Non-playing-Character
commited on
Update inference.py
Browse files- inference.py +359 -359
inference.py
CHANGED
@@ -1,359 +1,359 @@
|
|
1 |
-
from os import listdir, path
|
2 |
-
import numpy as np
|
3 |
-
import scipy, cv2, os, sys, argparse, audio
|
4 |
-
import json, subprocess, random, string
|
5 |
-
from tqdm import tqdm
|
6 |
-
from glob import glob
|
7 |
-
import torch, face_detection
|
8 |
-
from wav2lip_models import Wav2Lip
|
9 |
-
import platform
|
10 |
-
from face_parsing import init_parser, swap_regions
|
11 |
-
from basicsr.apply_sr import init_sr_model, enhance
|
12 |
-
|
13 |
-
parser = argparse.ArgumentParser(description='Inference code to lip-sync videos in the wild using Wav2Lip models')
|
14 |
-
|
15 |
-
parser.add_argument('--checkpoint_path', type=str,
|
16 |
-
help='Name of saved checkpoint to load weights from', required=True)
|
17 |
-
|
18 |
-
parser.add_argument('--segmentation_path', type=str,
|
19 |
-
help='Name of saved checkpoint of segmentation network', required=True)
|
20 |
-
|
21 |
-
parser.add_argument('--sr_path', type=str,
|
22 |
-
help='Name of saved checkpoint of super-resolution network', required=True)
|
23 |
-
|
24 |
-
parser.add_argument('--face', type=str,
|
25 |
-
help='Filepath of video/image that contains faces to use', required=True)
|
26 |
-
parser.add_argument('--audio', type=str,
|
27 |
-
help='Filepath of video/audio file to use as raw audio source', required=True)
|
28 |
-
parser.add_argument('--outfile', type=str, help='Video path to save result. See default for an e.g.',
|
29 |
-
default='results/result_voice.mp4')
|
30 |
-
|
31 |
-
|
32 |
-
parser.add_argument('--static', type=bool,
|
33 |
-
help='If True, then use only first video frame for inference', default=False)
|
34 |
-
parser.add_argument('--fps', type=float, help='Can be specified only if input is a static image (default: 25)',
|
35 |
-
default=25., required=False)
|
36 |
-
|
37 |
-
parser.add_argument('--pads', nargs='+', type=int, default=[0, 10, 0, 0],
|
38 |
-
help='Padding (top, bottom, left, right). Please adjust to include chin at least')
|
39 |
-
|
40 |
-
parser.add_argument('--face_det_batch_size', type=int,
|
41 |
-
help='Batch size for face detection', default=16)
|
42 |
-
parser.add_argument('--wav2lip_batch_size', type=int, help='Batch size for Wav2Lip model(s)', default=128)
|
43 |
-
|
44 |
-
parser.add_argument('--resize_factor', default=1, type=int,
|
45 |
-
help='Reduce the resolution by this factor. Sometimes, best results are obtained at 480p or 720p')
|
46 |
-
|
47 |
-
parser.add_argument('--crop', nargs='+', type=int, default=[0, -1, 0, -1],
|
48 |
-
help='Crop video to a smaller region (top, bottom, left, right). Applied after resize_factor and rotate arg. '
|
49 |
-
'Useful if multiple face present. -1 implies the value will be auto-inferred based on height, width')
|
50 |
-
|
51 |
-
parser.add_argument('--box', nargs='+', type=int, default=[-1, -1, -1, -1],
|
52 |
-
help='Specify a constant bounding box for the face. Use only as a last resort if the face is not detected.'
|
53 |
-
'Also, might work only if the face is not moving around much. Syntax: (top, bottom, left, right).')
|
54 |
-
|
55 |
-
parser.add_argument('--rotate', default=False, action='store_true',
|
56 |
-
help='Sometimes videos taken from a phone can be flipped 90deg. If true, will flip video right by 90deg.'
