
from skimage import draw
import glob
is_cuda = torch.cuda.is_available()
device = 'cuda' if is_cuda else 'cpu'
EMO_EMBED = 64
STG_EMBED = 512
INPUT_SIZE = 1024
ITER_NUM = 500
def train(
datapath: str,
stylegan2_checkpoint_path: str,
emo_mapping_checkpoint_path: str,
vggface2_checkpoint_path: str,
emonet_checkpoint_path: str,
log_path: str,
inversion_type: str,
output_path: str,
wplus: bool
):
stylegan = StyleGAN2(
checkpoint_path=stylegan2_checkpoint_path,
stylegan_size=INPUT_SIZE,
is_dnn=True,
is_pkl=True
)
# Fine Tuning StyleGAN2
stylegan.train().requires_grad_(True)
vggface2_net = VGGFace2(vggface2_checkpoint_path)
vggface2_net.eval()
landmark_net = FaceAlignment()
landmark_net.eval()
face_pool = torch.nn.AdaptiveAvgPool2d((256, 256))
ckpt_emo = torch.load(emonet_checkpoint_path)
ckpt_emo = { k.replace('module.',''): v for k,v in ckpt_emo.items() }
emonet = EmoNet(n_expression=8)
emonet.load_state_dict(ckpt_emo, strict=False)
emonet.eval()
ckpt_emo_mapping = torch.load(emo_mapping_checkpoint_path)
if wplus:
emo_mapping = EmoMappingWplus(INPUT_SIZE, EMO_EMBED, STG_EMBED)
else:
emo_mapping = EmoMappingW(EMO_EMBED, STG_EMBED)
emo_mapping.load_state_dict(ckpt_emo_mapping['emo_mapping_state_dict'])
emo_mapping.eval()
kwargs = {'num_workers': 1, 'pin_memory': False} if torch.cuda.is_available() else {}
train_loader = torch.utils.data.DataLoader(PersonalizedSyntheticDataset(datapath, inversion_type=inversion_type), batch_size=1, shuffle=False, **kwargs)
lpips = LPIPS(net='alex').to('cuda' if is_cuda else 'cpu').eval()
optimizer = torch.optim.Adam(stylegan.parameters(), lr=3e-4)
if is_cuda:
stylegan.cuda()
vggface2_net.cuda()
landmark_net.cuda()
emonet.cuda()
face_pool.cuda()
emo_mapping.cuda()
weight_id = 1.5
weight_emotion = 0.2
weight_background = 0.2
weight_landmark = 0.001
weight_reconstruction = 0.2
weight_l2 = 1
weight_lpips = 1
landmark_masks = {
'same_face': torch.BoolTensor(torch.ones([68]).bool()), # same face
'emo_face': torch.BoolTensor([1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, # lower face [0:16]
0, 0, 0, 0, 0, # left eyebrow [17:21]
0, 0, 0, 0, 0, # right eyebrow [22:26]
1, 1, 1, 1, 0, 0, 0, 0, 0, # nose
0, 0, 0, 0, 0, 0, # left eye
0, 0, 0, 0, 0, 0, # right eye
0, 0, 0, 0, 0, 0, # upper lip
0, 0, 0, 0, 0, 0, # lower lip
0, 0, 0, 0, 0, 0, 0, 0]),
}
if is_cuda:
landmark_masks['same_face'] = landmark_masks['same_face'].cuda()
landmark_masks['emo_face'] = landmark_masks['emo_face'].cuda()
iter = 1
for epoch in range(ITER_NUM):
for images, latents in train_loader:
if wplus:
latents = latents
else:
latents = latents[:, 0, :]
if is_cuda:
images, latents = images.cuda(), latents.cuda()
images = Variable(images)
latents = Variable(latents)
latents.requires_grad = True
images_scaled = face_pool(images)
batch_size = images.shape[0]
# images = (images + 1.) / 2.
same_face = iter % 10 < 2
#same_face = False
loss_log = {}
stylegan.zero_grad()
with torch.no_grad():
src_ids = vggface2_net(images_scaled)
orig_latents = latents
src_landmarks, src_landmarks_maxval = landmark_net(images_scaled)
background_masks, valid_bg_mask = compute_background_mask(images_scaled, src_landmarks, src_landmarks_maxval)
if same_face:
src_emotions = emonet(images)
else:
src_emotions = torch.FloatTensor(batch_size, 2).uniform_(-.7, .7).cuda() # limited the valence and arousal to [-7, 7] instead [-1, 1]
fake_latents = orig_latents + emo_mapping(orig_latents, src_emotions)
generated_images = stylegan.generate(fake_latents)
generated_images = (generated_images + 1.) / 2.
