Spaces:
Running
on
Zero
Running
on
Zero
sindhuhegde
commited on
Commit
•
b63403f
1
Parent(s):
2d17a01
Update app
Browse files
app_v1.py
DELETED
@@ -1,954 +0,0 @@
|
|
1 |
-
import gradio as gr
|
2 |
-
import argparse
|
3 |
-
import os, subprocess
|
4 |
-
from shutil import rmtree
|
5 |
-
|
6 |
-
import numpy as np
|
7 |
-
import cv2
|
8 |
-
import librosa
|
9 |
-
import torch
|
10 |
-
|
11 |
-
from utils.audio_utils import *
|
12 |
-
from utils.inference_utils import *
|
13 |
-
from sync_models.gestsync_models import *
|
14 |
-
|
15 |
-
import sys
|
16 |
-
if sys.version_info > (3, 0): long, unicode, basestring = int, str, str
|
17 |
-
|
18 |
-
from tqdm import tqdm
|
19 |
-
from scipy.io.wavfile import write
|
20 |
-
import mediapipe as mp
|
21 |
-
from protobuf_to_dict import protobuf_to_dict
|
22 |
-
mp_holistic = mp.solutions.holistic
|
23 |
-
from ultralytics import YOLO
|
24 |
-
from decord import VideoReader, cpu
|
25 |
-
|
26 |
-
import warnings
|
27 |
-
warnings.filterwarnings("ignore", category=DeprecationWarning)
|
28 |
-
warnings.filterwarnings("ignore", category=UserWarning)
|
29 |
-
|
30 |
-
# Set the path to checkpoint file
|
31 |
-
CHECKPOINT_PATH = "model_rgb.pth"
|
32 |
-
|
33 |
-
# Initialize global variables
|
34 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
35 |
-
use_cuda = torch.cuda.is_available()
|
36 |
-
n_negative_samples = 100
|
37 |
-
print("Using CUDA: ", use_cuda, device)
|
38 |
-
|
39 |
-
def preprocess_video(path, result_folder, apply_preprocess, padding=20):
|
40 |
-
|
41 |
-
'''
|
42 |
-
This function preprocesses the input video to extract the audio and crop the frames using YOLO model
|
43 |
-
|
44 |
-
Args:
|
45 |
-
- path (string) : Path of the input video file
|
46 |
-
- result_folder (string) : Path of the folder to save the extracted audio and cropped video
|
47 |
-
- padding (int) : Padding to add to the bounding box
|
48 |
-
Returns:
|
49 |
-
- wav_file (string) : Path of the extracted audio file
|
50 |
-
- fps (int) : FPS of the input video
|
51 |
-
- video_output (string) : Path of the cropped video file
|
52 |
-
- msg (string) : Message to be returned
|
53 |
-
'''
|
54 |
-
|
55 |
-
# Load all video frames
|
56 |
-
try:
|
57 |
-
vr = VideoReader(path, ctx=cpu(0))
|
58 |
-
fps = vr.get_avg_fps()
|
59 |
-
frame_count = len(vr)
|
60 |
-
except:
|
61 |
-
msg = "Oops! Could not load the video. Please check the input video and try again."
|
62 |
-
return None, None, None, msg
|
63 |
-
|
64 |
-
if frame_count < 25:
|
65 |
-
msg = "Not enough frames to process! Please give a longer video as input"
|
66 |
-
return None, None, None, msg
|
67 |
-
|
68 |
-
# Extract the audio from the input video file using ffmpeg
|
69 |
-
wav_file = os.path.join(result_folder, "audio.wav")
|
70 |
-
|
71 |
-
status = subprocess.call('ffmpeg -hide_banner -loglevel panic -y -i %s -async 1 -ac 1 -vn \
|
72 |
-
-acodec pcm_s16le -ar 16000 %s -y' % (path, wav_file), shell=True)
|
73 |
-
|
74 |
-
if status != 0:
|
75 |
-
msg = "Oops! Could not load the audio file. Please check the input video and try again."