|
57 |
-
'Use if you get a flipped result, despite feeding a normal looking video')
|
58 |
-
|
59 |
-
parser.add_argument('--nosmooth', default=False, action='store_true',
|
60 |
-
help='Prevent smoothing face detections over a short temporal window')
|
61 |
-
parser.add_argument('--no_segmentation', default=False, action='store_true',
|
62 |
-
help='Prevent using face segmentation')
|
63 |
-
parser.add_argument('--no_sr', default=False, action='store_true',
|
64 |
-
help='Prevent using super resolution')
|
65 |
-
|
66 |
-
parser.add_argument('--save_frames', default=
|
67 |
-
help='Save each frame as an image. Use with caution')
|
68 |
-
parser.add_argument('--gt_path', type=str,
|
69 |
-
help='Where to store saved ground truth frames', required=False)
|
70 |
-
parser.add_argument('--pred_path', type=str,
|
71 |
-
help='Where to store frames produced by algorithm', required=False)
|
72 |
-
parser.add_argument('--save_as_video', action="store_true", default=False,
|
73 |
-
help='Whether to save frames as video', required=False)
|
74 |
-
parser.add_argument('--image_prefix', type=str, default="",
|
75 |
-
help='Prefix to save frames with', required=False)
|
76 |
-
|
77 |
-
args = parser.parse_args()
|
78 |
-
args.img_size = 96
|
79 |
-
|
80 |
-
if os.path.isfile(args.face) and args.face.split('.')[1] in ['jpg', 'png', 'jpeg']:
|
81 |
-
args.static = True
|
82 |
-
|
83 |
-
def get_smoothened_boxes(boxes, T):
|
84 |
-
for i in range(len(boxes)):
|
85 |
-
if i + T > len(boxes):
|
86 |
-
window = boxes[len(boxes) - T:]
|
87 |
-
else:
|
88 |
-
window = boxes[i : i + T]
|
89 |
-
boxes[i] = np.mean(window, axis=0)
|
90 |
-
return boxes
|
91 |
-
|
92 |
-
def face_detect(images):
|
93 |
-
detector = face_detection.FaceAlignment(face_detection.LandmarksType._2D,
|
94 |
-
flip_input=False, device=device)
|
95 |
-
|
96 |
-
batch_size = args.face_det_batch_size
|
97 |
-
|
98 |
-
while 1:
|
99 |
-
predictions = []
|
100 |
-
try:
|
101 |
-
for i in range(0, len(images), batch_size):
|
102 |
-
predictions.extend(detector.get_detections_for_batch(np.array(images[i:i + batch_size])))
|
103 |
-
except RuntimeError:
|
104 |
-
if batch_size == 1:
|
105 |
-
raise RuntimeError('Image too big to run face detection on GPU. Please use the --resize_factor argument')
|
106 |
-
batch_size //= 2
|
107 |
-
print('Recovering from OOM error; New batch size: {}'.format(batch_size))
|
108 |
-
continue
|
109 |
-
break
|
110 |
-
|
111 |
-
results = []
|
112 |
-
pady1, pady2, padx1, padx2 = args.pads
|
113 |
-
for rect, image in zip(predictions, images):
|
114 |
-
if rect is None:
|
115 |
-
cv2.imwrite('temp/faulty_frame.jpg', image) # check this frame where the face was not detected.
|
116 |
-
raise ValueError('Face not detected! Ensure the video contains a face in all the frames.')
|
117 |
-
|
118 |
-
y1 = max(0, rect[1] - pady1)
|
119 |
-
y2 = min(image.shape[0], rect[3] + pady2)
|
120 |
-
x1 = max(0, rect[0] - padx1)
|
121 |
-
x2 = min(image.shape[1], rect[2] + padx2)
|
122 |
-
|
123 |
-
results.append([x1, y1, x2, y2])
|
124 |
-
|
125 |
-
boxes = np.array(results)
|
126 |
-
if not args.nosmooth: boxes = get_smoothened_boxes(boxes, T=5)
|
127 |
-
results = [[image[y1: y2, x1:x2], (y1, y2, x1, x2)] for image, (x1, y1, x2, y2) in zip(images, boxes)]
|
128 |
-
|
129 |
-
del detector
|
130 |
-
return results
|
131 |
-
|
132 |
-
def datagen(mels):
|
133 |
-
img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
|
134 |
-
|
135 |
-
"""
|
136 |
-
if args.box[0] == -1:
|
137 |
-
if not args.static:
|
138 |
-
face_det_results = face_detect(frames) # BGR2RGB for CNN face detection
|
139 |
-
else:
|
140 |
-
face_det_results = face_detect([frames[0]])
|
141 |
-
else:
|
142 |
-
print('Using the specified bounding box instead of face detection...')
