Spaces:
No application file
No application file
culture
commited on
Commit
·
2cce121
1
Parent(s):
e0804a6
Upload gfpgan/models/gfpgan_model.py
Browse files- gfpgan/models/gfpgan_model.py +580 -0
gfpgan/models/gfpgan_model.py
ADDED
@@ -0,0 +1,580 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
import os.path as osp
|
3 |
+
import torch
|
4 |
+
from basicsr.archs import build_network
|
5 |
+
from basicsr.losses import build_loss
|
6 |
+
from basicsr.losses.losses import r1_penalty
|
7 |
+
from basicsr.metrics import calculate_metric
|
8 |
+
from basicsr.models.base_model import BaseModel
|
9 |
+
from basicsr.utils import get_root_logger, imwrite, tensor2img
|
10 |
+
from basicsr.utils.registry import MODEL_REGISTRY
|
11 |
+
from collections import OrderedDict
|
12 |
+
from torch.nn import functional as F
|
13 |
+
from torchvision.ops import roi_align
|
14 |
+
from tqdm import tqdm
|
15 |
+
|
16 |
+
|
17 |
+
@MODEL_REGISTRY.register()
|
18 |
+
class GFPGANModel(BaseModel):
|
19 |
+
"""The GFPGAN model for Towards real-world blind face restoratin with generative facial prior"""
|
20 |
+
|
21 |
+
def __init__(self, opt):
|
22 |
+
super(GFPGANModel, self).__init__(opt)
|
23 |
+
self.idx = 0 # it is used for saving data for check
|
24 |
+
|
25 |
+
# define network
|
26 |
+
self.net_g = build_network(opt['network_g'])
|
27 |
+
self.net_g = self.model_to_device(self.net_g)
|
28 |
+
self.print_network(self.net_g)
|
29 |
+
|
30 |
+
# load pretrained model
|
31 |
+
load_path = self.opt['path'].get('pretrain_network_g', None)
|
32 |
+
if load_path is not None:
|
33 |
+
param_key = self.opt['path'].get('param_key_g', 'params')
|
34 |
+
self.load_network(self.net_g, load_path, self.opt['path'].get('strict_load_g', True), param_key)
|
35 |
+
|
36 |
+
self.log_size = int(math.log(self.opt['network_g']['out_size'], 2))
|
37 |
+
|
38 |
+
if self.is_train:
|
39 |
+
self.init_training_settings()
|
40 |
+
|
41 |
+
def init_training_settings(self):
|
42 |
+
train_opt = self.opt['train']
|
43 |
+
|
44 |
+
# ----------- define net_d ----------- #
|
45 |
+
self.net_d = build_network(self.opt['network_d'])
|
46 |
+
self.net_d = self.model_to_device(self.net_d)
|
47 |
+
self.print_network(self.net_d)
|
48 |
+
# load pretrained model
|
49 |
+
load_path = self.opt['path'].get('pretrain_network_d', None)
|
50 |
+
if load_path is not None:
|
51 |
+
self.load_network(self.net_d, load_path, self.opt['path'].get('strict_load_d', True))
|
52 |
+
|
53 |
+
# ----------- define net_g with Exponential Moving Average (EMA) ----------- #
|
54 |
+
# net_g_ema only used for testing on one GPU and saving. There is no need to wrap with DistributedDataParallel
|
55 |
+
self.net_g_ema = build_network(self.opt['network_g']).to(self.device)
|
56 |
+
# load pretrained model
|
57 |
+
load_path = self.opt['path'].get('pretrain_network_g', None)
|
58 |
+
if load_path is not None:
|
59 |
+
self.load_network(self.net_g_ema, load_path, self.opt['path'].get('strict_load_g', True), 'params_ema')
|
60 |
+
else:
|
61 |
+
self.model_ema(0) # copy net_g weight
|
62 |
+
|
63 |
+
self.net_g.train()
|
64 |
+
self.net_d.train()
|
65 |
+
self.net_g_ema.eval()
|
66 |
+
|
67 |
+
