-
Notifications
You must be signed in to change notification settings - Fork 1
/
inference.py
222 lines (174 loc) · 7.2 KB
/
inference.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
import os
import cv2
import lmdb
import math
import argparse
import numpy as np
from io import BytesIO
from PIL import Image
import pandas as pd
from tqdm import tqdm
import skvideo.io
import numpy as np
import torch
import torchvision.transforms.functional as F
import torchvision.transforms as transforms
import soundfile as sf
from util.logging import init_logging, make_logging_dir
from util.distributed import init_dist
from util.trainer import get_model_optimizer_and_scheduler, set_random_seed, get_trainer
from util.distributed import master_only_print as print
from data.vox_video_dataset import VoxVideoDataset
from config import Config
import torchvision
from midian_pool import MedianPool2d
import warnings
warnings.filterwarnings("ignore")
median = MedianPool2d()
from u2net import U2NET
seg_net = U2NET(3, 1)
seg_net.load_state_dict(torch.load("u2net_human_seg.pth"))
seg_net.cuda()
seg_net.eval()
parser = argparse.ArgumentParser(description="Training")
parser.add_argument("--config", default="./config/face.yaml")
parser.add_argument("--name", default=None)
parser.add_argument(
"--checkpoints_dir", default="result", help="Dir for saving logs and models."
)
parser.add_argument("--seed", type=int, default=0, help="Random seed.")
parser.add_argument("--cross_id", action="store_true")
parser.add_argument("--which_iter", type=int, default=None)
parser.add_argument("--no_resume", action="store_true")
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument("--single_gpu", action="store_true")
parser.add_argument("--input", required=True, type=str)
parser.add_argument("--output_dir", type=str)
args = parser.parse_args()
def normPRED(d):
ma = torch.max(d)
mi = torch.min(d)
dn = (d - mi) / (ma - mi)
return dn
def seg_inference(net, input):
input = input.clone()
input[:, 2] = (input[:, 2] - 0.406) / 0.225
input[:, 1] = (input[:, 1] - 0.456) / 0.224
input[:, 0] = (input[:, 0] - 0.485) / 0.229
input = torch.nn.functional.interpolate(input, size=(320, 320), mode="bicubic")
input2 = input.flip(3) # N, C, H, W
# inference
d1, d2, d3, d4, d5, d6, d7 = net(input)
d11, d2, d3, d4, d5, d6, d7 = net(input2)
d1 = (d1 + d11.flip(3)) / 2
d1 = torch.nn.functional.interpolate(d1, size=(256, 256), mode="bicubic")
# normalization
pred = 1.0 - d1[:, 0, :, :]
pred = normPRED(pred)
# convert torch tensor to numpy array
pred = pred.squeeze()
pred = pred.cpu().data.numpy()
del d1, d11, d2, d3, d4, d5, d6, d7
return pred
def transfera2b(a, b): # ref1 ref2 output
with torch.no_grad():
mb = torch.tensor(seg_inference(seg_net, b)).unsqueeze(0).unsqueeze(0).cuda()
ma = torch.tensor(seg_inference(seg_net, a)).unsqueeze(0).unsqueeze(0).cuda()
ma = torchvision.transforms.GaussianBlur(kernel_size=(7, 7), sigma=8.0)(ma)
ma = (ma > 0.8).float()
mb = (mb > 0.4).float()
m = ma * mb
return m, ma, mb
def write2video(results_dir, *video_list):
cat_video = None
for video in video_list:
video_numpy = video[:, :3, :, :].cpu().float().detach().numpy()
video_numpy = (np.transpose(video_numpy, (0, 2, 3, 1)) + 1) / 2.0 * 255.0
video_numpy = video_numpy.astype(np.uint8)
cat_video = (
np.concatenate([cat_video, video_numpy], 2)
if cat_video is not None
else video_numpy
)
gen_images = cat_video[:, :, 256:, ::-1] # N, h, w, c
gt_images = cat_video[:, :, :256, ::-1]
out_name = results_dir + ".mp4"
_, height, width, layers = gen_images.shape
size = (width, height)
out = cv2.VideoWriter(out_name, cv2.VideoWriter_fourcc(*"mp4v"), 30, size)
for i in range(gen_images.shape[0]):
frame = gen_images[i]
out.write(frame)
out.release()
if __name__ == "__main__":
set_random_seed(args.seed)
opt = Config(args.config, args, is_train=False)
opt.data.path = args.input
if not args.single_gpu:
opt.local_rank = args.local_rank
init_dist(opt.local_rank)
opt.device = torch.cuda.current_device()
# create a visualizer
date_uid, logdir = init_logging(opt)
opt.logdir = logdir
make_logging_dir(logdir, date_uid)
# create a model
net_G, net_G_ema, opt_G, sch_G = get_model_optimizer_and_scheduler(opt)
trainer = get_trainer(opt, net_G, net_G_ema, opt_G, sch_G, None)
current_epoch, current_iteration = trainer.load_checkpoint(opt, args.which_iter)
net_G = trainer.net_G_ema.eval()
output_dir = os.path.join(
args.output_dir,
"epoch_{:05}_iteration_{:09}".format(current_epoch, current_iteration),
)
os.makedirs(output_dir, exist_ok=True)
opt.data.cross_id = args.cross_id
dataset = VoxVideoDataset(opt.data, is_inference=True)
with torch.no_grad():
for video_index in tqdm(range(dataset.__len__())):
data = dataset.load_next_video()
input_source1 = data["source_image"][None].cuda()
name = data["video_name"]
loop_length = len(data["target_semantics"])
output_images, gt_images, warp_images = [], [], []
pred_list = []
mask_list = []
for frame_index in tqdm(range(loop_length)):
target_semantic = data["target_semantics"][frame_index][None].cuda()
down = input_source1.min()
up = input_source1.max()
input_source1 = (input_source1 - down) / (
up - down
) * 2 - 1 # =>[-1, 1]
a = (input_source1.cuda() + 1) / 2
output_dict1 = net_G(input_source1, target_semantic)
output_dict1["fake_image"] = (
output_dict1["fake_image"].clamp_(-1, 1) + 1
) / 2 * (up - down) + down
b = output_dict1["fake_image"]
b = (b.clamp_(-1, 1) + 1) / 2
m, ma, mb = transfera2b(a, b)
if frame_index == 0:
m = ma
output_dict1["fake_image"] = input_source1.cuda()
mask_list.append(m)
m = torch.median(torch.cat(mask_list[-5:]), 0)[0]
m = torchvision.transforms.GaussianBlur(kernel_size=(5, 5), sigma=8.0)(m)
output_dict1["fake_image"] = (
output_dict1["fake_image"] * (1 - m) + input_source1 * m
)
pred_list.append(output_dict1["fake_image"].cpu().clamp_(-1, 1))
if frame_index == 0:
mean = input_source1
else:
mean = torch.mean(torch.cat(pred_list[-3:], 0), 0).cuda()
output_images.append(output_dict1["fake_image"].cpu().clamp_(-1, 1))
warp_images.append(output_dict1["warp_image"].cpu().clamp_(-1, 1))
gt_images.append(data["target_image"][frame_index][None])
gen_images = torch.cat(output_images, 0)
gt_images = torch.cat(gt_images, 0)
warp_images = torch.cat(warp_images, 0)
write2video(
"{}/{}".format(output_dir, name), gt_images, warp_images, gen_images
)
print("write results to video {}/{}".format(output_dir, name))