|
143 |
-
y1, y2, x1, x2 = args.box
|
144 |
-
face_det_results = [[f[y1: y2, x1:x2], (y1, y2, x1, x2)] for f in frames]
|
145 |
-
"""
|
146 |
-
|
147 |
-
reader = read_frames()
|
148 |
-
|
149 |
-
for i, m in enumerate(mels):
|
150 |
-
try:
|
151 |
-
frame_to_save = next(reader)
|
152 |
-
except StopIteration:
|
153 |
-
reader = read_frames()
|
154 |
-
frame_to_save = next(reader)
|
155 |
-
|
156 |
-
face, coords = face_detect([frame_to_save])[0]
|
157 |
-
|
158 |
-
face = cv2.resize(face, (args.img_size, args.img_size))
|
159 |
-
|
160 |
-
img_batch.append(face)
|
161 |
-
mel_batch.append(m)
|
162 |
-
frame_batch.append(frame_to_save)
|
163 |
-
coords_batch.append(coords)
|
164 |
-
|
165 |
-
if len(img_batch) >= args.wav2lip_batch_size:
|
166 |
-
img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
|
167 |
-
|
168 |
-
img_masked = img_batch.copy()
|
169 |
-
img_masked[:, args.img_size//2:] = 0
|
170 |
-
|
171 |
-
img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
|
172 |
-
mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
|
173 |
-
|
174 |
-
yield img_batch, mel_batch, frame_batch, coords_batch
|
175 |
-
img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
|
176 |
-
|
177 |
-
if len(img_batch) > 0:
|
178 |
-
img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
|
179 |
-
|
180 |
-
img_masked = img_batch.copy()
|
181 |
-
img_masked[:, args.img_size//2:] = 0
|
182 |
-
|
183 |
-
img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
|
184 |
-
mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
|
185 |
-
|
186 |
-
yield img_batch, mel_batch, frame_batch, coords_batch
|
187 |
-
|
188 |
-
mel_step_size = 16
|
189 |
-
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
190 |
-
print('Using {} for inference.'.format(device))
|
191 |
-
|
192 |
-
def _load(checkpoint_path):
|
193 |
-
if device == 'cuda':
|
194 |
-
checkpoint = torch.load(checkpoint_path)
|
195 |
-
else:
|
196 |
-
checkpoint = torch.load(checkpoint_path,
|
197 |
-
map_location=lambda storage, loc: storage)
|
198 |
-
return checkpoint
|
199 |
-
|
200 |
-
def load_model(path):
|
201 |
-
model = Wav2Lip()
|
202 |
-
print("Load checkpoint from: {}".format(path))
|
203 |
-
checkpoint = _load(path)
|
204 |
-
s = checkpoint["state_dict"]
|
205 |
-
new_s = {}
|
206 |
-
for k, v in s.items():
|
207 |
-
new_s[k.replace('module.', '')] = v
|
208 |
-
model.load_state_dict(new_s)
|
209 |
-
|
210 |
-
model = model.to(device)
|
211 |
-
return model.eval()
|
212 |
-
|
213 |
-
def read_frames():
|
214 |
-
if args.face.split('.')[1] in ['jpg', 'png', 'jpeg']:
|
215 |
-
face = cv2.imread(args.face)
|
216 |
-
while 1:
|
217 |
-
yield face
|
218 |
-
|
219 |
-
video_stream = cv2.VideoCapture(args.face)
|
220 |
-
fps = video_stream.get(cv2.CAP_PROP_FPS)
|
221 |
-
|
222 |
-
print('Reading video frames from start...')
|
223 |
-
|
224 |
-
while 1:
|
225 |
-
still_reading, frame = video_stream.read()
|
226 |
-
if not still_reading:
|
227 |
-
video_stream.release()
|
228 |
-
break
|
229 |
-
if args.resize_factor > 1:
|
230 |
-
frame = cv2.resize(frame, (frame.shape[1]//args.resize_factor, frame.shape[0]//args.resize_factor))
|
231 |
-
|
232 |
-
if args.rotate:
|
233 |
-
frame = cv2.rotate(frame, cv2.cv2.ROTATE_90_CLOCKWISE)
|
234 |
-
|
235 |
-
y1, y2, x1, x2 = args.crop
|
236 |
-
if x2 == -1: x2 = frame.shape[1]
|
237 |
-
if y2 == -1: y2 = frame.shape[0]
|
238 |
-
|
239 |
-
frame = frame[y1:y2, x1:x2]
|
240 |
-
|
241 |
-
yield frame
|
242 |
-
|
243 |
-
def main():
|
244 |
-
if not os.path.isfile(args.face):
|
245 |
-
raise ValueError('--face argument must be a valid path to video/image file')
|
246 |
-
|
247 |
-
elif args.face.split('.')[1] in ['jpg', 'png', 'jpeg']:
|
248 |
-
fps = args.fps
|
249 |
-
else:
|
250 |
-
video_stream = cv2.VideoCapture(args.face)
|
251 |
-
fps = video_stream.get(cv2.CAP_PROP_FPS)
|
252 |
-
video_stream.release()
|
253 |
-
|
254 |
-
|
255 |
-
if not args.audio.endswith('.wav'):
|
256 |
-
print('Extracting raw audio...')