# ----------- facial component networks ----------- #
|
68 |
+
if ('network_d_left_eye' in self.opt and 'network_d_right_eye' in self.opt and 'network_d_mouth' in self.opt):
|
69 |
+
self.use_facial_disc = True
|
70 |
+
else:
|
71 |
+
self.use_facial_disc = False
|
72 |
+
|
73 |
+
if self.use_facial_disc:
|
74 |
+
# left eye
|
75 |
+
self.net_d_left_eye = build_network(self.opt['network_d_left_eye'])
|
76 |
+
self.net_d_left_eye = self.model_to_device(self.net_d_left_eye)
|
77 |
+
self.print_network(self.net_d_left_eye)
|
78 |
+
load_path = self.opt['path'].get('pretrain_network_d_left_eye')
|
79 |
+
if load_path is not None:
|
80 |
+
self.load_network(self.net_d_left_eye, load_path, True, 'params')
|
81 |
+
# right eye
|
82 |
+
self.net_d_right_eye = build_network(self.opt['network_d_right_eye'])
|
83 |
+
self.net_d_right_eye = self.model_to_device(self.net_d_right_eye)
|
84 |
+
self.print_network(self.net_d_right_eye)
|
85 |
+
load_path = self.opt['path'].get('pretrain_network_d_right_eye')
|
86 |
+
if load_path is not None:
|
87 |
+
self.load_network(self.net_d_right_eye, load_path, True, 'params')
|
88 |
+
# mouth
|
89 |
+
self.net_d_mouth = build_network(self.opt['network_d_mouth'])
|
90 |
+
self.net_d_mouth = self.model_to_device(self.net_d_mouth)
|
91 |
+
self.print_network(self.net_d_mouth)
|
92 |
+
load_path = self.opt['path'].get('pretrain_network_d_mouth')
|
93 |
+
if load_path is not None:
|
94 |
+
self.load_network(self.net_d_mouth, load_path, True, 'params')
|
95 |
+
|
96 |
+
self.net_d_left_eye.train()
|
97 |
+
self.net_d_right_eye.train()
|
98 |
+
self.net_d_mouth.train()
|
99 |
+
|
100 |
+
# ----------- define facial component gan loss ----------- #
|
101 |
+
self.cri_component = build_loss(train_opt['gan_component_opt']).to(self.device)
|
102 |
+
|
103 |
+
# ----------- define losses ----------- #
|
104 |
+
# pixel loss
|
105 |
+
if train_opt.get('pixel_opt'):
|
106 |
+
self.cri_pix = build_loss(train_opt['pixel_opt']).to(self.device)
|
107 |
+
else:
|
108 |
+
self.cri_pix = None
|
109 |
+
|
110 |
+
# perceptual loss
|
111 |
+
if train_opt.get('perceptual_opt'):
|
112 |
+
self.cri_perceptual = build_loss(train_opt['perceptual_opt']).to(self.device)
|
113 |
+
else:
|
114 |
+
self.cri_perceptual = None
|
115 |
+
|
116 |
+
# L1 loss is used in pyramid loss, component style loss and identity loss
|
117 |
+
self.cri_l1 = build_loss(train_opt['L1_opt']).to(self.device)
|
118 |
+
|
119 |
+
# gan loss (wgan)
|
120 |
+
self.cri_gan = build_loss(train_opt['gan_opt']).to(self.device)
|
121 |
+
|
122 |
+
# ----------- define identity loss ----------- #
|
123 |
+
if 'network_identity' in self.opt:
|
124 |
+
self.use_identity = True
|
125 |
+
else:
|
126 |
+
self.use_identity = False
|
127 |
+
|
128 |
+
if self.use_identity:
|
129 |
+
# define identity network
|
130 |
+
self.network_identity = build_network(self.opt['network_identity'])
|
131 |
+
self.network_identity = self.model_to_device(self.network_identity)
|
132 |
+
self.print_network(self.network_identity)
|
133 |
+
load_path = self.opt['path'].get('pretrain_network_identity')
|
134 |
+
if load_path is not None:
|
135 |
+
self.load_network(self.network_identity, load_path, True, None)