generated_images_scaled = face_pool(generated_images)
generated_images_ids = vggface2_net(generated_images_scaled)
id_loss = 0
id_loss = weight_id * \
compute_id_loss(src_ids, generated_images_ids)
loss_log['id_loss'] = "{:.4f}".format(id_loss.item())
generated_emotions = emonet(generated_images)
emotion_loss = weight_emotion * \
compute_emotion_loss(src_emotions, generated_emotions)
loss_log['emo_loss'] = "{:.4f}".format(emotion_loss.item())
generated_landmarks, generated_landmarks_maxval = landmark_net(generated_images_scaled)
landmark_mask = landmark_masks['same_face' if same_face else 'emo_face'].repeat(batch_size).view(batch_size, -1)
landmark_loss = weight_landmark * \
compute_landmark_loss(src_landmarks, src_landmarks_maxval, generated_landmarks, generated_landmarks_maxval, landmark_mask)
loss_log['landmarks'] = "{:.4f}".format(landmark_loss.item())
if same_face:
# Sameface loss 1
# latent_loss = weight_latent * \
# compute_latent_loss(orig_latents, fake_latents)
# loss_log['latent_loss'] = "{:.4f}".format(latent_loss.item())
# reconstruction_loss = weight_reconstruction * \
# compute_pixel_reconstruction_loss(images, generated_images)
# loss_log['recon_loss'] = "{:.4f}".format(reconstruction_loss.item())
# Sameface loss 2
l2_loss = weight_l2 * \
compute_l2_loss(images_scaled, generated_images)
lpips_loss = weight_lpips * \
torch.squeeze(lpips(generated_images, images_scaled))
loss_log['lpips_loss'] = "{:.4f}".format(lpips_loss.item())
reconstruction_loss = l2_loss + lpips_loss
background_loss = 0
else:
if valid_bg_mask:
background_loss = weight_background * \
compute_background_loss(background_masks, images_scaled, generated_images_scaled)
loss_log['bg_loss'] = "{:.4f}".format(background_loss.item())
else:
background_loss = 0
#latent_loss = 0.
reconstruction_loss = 0.
total_losses = reconstruction_loss + \
background_loss + \
landmark_loss + \
emotion_loss + \
id_loss
loss_log['sum_non_gan'] = "{:.4f}".format(total_losses.item())
optimizer.zero_grad()
total_losses.backward()
optimizer.step()
print(epoch, loss_log)
if iter % 10 == 8:
if same_face:
filename = '{:06}_same_face.png'.format(iter)
else:
filename = '{:06}_{:.2f}_{:.2f}_emo_face.png'.format(iter, src_emotions[0, 0].item(), src_emotions[0, 1].item())
if not os.path.exists(log_path):
os.makedirs(log_path)
if not os.path.exists(os.path.join(log_path, "personalized_training")):
os.makedirs(os.path.join(log_path, "personalized_training"))
save_image([images[0], generated_images[0]],
os.path.join(log_path, "personalized_training", filename),
normalize=True)
iter += 1
person_name = os.path.basename(os.path.normpath(datapath))
torch.save(stylegan,
os.path.join(output_path, 'EmoStyle_{}_{}_{}.pt'.format(inversion_type, ITER_NUM, person_name)))
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Personalized training script")
parser.add_argument("--datapath", type=str, default="experiments/personalized_single_4/")
parser.add_argument("--stylegan2_checkpoint_path", type=str, default="pretrained/ffhq2.pkl")
parser.add_argument("--emo_mapping_checkpoint_path", type=str, default="checkpoints/emo_mapping_wplus/emo_mapping_wplus_2.pt")
parser.add_argument("--vggface2_checkpoint_path", type=str, default="pretrained/resnet50_ft_weight.pkl")
parser.add_argument("--emonet_checkpoint_path", type=str, default="pretrained/emonet_8.pth")
parser.add_argument("--log_path", type=str, default="logs/personalized")
parser.add_argument("--inversion_type", type=str, default="e4e")
parser.add_argument("--output_path", type=str, default="checkpoints/")
parser.add_argument("--wplus", type=bool, default=True)
args = parser.parse_args()
train(
datapath=args.datapath,
stylegan2_checkpoint_path=args.stylegan2_checkpoint_path,
emo_mapping_checkpoint_path=args.emo_mapping_checkpoint_path,
vggface2_checkpoint_path=args.vggface2_checkpoint_path,
emonet_checkpoint_path=args.emonet_checkpoint_path,
log_path=args.log_path,
inversion_type=args.inversion_type,
output_path=args.output_path,
wplus=args.wplus
)