|
257 |
-
command = 'ffmpeg -y -i {} -strict -2 {}'.format(args.audio, 'temp/temp.wav')
|
258 |
-
|
259 |
-
subprocess.call(command, shell=True)
|
260 |
-
args.audio = 'temp/temp.wav'
|
261 |
-
|
262 |
-
wav = audio.load_wav(args.audio, 16000)
|
263 |
-
mel = audio.melspectrogram(wav)
|
264 |
-
print(mel.shape)
|
265 |
-
|
266 |
-
if np.isnan(mel.reshape(-1)).sum() > 0:
|
267 |
-
raise ValueError('Mel contains nan! Using a TTS voice? Add a small epsilon noise to the wav file and try again')
|
268 |
-
|
269 |
-
mel_chunks = []
|
270 |
-
mel_idx_multiplier = 80./fps
|
271 |
-
i = 0
|
272 |
-
while 1:
|
273 |
-
start_idx = int(i * mel_idx_multiplier)
|
274 |
-
if start_idx + mel_step_size > len(mel[0]):
|
275 |
-
mel_chunks.append(mel[:, len(mel[0]) - mel_step_size:])
|
276 |
-
break
|
277 |
-
mel_chunks.append(mel[:, start_idx : start_idx + mel_step_size])
|
278 |
-
i += 1
|
279 |
-
|
280 |
-
print("Length of mel chunks: {}".format(len(mel_chunks)))
|
281 |
-
|
282 |
-
batch_size = args.wav2lip_batch_size
|
283 |
-
gen = datagen(mel_chunks)
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
if args.save_as_video:
|
288 |
-
gt_out = cv2.VideoWriter("temp/gt.avi", cv2.VideoWriter_fourcc(*'DIVX'), fps, (384, 384))
|
289 |
-
pred_out = cv2.VideoWriter("temp/pred.avi", cv2.VideoWriter_fourcc(*'DIVX'), fps, (96, 96))
|
290 |
-
|
291 |
-
abs_idx = 0
|
292 |
-
for i, (img_batch, mel_batch, frames, coords) in enumerate(tqdm(gen,
|
293 |
-
total=int(np.ceil(float(len(mel_chunks))/batch_size)))):
|
294 |
-
if i == 0:
|
295 |
-
print("Loading segmentation network...")
|
296 |
-
seg_net = init_parser(args.segmentation_path)
|
297 |
-
|
298 |
-
print("Loading super resolution model...")
|
299 |
-
sr_net = init_sr_model(args.sr_path)
|
300 |
-
|
301 |
-
model = load_model(args.checkpoint_path)
|
302 |
-
print ("Model loaded")
|
303 |
-
|
304 |
-
frame_h, frame_w = next(read_frames()).shape[:-1]
|
305 |
-
out = cv2.VideoWriter('temp/result.avi',
|
306 |
-
cv2.VideoWriter_fourcc(*'DIVX'), fps, (frame_w, frame_h))
|
307 |
-
|
308 |
-
img_batch = torch.FloatTensor(np.transpose(img_batch, (0, 3, 1, 2))).to(device)
|
309 |
-
mel_batch = torch.FloatTensor(np.transpose(mel_batch, (0, 3, 1, 2))).to(device)
|
310 |
-
|
311 |
-
with torch.no_grad():
|
312 |
-
pred = model(mel_batch, img_batch)
|
313 |
-
|
314 |
-
pred = pred.cpu().numpy().transpose(0, 2, 3, 1) * 255.
|
315 |
-
|
316 |
-
for p, f, c in zip(pred, frames, coords):
|
317 |
-
y1, y2, x1, x2 = c
|
318 |
-
|
319 |
-
if args.save_frames:
|
320 |
-
print("saving frames or video...")
|
321 |
-
if args.save_as_video:
|
322 |
-
print("videos...")
|
323 |
-
pred_out.write(p.astype(np.uint8))
|
324 |
-
gt_out.write(cv2.resize(f[y1:y2, x1:x2], (384, 384)))
|
325 |
-
else:
|
326 |
-
print("frames...")
|
327 |
-
print(f"{args.
|
328 |
-
cv2.imwrite(f"{args.
|
329 |
-
cv2.imwrite(f"{args.
|
330 |
-
abs_idx += 1
|
331 |
-
|
332 |
-
if not args.no_sr:
|
333 |
-
p = enhance(sr_net, p)
|
334 |
-
p = cv2.resize(p.astype(np.uint8), (x2 - x1, y2 - y1))
|
335 |
-
|
336 |
-
if not args.no_segmentation:
|
337 |
-
p = swap_regions(f[y1:y2, x1:x2], p, seg_net)
|
338 |
-
|
339 |
-
f[y1:y2, x1:x2] = p
|
340 |
-
out.write(f)
|
341 |
-
|
342 |
-
out.release()
|
343 |
-
|
344 |
-
command = 'ffmpeg -y -i {} -i {} -strict -2 -q:v 1 {}'.format(args.audio, 'temp/result.avi', args.outfile)
|
345 |
-
subprocess.call(command, shell=platform.system() != 'Windows')
|
346 |
-
|
347 |
-
if args.save_frames and args.save_as_video:
|
348 |
-
gt_out.release()
|
349 |
-
pred_out.release()
|
350 |
-
|
351 |
-
command = 'ffmpeg -y -i {} -i {} -strict -2 -q:v 1 {}'.format(args.audio, 'temp/gt.avi', args.gt_path)
