Spaces:
Sleeping
Sleeping
wissemkarous
commited on
Commit
β’
8e606bb
1
Parent(s):
8c79f36
utils
Browse files- demo.py +242 -0
- two_stream_infer.py +38 -0
demo.py
ADDED
@@ -0,0 +1,242 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import os
|
3 |
+
from dataset import MyDataset
|
4 |
+
import numpy as np
|
5 |
+
import cv2
|
6 |
+
import face_alignment
|
7 |
+
import streamlit as st
|
8 |
+
|
9 |
+
|
10 |
+
def get_position(size, padding=0.25):
|
11 |
+
x = [
|
12 |
+
0.000213256,
|
13 |
+
0.0752622,
|
14 |
+
0.18113,
|
15 |
+
0.29077,
|
16 |
+
0.393397,
|
17 |
+
0.586856,
|
18 |
+
0.689483,
|
19 |
+
0.799124,
|
20 |
+
0.904991,
|
21 |
+
0.98004,
|
22 |
+
0.490127,
|
23 |
+
0.490127,
|
24 |
+
0.490127,
|
25 |
+
0.490127,
|
26 |
+
0.36688,
|
27 |
+
0.426036,
|
28 |
+
0.490127,
|
29 |
+
0.554217,
|
30 |
+
0.613373,
|
31 |
+
0.121737,
|
32 |
+
0.187122,
|
33 |
+
0.265825,
|
34 |
+
0.334606,
|
35 |
+
0.260918,
|
36 |
+
0.182743,
|
37 |
+
0.645647,
|
38 |
+
0.714428,
|
39 |
+
0.793132,
|
40 |
+
0.858516,
|
41 |
+
0.79751,
|
42 |
+
0.719335,
|
43 |
+
0.254149,
|
44 |
+
0.340985,
|
45 |
+
0.428858,
|
46 |
+
0.490127,
|
47 |
+
0.551395,
|
48 |
+
0.639268,
|
49 |
+
0.726104,
|
50 |
+
0.642159,
|
51 |
+
0.556721,
|
52 |
+
0.490127,
|
53 |
+
0.423532,
|
54 |
+
0.338094,
|
55 |
+
0.290379,
|
56 |
+
0.428096,
|
57 |
+
0.490127,
|
58 |
+
0.552157,
|
59 |
+
0.689874,
|
60 |
+
0.553364,
|
61 |
+
0.490127,
|
62 |
+
0.42689,
|
63 |
+
]
|
64 |
+
|
65 |
+
y = [
|
66 |
+
0.106454,
|
67 |
+
0.038915,
|
68 |
+
0.0187482,
|
69 |
+
0.0344891,
|
70 |
+
0.0773906,
|
71 |
+
0.0773906,
|
72 |
+
0.0344891,
|
73 |
+
0.0187482,
|
74 |
+
0.038915,
|
75 |
+
0.106454,
|
76 |
+
0.203352,
|
77 |
+
0.307009,
|
78 |
+
0.409805,
|
79 |
+
0.515625,
|
80 |
+
0.587326,
|
81 |
+
0.609345,
|
82 |
+
0.628106,
|
83 |
+
0.609345,
|
84 |
+
0.587326,
|
85 |
+
0.216423,
|
86 |
+
0.178758,
|
87 |
+
0.179852,
|
88 |
+
0.231733,
|
89 |
+
0.245099,
|
90 |
+
0.244077,
|
91 |
+
0.231733,
|
92 |
+
0.179852,
|
93 |
+
0.178758,
|
94 |
+
0.216423,
|
95 |
+
0.244077,
|
96 |
+
0.245099,
|
97 |
+
0.780233,
|
98 |
+
0.745405,
|
99 |
+
0.727388,
|
100 |
+
0.742578,
|
101 |
+
0.727388,
|
102 |
+
0.745405,
|
103 |
+
0.780233,
|
104 |
+
0.864805,
|
105 |
+
0.902192,
|
106 |
+
0.909281,
|
107 |
+
0.902192,
|
108 |
+
0.864805,
|
109 |
+
0.784792,
|
110 |
+
0.778746,
|
111 |
+
0.785343,
|
112 |
+
0.778746,
|
113 |
+
0.784792,
|
114 |
+
0.824182,
|
115 |
+
0.831803,
|
116 |
+
0.824182,
|
117 |
+
]
|
118 |
+
|
119 |
+
x, y = np.array(x), np.array(y)
|
120 |
+
|
121 |
+
x = (x + padding) / (2 * padding + 1)
|
122 |
+
y = (y + padding) / (2 * padding + 1)
|
123 |
+
x = x * size
|
124 |