|
76 |
-
return None, None, None, msg
|
77 |
-
print("Extracted the audio from the video")
|
78 |
-
|
79 |
-
if apply_preprocess=="True":
|
80 |
-
all_frames = []
|
81 |
-
for k in range(len(vr)):
|
82 |
-
all_frames.append(vr[k].asnumpy())
|
83 |
-
all_frames = np.asarray(all_frames)
|
84 |
-
print("Extracted the frames for pre-processing")
|
85 |
-
|
86 |
-
# Load YOLOv9 model (pre-trained on COCO dataset)
|
87 |
-
yolo_model = YOLO("yolov9s.pt")
|
88 |
-
print("Loaded the YOLO model")
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
person_videos = {}
|
93 |
-
person_tracks = {}
|
94 |
-
|
95 |
-
print("Processing the frames...")
|
96 |
-
for frame_idx in tqdm(range(frame_count)):
|
97 |
-
|
98 |
-
frame = all_frames[frame_idx]
|
99 |
-
|
100 |
-
# Perform person detection
|
101 |
-
results = yolo_model(frame, verbose=False)
|
102 |
-
detections = results[0].boxes
|
103 |
-
|
104 |
-
for i, det in enumerate(detections):
|
105 |
-
x1, y1, x2, y2 = det.xyxy[0]
|
106 |
-
cls = det.cls[0]
|
107 |
-
if int(cls) == 0: # Class 0 is 'person' in COCO dataset
|
108 |
-
|
109 |
-
x1 = max(0, int(x1) - padding)
|
110 |
-
y1 = max(0, int(y1) - padding)
|
111 |
-
x2 = min(frame.shape[1], int(x2) + padding)
|
112 |
-
y2 = min(frame.shape[0], int(y2) + padding)
|
113 |
-
|
114 |
-
if i not in person_videos:
|
115 |
-
person_videos[i] = []
|
116 |
-
person_tracks[i] = []
|
117 |
-
|
118 |
-
person_videos[i].append(frame)
|
119 |
-
person_tracks[i].append([x1,y1,x2,y2])
|
120 |
-
|
121 |
-
|
122 |
-
num_persons = 0
|
123 |
-
for i in person_videos.keys():
|
124 |
-
if len(person_videos[i]) >= frame_count//2:
|
125 |
-
num_persons+=1
|
126 |
-
|
127 |
-
if num_persons==0:
|
128 |
-
msg = "No person detected in the video! Please give a video with one person as input"
|
129 |
-
return None, None, None, msg
|
130 |
-
if num_persons>1:
|
131 |
-
msg = "More than one person detected in the video! Please give a video with only one person as input"
|
132 |
-
return None, None, None, msg
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
# For the person detected, crop the frame based on the bounding box
|
137 |
-
if len(person_videos[0]) > frame_count-10:
|
138 |
-
crop_filename = os.path.join(result_folder, "preprocessed_video.avi")
|
139 |
-
fourcc = cv2.VideoWriter_fourcc(*'DIVX')
|
140 |
-
|
141 |
-
# Get bounding box coordinates based on person_tracks[i]
|
142 |
-
max_x1 = min([track[0] for track in person_tracks[0]])
|
143 |
-
max_y1 = min([track[1] for track in person_tracks[0]])
|
144 |
-
max_x2 = max([track[2] for track in person_tracks[0]])
|
145 |
-
max_y2 = max([track[3] for track in person_tracks[0]])
|
146 |
-
|
147 |
-
max_width = max_x2 - max_x1
|
148 |
-
max_height = max_y2 - max_y1
|
149 |
-
|
150 |
-
out = cv2.VideoWriter(crop_filename, fourcc, fps, (max_width, max_height))
|
151 |
-
for frame in person_videos[0]:
|
152 |
-
crop = frame[max_y1:max_y2, max_x1:max_x2]
|
153 |
-
crop = cv2.cvtColor(crop, cv2.COLOR_BGR2RGB)
|
154 |
-
out.write(crop)
|
155 |
-
out.release()
|
156 |
-
|
157 |
-
no_sound_video = crop_filename.split('.')[0] + '_nosound.mp4'
|
158 |
-
status = subprocess.call('ffmpeg -hide_banner -loglevel panic -y -i %s -c copy -an -strict -2 %s' % (crop_filename, no_sound_video), shell=True)
|
159 |
-
if status != 0:
|
160 |
-
msg = "Oops! Could not preprocess the video. Please check the input video and try again."