|
352 |
-
subprocess.call(command, shell=platform.system() != 'Windows')
|
353 |
-
|
354 |
-
command = 'ffmpeg -y -i {} -i {} -strict -2 -q:v 1 {}'.format(args.audio, 'temp/pred.avi', args.pred_path)
|
355 |
-
subprocess.call(command, shell=platform.system() != 'Windows')
|
356 |
-
|
357 |
-
|
358 |
-
if __name__ == '__main__':
|
359 |
-
main()
|
|
|
1 |
+
from os import listdir, path
|
2 |
+
import numpy as np
|
3 |
+
import scipy, cv2, os, sys, argparse, audio
|
4 |
+
import json, subprocess, random, string
|
5 |
+
from tqdm import tqdm
|
6 |
+
from glob import glob
|
7 |
+
import torch, face_detection
|
8 |
+
from wav2lip_models import Wav2Lip
|
9 |
+
import platform
|
10 |
+
from face_parsing import init_parser, swap_regions
|
11 |
+
from basicsr.apply_sr import init_sr_model, enhance
|
12 |
+
|
13 |
+
parser = argparse.ArgumentParser(description='Inference code to lip-sync videos in the wild using Wav2Lip models')
|
14 |
+
|
15 |
+
parser.add_argument('--checkpoint_path', type=str,
|
16 |
+
help='Name of saved checkpoint to load weights from', required=True)
|
17 |
+
|
18 |
+
parser.add_argument('--segmentation_path', type=str,
|
19 |
+
help='Name of saved checkpoint of segmentation network', required=True)
|
20 |
+
|
21 |
+
parser.add_argument('--sr_path', type=str,
|
22 |
+
help='Name of saved checkpoint of super-resolution network', required=True)
|
23 |
+
|
24 |
+
parser.add_argument('--face', type=str,
|
25 |
+
help='Filepath of video/image that contains faces to use', required=True)
|
26 |
+
parser.add_argument('--audio', type=str,
|
27 |
+
help='Filepath of video/audio file to use as raw audio source', required=True)
|
28 |
+
parser.add_argument('--outfile', type=str, help='Video path to save result. See default for an e.g.',
|
29 |
+
default='results/result_voice.mp4')
|
30 |
+
|
31 |
+
|
32 |
+
parser.add_argument('--static', type=bool,
|
33 |
+
help='If True, then use only first video frame for inference', default=False)
|
34 |
+
parser.add_argument('--fps', type=float, help='Can be specified only if input is a static image (default: 25)',
|
35 |
+
default=25., required=False)
|
36 |
+
|
37 |
+
parser.add_argument('--pads', nargs='+', type=int, default=[0, 10, 0, 0],
|
38 |
+
help='Padding (top, bottom, left, right). Please adjust to include chin at least')
|
39 |
+
|
40 |
+
parser.add_argument('--face_det_batch_size', type=int,
|
41 |
+
help='Batch size for face detection', default=16)
|
42 |
+
parser.add_argument('--wav2lip_batch_size', type=int, help='Batch size for Wav2Lip model(s)', default=128)
|
43 |
+
|
44 |
+
parser.add_argument('--resize_factor', default=1, type=int,
|
45 |
+
help='Reduce the resolution by this factor. Sometimes, best results are obtained at 480p or 720p')
|
46 |
+
|
47 |
+
parser.add_argument('--crop', nargs='+', type=int, default=[0, -1, 0, -1],
|
48 |
+
help='Crop video to a smaller region (top, bottom, left, right). Applied after resize_factor and rotate arg. '
|
49 |
+
'Useful if multiple face present. -1 implies the value will be auto-inferred based on height, width')
|
50 |
+
|
51 |
+
parser.add_argument('--box', nargs='+', type=int, default=[-1, -1, -1, -1],
|
52 |
+
help='Specify a constant bounding box for the face. Use only as a last resort if the face is not detected.'
|
53 |
+
'Also, might work only if the face is not moving around much. Syntax: (top, bottom, left, right).')
|
54 |
+
|
55 |
+
parser.add_argument('--rotate', default=False, action='store_true',
|
56 |
+
help='Sometimes videos taken from a phone can be flipped 90deg. If true, will flip video right by 90deg.'