|
136 |
+
self.network_identity.eval()
|
137 |
+
for param in self.network_identity.parameters():
|
138 |
+
param.requires_grad = False
|
139 |
+
|
140 |
+
# regularization weights
|
141 |
+
self.r1_reg_weight = train_opt['r1_reg_weight'] # for discriminator
|
142 |
+
self.net_d_iters = train_opt.get('net_d_iters', 1)
|
143 |
+
self.net_d_init_iters = train_opt.get('net_d_init_iters', 0)
|
144 |
+
self.net_d_reg_every = train_opt['net_d_reg_every']
|
145 |
+
|
146 |
+
# set up optimizers and schedulers
|
147 |
+
self.setup_optimizers()
|
148 |
+
self.setup_schedulers()
|
149 |
+
|
150 |
+
def setup_optimizers(self):
|
151 |
+
train_opt = self.opt['train']
|
152 |
+
|
153 |
+
# ----------- optimizer g ----------- #
|
154 |
+
net_g_reg_ratio = 1
|
155 |
+
normal_params = []
|
156 |
+
for _, param in self.net_g.named_parameters():
|
157 |
+
normal_params.append(param)
|
158 |
+
optim_params_g = [{ # add normal params first
|
159 |
+
'params': normal_params,
|
160 |
+
'lr': train_opt['optim_g']['lr']
|
161 |
+
}]
|
162 |
+
optim_type = train_opt['optim_g'].pop('type')
|
163 |
+
lr = train_opt['optim_g']['lr'] * net_g_reg_ratio
|
164 |
+
betas = (0**net_g_reg_ratio, 0.99**net_g_reg_ratio)
|
165 |
+
self.optimizer_g = self.get_optimizer(optim_type, optim_params_g, lr, betas=betas)
|
166 |
+
self.optimizers.append(self.optimizer_g)
|
167 |
+
|
168 |
+
# ----------- optimizer d ----------- #
|
169 |
+
net_d_reg_ratio = self.net_d_reg_every / (self.net_d_reg_every + 1)
|
170 |
+
normal_params = []
|
171 |
+
for _, param in self.net_d.named_parameters():
|
172 |
+
normal_params.append(param)
|
173 |
+
optim_params_d = [{ # add normal params first
|
174 |
+
'params': normal_params,
|
175 |
+
'lr': train_opt['optim_d']['lr']
|
176 |
+
}]
|
177 |
+
optim_type = train_opt['optim_d'].pop('type')
|
178 |
+
lr = train_opt['optim_d']['lr'] * net_d_reg_ratio
|
179 |
+
betas = (0**net_d_reg_ratio, 0.99**net_d_reg_ratio)
|
180 |
+
self.optimizer_d = self.get_optimizer(optim_type, optim_params_d, lr, betas=betas)
|
181 |
+
self.optimizers.append(self.optimizer_d)
|
182 |
+
|
183 |
+
# ----------- optimizers for facial component networks ----------- #
|
184 |
+
if self.use_facial_disc:
|
185 |
+
# setup optimizers for facial component discriminators
|
186 |
+
optim_type = train_opt['optim_component'].pop('type')
|
187 |
+
lr = train_opt['optim_component']['lr']
|
188 |
+
# left eye
|
189 |
+
self.optimizer_d_left_eye = self.get_optimizer(
|
190 |
+
optim_type, self.net_d_left_eye.parameters(), lr, betas=(0.9, 0.99))
|
191 |
+
self.optimizers.append(self.optimizer_d_left_eye)
|
192 |
+
# right eye
|
193 |
+
self.optimizer_d_right_eye = self.get_optimizer(
|
194 |
+
optim_type, self.net_d_right_eye.parameters(), lr, betas=(0.9, 0.99))
|
195 |
+
self.optimizers.append(self.optimizer_d_right_eye)
|
196 |
+
# mouth
|
197 |
+
self.optimizer_d_mouth = self.get_optimizer(
|
198 |
+
optim_type, self.net_d_mouth.parameters(), lr, betas=(0.9, 0.99))
|
199 |
+
self.optimizers.append(self.optimizer_d_mouth)
|
200 |
+
|
201 |
+
def feed_data(self, data):
|
202 |
+
self.lq = data['lq'].to(self.device)
|
203 |
+
if 'gt' in data:
|
204 |
+
self.gt = data['gt'].to(self.device)