+
y = y * size
|
125 |
+
return np.array(list(zip(x, y)))
|
126 |
+
|
127 |
+
|
128 |
+
def output_video(p, txt, output_path):
|
129 |
+
files = os.listdir(p)
|
130 |
+
files = sorted(files, key=lambda x: int(os.path.splitext(x)[0]))
|
131 |
+
|
132 |
+
font = cv2.FONT_HERSHEY_SIMPLEX
|
133 |
+
|
134 |
+
for file, line in zip(files, txt):
|
135 |
+
img = cv2.imread(os.path.join(p, file))
|
136 |
+
h, w, _ = img.shape
|
137 |
+
img = cv2.putText(
|
138 |
+
img, line, (w // 8, 11 * h // 12), font, 1.2, (0, 0, 0), 3, cv2.LINE_AA
|
139 |
+
)
|
140 |
+
img = cv2.putText(
|
141 |
+
img,
|
142 |
+
line,
|
143 |
+
(w // 8, 11 * h // 12),
|
144 |
+
font,
|
145 |
+
1.2,
|
146 |
+
(255, 255, 255),
|
147 |
+
0,
|
148 |
+
cv2.LINE_AA,
|
149 |
+
)
|
150 |
+
h = h // 2
|
151 |
+
w = w // 2
|
152 |
+
img = cv2.resize(img, (w, h))
|
153 |
+
cv2.imwrite(os.path.join(p, file), img)
|
154 |
+
|
155 |
+
# create the output_videos directory if it doesn't exist
|
156 |
+
if not os.path.exists(output_path):
|
157 |
+
os.makedirs(output_path)
|
158 |
+
|
159 |
+
output = os.path.join(output_path, "output.mp4")
|
160 |
+
cmd = "ffmpeg -hide_banner -loglevel error -y -i {}/%04d.jpg -r 25 {}".format(
|
161 |
+
p, output
|
162 |
+
)
|
163 |
+
os.system(cmd)
|
164 |
+
|
165 |
+
|
166 |
+
def transformation_from_points(points1, points2):
|
167 |
+
points1 = points1.astype(np.float64)
|
168 |
+
points2 = points2.astype(np.float64)
|
169 |
+
|
170 |
+
c1 = np.mean(points1, axis=0)
|
171 |
+
c2 = np.mean(points2, axis=0)
|
172 |
+
points1 -= c1
|
173 |
+
points2 -= c2
|
174 |
+
s1 = np.std(points1)
|
175 |
+
s2 = np.std(points2)
|
176 |
+
points1 /= s1
|
177 |
+
points2 /= s2
|
178 |
+
|
179 |
+
U, S, Vt = np.linalg.svd(points1.T * points2)
|
180 |
+
R = (U * Vt).T
|
181 |
+
return np.vstack(
|
182 |
+
[
|
183 |
+
np.hstack(((s2 / s1) * R, c2.T - (s2 / s1) * R * c1.T)),
|
184 |
+
np.matrix([0.0, 0.0, 1.0]),
|
185 |
+
]
|
186 |
+
)
|
187 |
+
|
188 |
+
|
189 |
+
@st.cache_data(show_spinner=False, persist=True)
|
190 |
+
def load_video(file, device: str):
|
191 |
+
video_name = file.split(".")[0]
|
192 |
+
# create the samples directory if it doesn't exist
|
193 |
+
if not os.path.exists(f"{video_name}_samples"):
|
194 |
+
os.makedirs(f"{video_name}_samples")
|
195 |
+
|
196 |
+
p = os.path.join(f"{video_name}_samples")
|
197 |
+
output = os.path.join(f"{video_name}_samples", "%04d.jpg")
|
198 |
+
cmd = "ffmpeg -hide_banner -loglevel error -i {} -qscale:v 2 -r 25 {}".format(
|
199 |
+
file, output
|
200 |
+
)
|
201 |
+
os.system(cmd)
|
202 |
+
|
203 |
+
files = os.listdir(p)
|
204 |
+
files = sorted(files, key=lambda x: int(os.path.splitext(x)[0]))
|