|
57 |
+
'Use if you get a flipped result, despite feeding a normal looking video')
|
58 |
+
|
59 |
+
parser.add_argument('--nosmooth', default=False, action='store_true',
|
60 |
+
help='Prevent smoothing face detections over a short temporal window')
|
61 |
+
parser.add_argument('--no_segmentation', default=False, action='store_true',
|
62 |
+
help='Prevent using face segmentation')
|
63 |
+
parser.add_argument('--no_sr', default=False, action='store_true',
|
64 |
+
help='Prevent using super resolution')
|
65 |
+
|
66 |
+
parser.add_argument('--save_frames', default=True, action='store_true',
|
67 |
+
help='Save each frame as an image. Use with caution')
|
68 |
+
parser.add_argument('--gt_path', type=str,
|
69 |
+
help='Where to store saved ground truth frames', required=False)
|
70 |
+
parser.add_argument('--pred_path', type=str,
|
71 |
+
help='Where to store frames produced by algorithm', required=False)
|
72 |
+
parser.add_argument('--save_as_video', action="store_true", default=False,
|
73 |
+
help='Whether to save frames as video', required=False)
|
74 |
+
parser.add_argument('--image_prefix', type=str, default="",
|
75 |
+
help='Prefix to save frames with', required=False)
|
76 |
+
|
77 |
+
args = parser.parse_args()
|
78 |
+
args.img_size = 96
|
79 |
+
|
80 |
+
if os.path.isfile(args.face) and args.face.split('.')[1] in ['jpg', 'png', 'jpeg']:
|
81 |
+
args.static = True
|
82 |
+
|
83 |
+
def get_smoothened_boxes(boxes, T):
|
84 |
+
for i in range(len(boxes)):
|
85 |
+
if i + T > len(boxes):
|
86 |
+
window = boxes[len(boxes) - T:]
|
87 |
+
else:
|
88 |
+
window = boxes[i : i + T]
|
89 |
+
boxes[i] = np.mean(window, axis=0)
|
90 |
+
return boxes
|
91 |
+
|
92 |
+
def face_detect(images):
|
93 |
+
detector = face_detection.FaceAlignment(face_detection.LandmarksType._2D,
|
94 |
+
flip_input=False, device=device)
|
95 |
+
|
96 |
+
batch_size = args.face_det_batch_size
|
97 |
+
|
98 |
+
while 1:
|
99 |
+
predictions = []
|
100 |
+
try:
|
101 |
+
for i in range(0, len(images), batch_size):
|
102 |
+
predictions.extend(detector.get_detections_for_batch(np.array(images[i:i + batch_size])))
|
103 |
+
except RuntimeError:
|
104 |
+
if batch_size == 1:
|
105 |
+
raise RuntimeError('Image too big to run face detection on GPU. Please use the --resize_factor argument')
|
106 |
+
batch_size //= 2
|
107 |
+
print('Recovering from OOM error; New batch size: {}'.format(batch_size))
|
108 |
+
continue
|
109 |
+
break
|
110 |
+
|
111 |
+
results = []
|
112 |
+
pady1, pady2, padx1, padx2 = args.pads
|
113 |
+
for rect, image in zip(predictions, images):
|
114 |
+
if rect is None:
|
115 |
+
cv2.imwrite('temp/faulty_frame.jpg', image) # check this frame where the face was not detected.
|
116 |
+
raise ValueError('Face not detected! Ensure the video contains a face in all the frames.')
|
117 |
+
|
118 |
+
y1 = max(0, rect[1] - pady1)
|
119 |
+
y2 = min(image.shape[0], rect[3] + pady2)
|
120 |
+
x1 = max(0, rect[0] - padx1)
|
121 |
+
x2 = min(image.shape[1], rect[2] + padx2)
|
122 |
+
|
123 |
+
results.append([x1, y1, x2, y2])
|
124 |
+
|
125 |
+
boxes = np.array(results)
|
126 |
+
if not args.nosmooth: boxes = get_smoothened_boxes(boxes, T=5)
|
127 |
+
results = [[image[y1: y2, x1:x2], (y1, y2, x1, x2)] for image, (x1, y1, x2, y2) in zip(images, boxes)]
|
128 |
+
|
129 |
+
del detector
|
130 |
+
return results
|
131 |
+
|
132 |
+
def datagen(mels):
|
133 |
+
img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
|
134 |
+
|
135 |
+
"""
|
136 |
+
if args.box[0] == -1:
|
137 |
+
if not args.static:
|
138 |
+
face_det_results = face_detect(frames) # BGR2RGB for CNN face detection
|
139 |
+
else:
|
140 |
+
face_det_results = face_detect([frames[0]])
|
141 |
+
else:
|
142 |
+
print('Using the specified bounding box instead of face detection...')
|
143 |
+
y1, y2, x1, x2 = args.box
|
144 |
+
face_det_results = [[f[y1: y2, x1:x2], (y1, y2, x1, x2)] for f in frames]
|
145 |
+
"""
|
146 |
+
|
147 |
+
reader = read_frames()
|
148 |
+
|
149 |
+
for i, m in enumerate(mels):
|
150 |
+
try:
|
151 |
+
frame_to_save = next(reader)
|
152 |
+
except StopIteration:
|
153 |
+
reader = read_frames()
|
154 |
+
frame_to_save = next(reader)
|
155 |
+
|
156 |
+
face, coords = face_detect([frame_to_save])[0]
|
157 |
+
|
158 |
+
face = cv2.resize(face, (args.img_size, args.img_size))
|
159 |
+
|
160 |
+
img_batch.append(face)
|
161 |
+
mel_batch.append(m)
|
162 |
+
frame_batch.append(frame_to_save)
|
163 |
+
coords_batch.append(coords)
|
164 |
+
|
165 |
+
if len(img_batch) >= args.wav2lip_batch_size:
|
166 |
+
img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
|
167 |
+
|
168 |
+
img_masked = img_batch.copy()
|
169 |