|
205 |
+
|
206 |
+
if 'loc_left_eye' in data:
|
207 |
+
# get facial component locations, shape (batch, 4)
|
208 |
+
self.loc_left_eyes = data['loc_left_eye']
|
209 |
+
self.loc_right_eyes = data['loc_right_eye']
|
210 |
+
self.loc_mouths = data['loc_mouth']
|
211 |
+
|
212 |
+
# uncomment to check data
|
213 |
+
# import torchvision
|
214 |
+
# if self.opt['rank'] == 0:
|
215 |
+
# import os
|
216 |
+
# os.makedirs('tmp/gt', exist_ok=True)
|
217 |
+
# os.makedirs('tmp/lq', exist_ok=True)
|
218 |
+
# print(self.idx)
|
219 |
+
# torchvision.utils.save_image(
|
220 |
+
# self.gt, f'tmp/gt/gt_{self.idx}.png', nrow=4, padding=2, normalize=True, range=(-1, 1))
|
221 |
+
# torchvision.utils.save_image(
|
222 |
+
# self.lq, f'tmp/lq/lq{self.idx}.png', nrow=4, padding=2, normalize=True, range=(-1, 1))
|
223 |
+
# self.idx = self.idx + 1
|
224 |
+
|
225 |
+
def construct_img_pyramid(self):
|
226 |
+
"""Construct image pyramid for intermediate restoration loss"""
|
227 |
+
pyramid_gt = [self.gt]
|
228 |
+
down_img = self.gt
|
229 |
+
for _ in range(0, self.log_size - 3):
|
230 |
+
down_img = F.interpolate(down_img, scale_factor=0.5, mode='bilinear', align_corners=False)
|
231 |
+
pyramid_gt.insert(0, down_img)
|
232 |
+
return pyramid_gt
|
233 |
+
|
234 |
+
def get_roi_regions(self, eye_out_size=80, mouth_out_size=120):
|
235 |
+
face_ratio = int(self.opt['network_g']['out_size'] / 512)
|
236 |
+
eye_out_size *= face_ratio
|
237 |
+
mouth_out_size *= face_ratio
|
238 |
+
|
239 |
+
rois_eyes = []
|
240 |
+
rois_mouths = []
|
241 |
+
for b in range(self.loc_left_eyes.size(0)): # loop for batch size
|
242 |
+
# left eye and right eye
|
243 |
+
img_inds = self.loc_left_eyes.new_full((2, 1), b)
|
244 |
+
bbox = torch.stack([self.loc_left_eyes[b, :], self.loc_right_eyes[b, :]], dim=0) # shape: (2, 4)
|
245 |
+
rois = torch.cat([img_inds, bbox], dim=-1) # shape: (2, 5)
|
246 |
+
rois_eyes.append(rois)
|
247 |
+
# mouse
|
248 |
+
img_inds = self.loc_left_eyes.new_full((1, 1), b)
|
249 |
+
rois = torch.cat([img_inds, self.loc_mouths[b:b + 1, :]], dim=-1) # shape: (1, 5)
|
250 |
+
rois_mouths.append(rois)
|
251 |
+
|
252 |
+
rois_eyes = torch.cat(rois_eyes, 0).to(self.device)
|
253 |
+
rois_mouths = torch.cat(rois_mouths, 0).to(self.device)
|
254 |
+
|
255 |
+
# real images
|
256 |
+
all_eyes = roi_align(self.gt, boxes=rois_eyes, output_size=eye_out_size) * face_ratio
|
257 |
+
self.left_eyes_gt = all_eyes[0::2, :, :, :]
|
258 |
+
self.right_eyes_gt = all_eyes[1::2, :, :, :]
|
259 |
+
self.mouths_gt = roi_align(self.gt, boxes=rois_mouths, output_size=mouth_out_size) * face_ratio
|
260 |
+
# output
|
261 |
+
all_eyes = roi_align(self.output, boxes=rois_eyes, output_size=eye_out_size) * face_ratio
|
262 |
+
self.left_eyes = all_eyes[0::2, :, :, :]
|
263 |
+
self.right_eyes = all_eyes[1::2, :, :, :]
|
264 |
+
self.mouths = roi_align(self.output, boxes=rois_mouths, output_size=mouth_out_size) * face_ratio
|
265 |
+
|
266 |
+
def _gram_mat(self, x):
|
267 |
+
"""Calculate Gram matrix.
|
268 |
+
|
269 |
+
Args:
|
270 |
+
x (torch.Tensor): Tensor with shape of (n, c, h, w).
|
271 |
+
|
272 |
+
Returns:
|
273 |
+
torch.Tensor: Gram matrix.