205 |
+
|
206 |
+
array = [cv2.imread(os.path.join(p, file)) for file in files]
|
207 |
+
|
208 |
+
array = list(filter(lambda im: not im is None, array))
|
209 |
+
|
210 |
+
fa = face_alignment.FaceAlignment(
|
211 |
+
face_alignment.LandmarksType._2D, flip_input=False, device=device
|
212 |
+
)
|
213 |
+
points = [fa.get_landmarks(I) for I in array]
|
214 |
+
|
215 |
+
front256 = get_position(256)
|
216 |
+
video = []
|
217 |
+
for point, scene in zip(points, array):
|
218 |
+
if point is not None:
|
219 |
+
shape = np.array(point[0])
|
220 |
+
shape = shape[17:]
|
221 |
+
M = transformation_from_points(np.matrix(shape), np.matrix(front256))
|
222 |
+
|
223 |
+
img = cv2.warpAffine(scene, M[:2], (256, 256))
|
224 |
+
(x, y) = front256[-20:].mean(0).astype(np.int32)
|
225 |
+
w = 160 // 2
|
226 |
+
img = img[y - w // 2 : y + w // 2, x - w : x + w, ...]
|
227 |
+
img = cv2.resize(img, (128, 64))
|
228 |
+
video.append(img)
|
229 |
+
|
230 |
+
video = np.stack(video, axis=0).astype(np.float32)
|
231 |
+
video = torch.FloatTensor(video.transpose(3, 0, 1, 2)) / 255.0
|
232 |
+
|
233 |
+
return video, p, files
|
234 |
+
|
235 |
+
|
236 |
+
def ctc_decode(y):
|
237 |
+
y = y.argmax(-1)
|
238 |
+
t = y.size(0)
|
239 |
+
result = []
|
240 |
+
for i in range(t + 1):
|
241 |
+
result.append(MyDataset.ctc_arr2txt(y[:i], start=1))
|
242 |
+
return result
|
two_stream_infer.py
ADDED
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from models.two_stream_lipnet import TwoStreamLipNet
|
2 |
+
import options as opt
|
3 |
+
import os
|
4 |
+
import torch
|
5 |
+
import streamlit as st
|
6 |
+
|
7 |
+
os.environ["CUDA_VISIBLE_DEVICES"] = opt.gpu
|
8 |
+
|
9 |
+
|
10 |
+
@st.cache_resource
|
11 |
+
def load_model():
|
12 |
+
model = TwoStreamLipNet()
|
13 |
+
model = model.to(opt.device)
|
14 |
+
|
15 |
+
# load the pretrained weights
|
16 |
+
if hasattr(opt, "two_stream_weights"):
|
17 |
+
pretrained_dict = torch.load(
|
18 |
+
opt.two_stream_weights, map_location=torch.device(opt.device)
|
19 |
+
)
|
20 |
+
model_dict = model.state_dict()
|
21 |
+
pretrained_dict = {
|
22 |
+
k: v
|
23 |
+
for k, v in pretrained_dict.items()
|
24 |
+
if k in model_dict.keys() and v.size() == model_dict[k].size()
|
25 |
+
}
|
26 |
+
missed_params = [
|
27 |
+
k for k, v in model_dict.items() if not k in pretrained_dict.keys()
|
28 |
+
]
|
29 |
+
print(
|
30 |
+
"loaded params/tot params:{}/{}".format(
|
31 |
+
len(pretrained_dict), len(model_dict)
|
32 |
+
)
|
33 |
+
)
|
34 |
+
print("miss matched params:{}".format(missed_params))
|
35 |
+
model_dict.update(pretrained_dict)
|
36 |
+
model.load_state_dict(model_dict)
|
37 |
+
|
38 |
+
return model
|