+
img_masked[:, args.img_size//2:] = 0
|
170 |
+
|
171 |
+
img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
|
172 |
+
mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
|
173 |
+
|
174 |
+
yield img_batch, mel_batch, frame_batch, coords_batch
|
175 |
+
img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
|
176 |
+
|
177 |
+
if len(img_batch) > 0:
|
178 |
+
img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
|
179 |
+
|
180 |
+
img_masked = img_batch.copy()
|
181 |
+
img_masked[:, args.img_size//2:] = 0
|
182 |
+
|
183 |
+
img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
|
184 |
+
mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
|
185 |
+
|
186 |
+
yield img_batch, mel_batch, frame_batch, coords_batch
|
187 |
+
|
188 |
+
mel_step_size = 16
|
189 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
190 |
+
print('Using {} for inference.'.format(device))
|
191 |
+
|
192 |
+
def _load(checkpoint_path):
|
193 |
+
if device == 'cuda':
|
194 |
+
checkpoint = torch.load(checkpoint_path)
|
195 |
+
else:
|
196 |
+
checkpoint = torch.load(checkpoint_path,
|
197 |
+
map_location=lambda storage, loc: storage)
|
198 |
+
return checkpoint
|
199 |
+
|
200 |
+
def load_model(path):
|
201 |
+
model = Wav2Lip()
|
202 |
+
print("Load checkpoint from: {}".format(path))
|
203 |
+
checkpoint = _load(path)
|
204 |
+
s = checkpoint["state_dict"]
|
205 |
+
new_s = {}
|
206 |
+
for k, v in s.items():
|
207 |
+
new_s[k.replace('module.', '')] = v
|
208 |
+
model.load_state_dict(new_s)
|
209 |
+
|
210 |
+
model = model.to(device)
|
211 |
+
return model.eval()
|
212 |
+
|
213 |
+
def read_frames():
|
214 |
+
if args.face.split('.')[1] in ['jpg', 'png', 'jpeg']:
|
215 |
+
face = cv2.imread(args.face)
|
216 |
+
while 1:
|
217 |
+
yield face
|
218 |
+
|
219 |
+
video_stream = cv2.VideoCapture(args.face)
|
220 |
+
fps = video_stream.get(cv2.CAP_PROP_FPS)
|
221 |
+
|
222 |
+
print('Reading video frames from start...')
|
223 |
+
|
224 |
+
while 1:
|
225 |
+
still_reading, frame = video_stream.read()
|
226 |
+
if not still_reading:
|
227 |
+
video_stream.release()
|
228 |
+
break
|
229 |
+
if args.resize_factor > 1:
|
230 |
+
frame = cv2.resize(frame, (frame.shape[1]//args.resize_factor, frame.shape[0]//args.resize_factor))
|
231 |
+
|
232 |
+
if args.rotate:
|
233 |
+
frame = cv2.rotate(frame, cv2.cv2.ROTATE_90_CLOCKWISE)
|
234 |
+
|
235 |
+
y1, y2, x1, x2 = args.crop
|
236 |
+
if x2 == -1: x2 = frame.shape[1]
|
237 |
+
if y2 == -1: y2 = frame.shape[0]
|
238 |
+
|
239 |
+
frame = frame[y1:y2, x1:x2]
|
240 |
+
|
241 |
+
yield frame
|
242 |
+
|
243 |
+
def main():
|
244 |
+
if not os.path.isfile(args.face):
|
245 |
+
raise ValueError('--face argument must be a valid path to video/image file')
|
246 |
+
|
247 |
+
elif args.face.split('.')[1] in ['jpg', 'png', 'jpeg']:
|
248 |
+
fps = args.fps
|
249 |
+
else:
|
250 |
+
video_stream = cv2.VideoCapture(args.face)
|
251 |
+
fps = video_stream.get(cv2.CAP_PROP_FPS)
|
252 |
+
video_stream.release()
|
253 |
+
|
254 |
+
|
255 |
+
if not args.audio.endswith('.wav'):
|
256 |
+
print('Extracting raw audio...')
|
257 |
+
command = 'ffmpeg -y -i {} -strict -2 {}'.format(args.audio, 'temp/temp.wav')
|
258 |
+
|
259 |
+
subprocess.call(command, shell=True)
|
260 |
+
args.audio = 'temp/temp.wav'
|
261 |
+
|
262 |
+
wav = audio.load_wav(args.audio, 16000)
|
263 |
+
mel = audio.melspectrogram(wav)
|
264 |
+
print(mel.shape)
|
265 |
+
|
266 |
+
if np.isnan(mel.reshape(-1)).sum() > 0:
|
267 |
+
raise ValueError('Mel contains nan! Using a TTS voice? Add a small epsilon noise to the wav file and try again')
|
268 |
+
|
269 |
+
mel_chunks = []
|
270 |
+
mel_idx_multiplier = 80./fps
|
271 |
+
i = 0
|
272 |
+
while 1:
|
273 |
+
start_idx = int(i * mel_idx_multiplier)
|
274 |
+
if start_idx + mel_step_size > len(mel[0]):
|
275 |
+
mel_chunks.append(mel[:, len(mel[0]) - mel_step_size:])
|
276 |
+
break
|
277 |
+
mel_chunks.append(mel[:, start_idx : start_idx + mel_step_size])
|
278 |
+
i += 1
|
279 |
+
|
280 |
+
print("Length of mel chunks: {}".format(len(mel_chunks)))
|
281 |
+
|
282 |
+
batch_size = args.wav2lip_batch_size
|
283 |
+
gen = datagen(mel_chunks)
|
284 |
+
|
285 |
+
|
286 |
+
|
287 |
+
if args.save_as_video:
|
288 |
+
gt_out = cv2.VideoWriter("temp/gt.avi", cv2.VideoWriter_fourcc(*'DIVX'), fps, (384, 384))
|
289 |
+
pred_out = cv2.VideoWriter("temp/pred.avi", cv2.VideoWriter_fourcc(*'DIVX'), fps, (96, 96))
|
290 |
+
|
291 |
+
abs_idx = 0
|
292 |
+
for i, (img_batch, mel_batch, frames, coords) in enumerate(tqdm(gen,
|
293 |
+
total=int(np.ceil(float(len(mel_chunks))/batch_size)))):
|
294 |
+
if i == 0:
|
295 |
+
print("Loading segmentation network...")