|
274 |
+
"""
|
275 |
+
n, c, h, w = x.size()
|
276 |
+
features = x.view(n, c, w * h)
|
277 |
+
features_t = features.transpose(1, 2)
|
278 |
+
gram = features.bmm(features_t) / (c * h * w)
|
279 |
+
return gram
|
280 |
+
|
281 |
+
def gray_resize_for_identity(self, out, size=128):
|
282 |
+
out_gray = (0.2989 * out[:, 0, :, :] + 0.5870 * out[:, 1, :, :] + 0.1140 * out[:, 2, :, :])
|
283 |
+
out_gray = out_gray.unsqueeze(1)
|
284 |
+
out_gray = F.interpolate(out_gray, (size, size), mode='bilinear', align_corners=False)
|
285 |
+
return out_gray
|
286 |
+
|
287 |
+
def optimize_parameters(self, current_iter):
|
288 |
+
# optimize net_g
|
289 |
+
for p in self.net_d.parameters():
|
290 |
+
p.requires_grad = False
|
291 |
+
self.optimizer_g.zero_grad()
|
292 |
+
|
293 |
+
# do not update facial component net_d
|
294 |
+
if self.use_facial_disc:
|
295 |
+
for p in self.net_d_left_eye.parameters():
|
296 |
+
p.requires_grad = False
|
297 |
+
for p in self.net_d_right_eye.parameters():
|
298 |
+
p.requires_grad = False
|
299 |
+
for p in self.net_d_mouth.parameters():
|
300 |
+
p.requires_grad = False
|
301 |
+
|
302 |
+
# image pyramid loss weight
|
303 |
+
if current_iter < self.opt['train'].get('remove_pyramid_loss', float('inf')):
|
304 |
+
pyramid_loss_weight = self.opt['train'].get('pyramid_loss_weight', 1)
|
305 |
+
else:
|
306 |
+
pyramid_loss_weight = 1e-12 # very small loss
|
307 |
+
if pyramid_loss_weight > 0:
|
308 |
+
self.output, out_rgbs = self.net_g(self.lq, return_rgb=True)
|
309 |
+
pyramid_gt = self.construct_img_pyramid()
|
310 |
+
else:
|
311 |
+
self.output, out_rgbs = self.net_g(self.lq, return_rgb=False)
|
312 |
+
|
313 |
+
# get roi-align regions
|
314 |
+
if self.use_facial_disc:
|
315 |
+
self.get_roi_regions(eye_out_size=80, mouth_out_size=120)
|
316 |
+
|
317 |
+
l_g_total = 0
|
318 |
+
loss_dict = OrderedDict()
|
319 |
+
if (current_iter % self.net_d_iters == 0 and current_iter > self.net_d_init_iters):
|
320 |
+
# pixel loss
|
321 |
+
if self.cri_pix:
|
322 |
+
l_g_pix = self.cri_pix(self.output, self.gt)
|
323 |
+
l_g_total += l_g_pix
|
324 |
+
loss_dict['l_g_pix'] = l_g_pix
|
325 |
+
|
326 |
+
# image pyramid loss
|
327 |
+
if pyramid_loss_weight > 0:
|
328 |
+
for i in range(0, self.log_size - 2):
|
329 |
+
l_pyramid = self.cri_l1(out_rgbs[i], pyramid_gt[i]) * pyramid_loss_weight
|
330 |
+
l_g_total += l_pyramid
|
331 |
+
loss_dict[f'l_p_{2**(i+3)}'] = l_pyramid
|
332 |
+
|
333 |
+
# perceptual loss
|
334 |
+
if self.cri_perceptual:
|
335 |
+
l_g_percep, l_g_style = self.cri_perceptual(self.output, self.gt)
|
336 |
+
if l_g_percep is not None:
|
337 |
+
l_g_total += l_g_percep
|
338 |
+
loss_dict['l_g_percep'] = l_g_percep
|
339 |
+
if l_g_style is not None:
|
340 |
+
l_g_total += l_g_style
|
341 |
+
loss_dict['l_g_style'] = l_g_style
|
342 |
+
|
343 |
+
# gan loss
|
344 |
+
fake_g_pred = self.net_d(self.output)
|
345 |
+
l_g_gan = self.cri_gan(fake_g_pred, True, is_disc=False)
|
346 |
+
l_g_total += l_g_gan
|
347 |
+
loss_dict['l_g_gan'] = l_g_gan
|
348 |
+
|
349 |
+
# facial component loss
|
350 |
+
if self.use_facial_disc:
|
351 |
+
# left eye
|
352 |
+
fake_left_eye, fake_left_eye_feats = self.net_d_left_eye(self.left_eyes, return_feats=True)
|
353 |
+
l_g_gan = self.cri_component(fake_left_eye, True, is_disc=False)
|
354 |
+
l_g_total += l_g_gan
|
355 |
+
loss_dict['l_g_gan_left_eye'] = l_g_gan
|
356 |
+
# right eye
|
357 |
+
