-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathtrainer.py
181 lines (139 loc) · 5.72 KB
/
trainer.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
from pathlib import Path
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision.utils import save_image
from tqdm import tqdm
from utils import images2grid
from models.clip.model import CLIP
from models.stylegan.model import load_gan
from models.e4e.model import load_e4e
from models.densepose.model import DenseNet, set_instances
from models.segmentation.model import SegModel
def predict_w(img):
e4e = load_e4e()
e4e = e4e.to(img.device)
w = e4e(img)
return w
def img_loss_fn(im1, im2):
loss = F.mse_loss(im1, im2, reduction='none')
loss = loss.mean((1,2,3)).sum()
return loss
def shape_loss_fn(im1, im2):
loss = F.mse_loss(im1, im2, reduction='none')
loss = loss.mean((1,2,3)).sum()
return loss
def clip_loss_fn(clip_similarity):
loss = (1 - clip_similarity).sum()
return loss
def w_delta_loss_fn(w):
N = w.size(1)
w_ref = w[:, 0].unsqueeze(1).repeat(1, N-1, 1)
w_tar = w[:, 1:]
loss = F.mse_loss(w_ref, w_tar, reduction='none')
loss = loss.mean((1,2)).sum()
return loss
def resize(x):
return F.interpolate(x, (256, 256), mode='bilinear')
class Model(nn.Module):
def __init__(self, clip_model) -> None:
super().__init__()
self.gan = load_gan()
self.clip = CLIP(model_name=clip_model)
self.densenet = DenseNet()
self.segnet = SegModel()
w_mean, w_std = self.gan.get_latent_stats()
self.register_buffer('w_mean', w_mean)
self.register_buffer('w_std', w_std)
def clip_similarity(self, imgs, text_feats):
img_features = self.clip.encode_image(imgs)
clip_similarity = self.clip.compute_similarity(img_features, text_feats)
return clip_similarity
def densenet_forward(self, imgs):
bs = imgs.size(0)
box = torch.tensor([[0, 0, 256, 256]], device=imgs.device)
instances = set_instances(box, bs)
body = self.densenet(imgs, instances)
return body
def deeplab_seg_head(self, imgs):
_, _, head = self.segnet(imgs)
return head
class Trainer():
def __init__(self, opt={}):
self.opt = self.set_defaults(opt)
self.model = Model(clip_model=opt.clip_model).eval().requires_grad_(False)
def set_defaults(self, opt):
weights_dict = {'clip': 1,
'img': 30,
'shape': 10,
'head_shape': 1,
'w_delta': 1}
opt_default = dict(weights_dict=weights_dict,
lr=5e-2,
n_iters=500,
clip_model='RN50x4')
for key in opt_default:
if key not in opt:
opt[key] = opt_default[key]
return opt
def to(self, device):
self.model.to(device)
return self
@property
def device(self):
return next(self.model.parameters()).device
@torch.no_grad()
def init(self, imgs_batch, sentence, img_save_dir=None):
self.imgs_batch = imgs_batch.to(self.device)
with torch.no_grad():
text_feats = self.model.clip.encode_text(sentence).view(1, -1) # check the dimensionality
shape_real = self.model.densenet_forward(imgs_batch)
_, body_mask, head_mask = self.model.segnet(imgs_batch)
self.shape_real = shape_real
self.img_mask = 1 - body_mask
self.blend_mask = 1 - head_mask
self.text_feats = text_feats
self.head_shape_init = head_mask
w = predict_w(imgs_batch).cpu()
self.w = w.clone().to(self.device).requires_grad_(True)
self.optimizer = torch.optim.Adam([self.w], lr=self.opt.lr)
if img_save_dir is not None:
img_save_dir = Path(img_save_dir)
save_image(self.img_mask, img_save_dir/'img_mask.jpg')
save_image(self.blend_mask, img_save_dir/'blend_mask.jpg')
def forward(self, n_log=None):
pbar = tqdm(range(self.opt.n_iters+1))
for i in pbar:
w = self.w
imgs_gen = self.model.gan.gen_w(w)
body_shape = self.model.densenet_forward(imgs_gen)
head_shape = self.model.deeplab_seg_head(imgs_gen)
loss_dict = {}
loss_dict['img'] = img_loss_fn(self.img_mask*imgs_gen, self.img_mask*self.imgs_batch)
loss_dict['shape'] = shape_loss_fn(self.shape_real, body_shape)
loss_dict['head_shape'] = shape_loss_fn(self.head_shape_init, head_shape)
loss_dict['w_delta'] = w_delta_loss_fn(w)
clip_sim = self.model.clip_similarity(imgs_gen, self.text_feats)
loss_dict['clip'] = clip_loss_fn(clip_sim)
loss_dict_scaled = {k: loss_dict[k]*self.opt.weights_dict[k] for k in loss_dict}
loss = sum(loss_dict_scaled.values())
loss.backward()
self.optimizer.step()
self.optimizer.zero_grad()
pbar.set_description(f'{loss.item():.3f}')
if n_log is not None and i % n_log == 0:
##### image composition
with torch.no_grad():
imgs_final = self.blend_mask * imgs_gen + (1-self.blend_mask) * self.imgs_batch
yield imgs_final, imgs_gen, loss_dict_scaled
with torch.no_grad():
imgs_final = self.blend_mask * imgs_gen + (1-self.blend_mask) * self.imgs_batch
yield imgs_final, imgs_gen, loss_dict_scaled
def process_gen(self, n_log=20):
for imgs_final, imgs_gen, loss_dict in self.forward(n_log=n_log):
grid = images2grid(self.imgs_batch, imgs_gen, imgs_final)
yield grid, loss_dict
def process(self):
for imgs_final, _, _ in self.forward(n_log=None):
pass
return imgs_final