|
296 |
+
seg_net = init_parser(args.segmentation_path)
|
297 |
+
|
298 |
+
print("Loading super resolution model...")
|
299 |
+
sr_net = init_sr_model(args.sr_path)
|
300 |
+
|
301 |
+
model = load_model(args.checkpoint_path)
|
302 |
+
print ("Model loaded")
|
303 |
+
|
304 |
+
frame_h, frame_w = next(read_frames()).shape[:-1]
|
305 |
+
out = cv2.VideoWriter('temp/result.avi',
|
306 |
+
cv2.VideoWriter_fourcc(*'DIVX'), fps, (frame_w, frame_h))
|
307 |
+
|
308 |
+
img_batch = torch.FloatTensor(np.transpose(img_batch, (0, 3, 1, 2))).to(device)
|
309 |
+
mel_batch = torch.FloatTensor(np.transpose(mel_batch, (0, 3, 1, 2))).to(device)
|
310 |
+
|
311 |
+
with torch.no_grad():
|
312 |
+
pred = model(mel_batch, img_batch)
|
313 |
+
|
314 |
+
pred = pred.cpu().numpy().transpose(0, 2, 3, 1) * 255.
|
315 |
+
|
316 |
+
for p, f, c in zip(pred, frames, coords):
|
317 |
+
y1, y2, x1, x2 = c
|
318 |
+
|
319 |
+
if args.save_frames:
|
320 |
+
print("saving frames or video...")
|
321 |
+
if args.save_as_video:
|
322 |
+
print("videos...")
|
323 |
+
pred_out.write(p.astype(np.uint8))
|
324 |
+
gt_out.write(cv2.resize(f[y1:y2, x1:x2], (384, 384)))
|
325 |
+
else:
|
326 |
+
print("frames...")
|
327 |
+
print(f"{args.pred_path}/{args.image_prefix}{abs_idx}.png")
|
328 |
+
cv2.imwrite(f"{args.pred_path}/{args.image_prefix}{abs_idx:05d}.png", p)
|
329 |
+
cv2.imwrite(f"{args.gt_path}/{args.image_prefix}{abs_idx}.png", f[y1:y2, x1:x2])
|
330 |
+
abs_idx += 1
|
331 |
+
|
332 |
+
if not args.no_sr:
|
333 |
+
p = enhance(sr_net, p)
|
334 |
+
p = cv2.resize(p.astype(np.uint8), (x2 - x1, y2 - y1))
|
335 |
+
|
336 |
+
if not args.no_segmentation:
|
337 |
+
p = swap_regions(f[y1:y2, x1:x2], p, seg_net)
|
338 |
+
|
339 |
+
f[y1:y2, x1:x2] = p
|
340 |
+
out.write(f)
|
341 |
+
|
342 |
+
out.release()
|
343 |
+
|
344 |
+
command = 'ffmpeg -y -i {} -i {} -strict -2 -q:v 1 {}'.format(args.audio, 'temp/result.avi', args.outfile)
|
345 |
+
subprocess.call(command, shell=platform.system() != 'Windows')
|
346 |
+
|
347 |
+
if args.save_frames and args.save_as_video:
|
348 |
+
gt_out.release()
|
349 |
+
pred_out.release()
|
350 |
+
|
351 |
+
command = 'ffmpeg -y -i {} -i {} -strict -2 -q:v 1 {}'.format(args.audio, 'temp/gt.avi', args.gt_path)
|
352 |
+
subprocess.call(command, shell=platform.system() != 'Windows')
|
353 |
+
|
354 |
+
command = 'ffmpeg -y -i {} -i {} -strict -2 -q:v 1 {}'.format(args.audio, 'temp/pred.avi', args.pred_path)
|
355 |
+
subprocess.call(command, shell=platform.system() != 'Windows')
|
356 |
+
|
357 |
+
|
358 |
+
if __name__ == '__main__':
|
359 |
+
main()
|