fake_right_eye, fake_right_eye_feats = self.net_d_right_eye(self.right_eyes, return_feats=True)
|
358 |
+
l_g_gan = self.cri_component(fake_right_eye, True, is_disc=False)
|
359 |
+
l_g_total += l_g_gan
|
360 |
+
loss_dict['l_g_gan_right_eye'] = l_g_gan
|
361 |
+
# mouth
|
362 |
+
fake_mouth, fake_mouth_feats = self.net_d_mouth(self.mouths, return_feats=True)
|
363 |
+
l_g_gan = self.cri_component(fake_mouth, True, is_disc=False)
|
364 |
+
l_g_total += l_g_gan
|
365 |
+
loss_dict['l_g_gan_mouth'] = l_g_gan
|
366 |
+
|
367 |
+
if self.opt['train'].get('comp_style_weight', 0) > 0:
|
368 |
+
# get gt feat
|
369 |
+
_, real_left_eye_feats = self.net_d_left_eye(self.left_eyes_gt, return_feats=True)
|
370 |
+
_, real_right_eye_feats = self.net_d_right_eye(self.right_eyes_gt, return_feats=True)
|
371 |
+
_, real_mouth_feats = self.net_d_mouth(self.mouths_gt, return_feats=True)
|
372 |
+
|
373 |
+
def _comp_style(feat, feat_gt, criterion):
|
374 |
+
return criterion(self._gram_mat(feat[0]), self._gram_mat(
|
375 |
+
feat_gt[0].detach())) * 0.5 + criterion(
|
376 |
+
self._gram_mat(feat[1]), self._gram_mat(feat_gt[1].detach()))
|
377 |
+
|
378 |
+
# facial component style loss
|
379 |
+
comp_style_loss = 0
|
380 |
+
comp_style_loss += _comp_style(fake_left_eye_feats, real_left_eye_feats, self.cri_l1)
|
381 |
+
comp_style_loss += _comp_style(fake_right_eye_feats, real_right_eye_feats, self.cri_l1)
|
382 |
+
comp_style_loss += _comp_style(fake_mouth_feats, real_mouth_feats, self.cri_l1)
|
383 |
+
comp_style_loss = comp_style_loss * self.opt['train']['comp_style_weight']
|
384 |
+
l_g_total += comp_style_loss
|
385 |
+
loss_dict['l_g_comp_style_loss'] = comp_style_loss
|
386 |
+
|
387 |
+
# identity loss
|
388 |
+
if self.use_identity:
|
389 |
+
identity_weight = self.opt['train']['identity_weight']
|
390 |
+
# get gray images and resize
|
391 |
+
out_gray = self.gray_resize_for_identity(self.output)
|
392 |
+
gt_gray = self.gray_resize_for_identity(self.gt)
|
393 |
+
|
394 |
+
identity_gt = self.network_identity(gt_gray).detach()
|
395 |
+
identity_out = self.network_identity(out_gray)
|
396 |
+
l_identity = self.cri_l1(identity_out, identity_gt) * identity_weight
|
397 |
+
l_g_total += l_identity
|
398 |
+
loss_dict['l_identity'] = l_identity
|
399 |
+
|
400 |
+
l_g_total.backward()
|
401 |
+
self.optimizer_g.step()
|
402 |
+
|
403 |
+
# EMA
|
404 |
+
self.model_ema(decay=0.5**(32 / (10 * 1000)))
|
405 |
+
|
406 |
+
# ----------- optimize net_d ----------- #
|
407 |
+
for p in self.net_d.parameters():
|
408 |
+
p.requires_grad = True
|
409 |
+
self.optimizer_d.zero_grad()
|
410 |
+
if self.use_facial_disc:
|
411 |
+
for p in self.net_d_left_eye.parameters():
|
412 |
+
p.requires_grad = True
|
413 |
+
for p in self.net_d_right_eye.parameters():
|
414 |
+
p.requires_grad = True
|
415 |
+
for p in self.net_d_mouth.parameters():
|
416 |
+
p.requires_grad = True
|
417 |
+
self.optimizer_d_left_eye.zero_grad()
|
418 |
+
self.optimizer_d_right_eye.zero_grad()
|
419 |
+
self.optimizer_d_mouth.zero_grad()
|
420 |
+
|
421 |
+
fake_d_pred = self.net_d(self.output.detach())
|
422 |
+
real_d_pred = self.net_d(self.gt)
|
423 |
+
l_d = self.cri_gan(real_d_pred, True, is_disc=True) + self.cri_gan(fake_d_pred, False, is_disc=True)
|
424 |
+
loss_dict['l_d'] = l_d
|
425 |
+
# In WGAN, real_score should be positive and fake_score should be negative
|
426 |
+
loss_dict['real_score'] = real_d_pred.detach().mean()
|
427 |
+
loss_dict['fake_score'] = fake_d_pred.detach().mean()
|
428 |
+
l_d.backward()
|
429 |
+
|
430 |
+
# regularization loss
|
431 |
+
if current_iter % self.net_d_reg_every == 0:
|
432 |
+
self.gt.requires_grad = True
|
433 |
+
real_pred = self.net_d(self.gt)
|
434 |
+
l_d_r1 = r1_penalty(real_pred, self.gt)
|
435 |
+
l_d_r1 = (self.r1_reg_weight / 2 * l_d_r1 * self.net_d_reg_every + 0 * real_pred[0])
|
436 |
+
loss_dict['l_d_r1'] = l_d_r1.detach().mean()
|
437 |
+
l_d_r1.backward()
|
438 |
+
|
439 |
+
self.optimizer_d.step()
|
440 |
+
|
441 |
+
# optimize facial component discriminators
|
442 |
+
if self.use_facial_disc:
|
443 |
+
# left eye
|
444 |
+
fake_d_pred, _ = self.net_d_left_eye(self.left_eyes.detach())
|
445 |
+
real_d_pred, _ = self.net_d_left_eye(self.left_eyes_gt)
|
446 |
+
l_d_left_eye = self.cri_component(
|
447 |
+
real_d_pred, True, is_disc=True) + self.cri_gan(
|
448 |
+
fake_d_pred, False, is_disc=True)
|
449 |
+
loss_dict['l_d_left_eye'] = l_d_left_eye
|
450 |
+
l_d_left_eye.backward()
|
451 |
+
# right eye
|
452 |
+
fake_d_pred, _ = self.net_d_right_eye(self.right_eyes.detach())
|
453 |
+
real_d_pred, _ = self.net_d_right_eye(self.right_eyes_gt)
|
454 |
+
l_d_right_eye = self.cri_component(
|
455 |
+
real_d_pred, True, is_disc=True) + self.cri_gan(
|
456 |
+
fake_d_pred, False, is_disc=True)
|
457 |
+
loss_dict['l_d_right_eye'] = l_d_right_eye
|
458 |
+
l_d_right_eye.backward()
|
459 |
+
# mouth
|
460 |
+
fake_d_pred, _ = self.net_d_mouth(self.mouths.detach())
|
461 |
+
real_d_pred, _ = self.net_d_mouth(self.mouths_gt)
|
462 |
+
l_d_mouth = self.cri_component(
|
463 |
+
real_d_pred, True, is_disc=True) + self.cri_gan(
|
464 |
+
fake_d_pred, False, is_disc=True)
|
465 |
+
loss_dict['l_d_mouth'] = l_d_mouth
|
466 |
+
l_d_mouth.backward()
|
467 |
+
|
468 |
+
self.optimizer_d_left_eye.step()
|
469 |
+
self.optimizer_d_right_eye.step()
|
470 |
+
self.optimizer_d_mouth.step()
|
471 |
+
|
472 |
+
self.log_dict = self.reduce_loss_dict(loss_dict)
|
473 |
+
|
474 |
+
def test(self):
|
475 |
+
with torch.no_grad():
|
476 |
+
if hasattr(self, 'net_g_ema'):
|
477 |
+
self.net_g_ema.eval()
|
478 |
+
self.output, _ = self.net_g_ema(self.lq)
|
479 |
+
else:
|
480 |
+
logger = get_root_logger()
|
481 |
+
logger.warning('Do not have self.net_g_ema, use self.net_g.')
|
482 |
+
self.net_g.eval()
|
483 |
+
self.output, _ = self.net_g(self.lq)
|
484 |
+
self.net_g.train()
|
485 |
+
|
486 |
+
def dist_validation(self, dataloader, current_iter, tb_logger, save_img):
|
487 |
+
if self.opt['rank'] == 0:
|
488 |
+
self.nondist_validation(dataloader, current_iter, tb_logger, save_img)
|
489 |
+
|
490 |
+
def nondist_validation(self, dataloader, current_iter, tb_logger, save_img):
|
491 |
+
dataset_name = dataloader.dataset.opt['name']
|
492 |
+
with_metrics = self.opt['val'].get('metrics') is not None
|
493 |
+
use_pbar = self.opt['val'].get('pbar', False)
|
494 |
+
|
495 |
+
if with_metrics:
|
496 |
+
if not hasattr(self, 'metric_results'): # only execute in the first run
|
497 |
+
self.metric_results = {metric: 0 for metric in self.opt['val']['metrics'].keys()}
|
498 |
+
# initialize the best metric results for each dataset_name (supporting multiple validation datasets)
|
499 |
+
self._initialize_best_metric_results(dataset_name)
|
500 |
+
# zero self.metric_results
|
501 |
+
self.metric_results = {metric: 0 for metric in self.metric_results}
|
502 |
+
|
503 |
+
metric_data = dict()
|
504 |
+
if use_pbar:
|
505 |
+
pbar = tqdm(total=len(dataloader), unit='image')
|
506 |
+
|
507 |
+
for idx, val_data in enumerate(dataloader):
|
508 |
+
img_name = osp.splitext(osp.basename(val_data['lq_path'][0]))[0]
|
509 |
+
self.feed_data(val_data)
|
510 |
+
self.test()
|
511 |
+
|
512 |
+
sr_img = tensor2img(self.output.detach().cpu(), min_max=(-1, 1))
|
513 |
+
metric_data['img'] = sr_img
|
514 |
+
if hasattr(self, 'gt'):
|
515 |
+
gt_img = tensor2img(self.gt.detach().cpu(), min_max=(-1, 1))
|
516 |
+
metric_data['img2'] = gt_img
|
517 |
+
del self.gt
|
518 |
+
|
519 |
+
# tentative for out of GPU memory
|
520 |
+
del self.lq
|
521 |
+
del self.output
|
522 |
+
torch.cuda.empty_cache()
|
523 |
+
|
524 |
+
if save_img:
|
525 |
+
if self.opt['is_train']:
|
526 |
+
save_img_path = osp.join(self.opt['path']['visualization'], img_name,
|
527 |
+
f'{img_name}_{current_iter}.png')
|
528 |
+
else:
|
529 |
+
if self.opt['val']['suffix']:
|
530 |
+
save_img_path = osp.join(self.opt['path']['visualization'], dataset_name,
|
531 |
+
f'{img_name}_{self.opt["val"]["suffix"]}.png')
|
532 |
+
else:
|
533 |
+
save_img_path = osp.join(self.opt['path']['visualization'], dataset_name,
|
534 |
+
f'{img_name}_{self.opt["name"]}.png')
|
535 |
+
imwrite(sr_img, save_img_path)
|
536 |
+
|
537 |
+
if with_metrics:
|
538 |
+
# calculate metrics
|
539 |
+
for name, opt_ in self.opt['val']['metrics'].items():
|
540 |
+
self.metric_results[name] += calculate_metric(metric_data, opt_)
|
541 |
+
if use_pbar:
|
542 |
+
pbar.update(1)
|
543 |
+
pbar.set_description(f'Test {img_name}')
|
544 |
+
if use_pbar:
|
545 |
+
pbar.close()
|
546 |
+
|
547 |
+
if with_metrics:
|
548 |
+
for metric in self.metric_results.keys():
|
549 |
+
self.metric_results[metric] /= (idx + 1)
|
550 |
+
# update the best metric result
|
551 |
+
self._update_best_metric_result(dataset_name, metric, self.metric_results[metric], current_iter)
|
552 |
+
|
553 |
+
self._log_validation_metric_values(current_iter, dataset_name, tb_logger)
|
554 |
+
|
555 |
+
def _log_validation_metric_values(self, current_iter, dataset_name, tb_logger):
|
556 |
+
log_str = f'Validation {dataset_name}\n'
|
557 |
+
for metric, value in self.metric_results.items():
|
558 |
+
log_str += f'\t # {metric}: {value:.4f}'
|
559 |
+
if hasattr(self, 'best_metric_results'):
|
560 |
+
log_str += (f'\tBest: {self.best_metric_results[dataset_name][metric]["val"]:.4f} @ '
|
561 |
+
f'{self.best_metric_results[dataset_name][metric]["iter"]} iter')
|
562 |
+
log_str += '\n'
|
563 |
+
|
564 |
+
logger = get_root_logger()
|
565 |
+
logger.info(log_str)
|
566 |
+
if tb_logger:
|
567 |
+
for metric, value in self.metric_results.items():
|
568 |
+
tb_logger.add_scalar(f'metrics/{dataset_name}/{metric}', value, current_iter)
|
569 |
+
|
570 |
+
def save(self, epoch, current_iter):
|
571 |
+
# save net_g and net_d
|
572 |
+
self.save_network([self.net_g, self.net_g_ema], 'net_g', current_iter, param_key=['params', 'params_ema'])
|
573 |
+
self.save_network(self.net_d, 'net_d', current_iter)
|
574 |
+
# save component discriminators
|
575 |
+
if self.use_facial_disc:
|
576 |
+
self.save_network(self.net_d_left_eye, 'net_d_left_eye', current_iter)
|
577 |
+
self.save_network(self.net_d_right_eye, 'net_d_right_eye', current_iter)
|
578 |
+
self.save_network(self.net_d_mouth, 'net_d_mouth', current_iter)
|
579 |
+
# save training state
|
580 |
+
self.save_training_state(epoch, current_iter)
|