|
161 |
-
return None, None, None, msg
|
162 |
-
|
163 |
-
video_output = crop_filename.split('.')[0] + '.mp4'
|
164 |
-
status = subprocess.call('ffmpeg -hide_banner -loglevel panic -y -i %s -i %s -strict -2 -q:v 1 %s' %
|
165 |
-
(wav_file , no_sound_video, video_output), shell=True)
|
166 |
-
if status != 0:
|
167 |
-
msg = "Oops! Could not preprocess the video. Please check the input video and try again."
|
168 |
-
return None, None, None, msg
|
169 |
-
|
170 |
-
os.remove(crop_filename)
|
171 |
-
os.remove(no_sound_video)
|
172 |
-
|
173 |
-
print("Successfully saved the pre-processed video: ", video_output)
|
174 |
-
else:
|
175 |
-
msg = "Could not track the person in the full video! Please give a single-speaker video as input"
|
176 |
-
return None, None, None, msg
|
177 |
-
|
178 |
-
else:
|
179 |
-
video_output = path
|
180 |
-
|
181 |
-
return wav_file, fps, video_output, "success"
|
182 |
-
|
183 |
-
def resample_video(video_file, video_fname, result_folder):
|
184 |
-
|
185 |
-
'''
|
186 |
-
This function resamples the video to 25 fps
|
187 |
-
|
188 |
-
Args:
|
189 |
-
- video_file (string) : Path of the input video file
|
190 |
-
- video_fname (string) : Name of the input video file
|
191 |
-
- result_folder (string) : Path of the folder to save the resampled video
|
192 |
-
Returns:
|
193 |
-
- video_file_25fps (string) : Path of the resampled video file
|
194 |
-
'''
|
195 |
-
video_file_25fps = os.path.join(result_folder, '{}.mp4'.format(video_fname))
|
196 |
-
|
197 |
-
# Resample the video to 25 fps
|
198 |
-
command = ("ffmpeg -hide_banner -loglevel panic -y -i {} -q:v 1 -filter:v fps=25 {}".format(video_file, video_file_25fps))
|
199 |
-
from subprocess import call
|
200 |
-
cmd = command.split(' ')
|
201 |
-
print('Resampled the video to 25 fps: {}'.format(video_file_25fps))
|
202 |
-
call(cmd)
|
203 |
-
|
204 |
-
return video_file_25fps
|
205 |
-
|
206 |
-
def load_checkpoint(path, model):
|
207 |
-
'''
|
208 |
-
This function loads the trained model from the checkpoint
|
209 |
-
|
210 |
-
Args:
|
211 |
-
- path (string) : Path of the checkpoint file
|
212 |
-
- model (object) : Model object
|
213 |
-
Returns:
|
214 |
-
- model (object) : Model object with the weights loaded from the checkpoint
|
215 |
-
'''
|
216 |
-
|
217 |
-
# Load the checkpoint
|
218 |
-
if use_cuda:
|
219 |
-
checkpoint = torch.load(path)
|
220 |
-
else:
|
221 |
-
checkpoint = torch.load(path, map_location="cpu")
|
222 |
-
|
223 |
-
s = checkpoint["state_dict"]
|
224 |
-
new_s = {}
|
225 |
-
|
226 |
-
for k, v in s.items():
|
227 |
-
new_s[k.replace('module.', '')] = v
|
228 |
-
model.load_state_dict(new_s)
|
229 |
-
|
230 |
-
if use_cuda:
|
231 |
-
model.cuda()
|
232 |
-
|
233 |
-
print("Loaded checkpoint from: {}".format(path))
|
234 |
-
|
235 |
-
return model.eval()
|
236 |
-
|
237 |
-
|
238 |
-
def load_video_frames(video_file):
|
239 |
-
'''
|
240 |
-
This function extracts the frames from the video
|
241 |
-
|
242 |
-
Args:
|
243 |
-
- video_file (string) : Path of the video file
|
244 |
-
Returns:
|
245 |
-
- frames (list) : List of frames extracted from the video
|
246 |
-
- msg (string) : Message to be returned
|
247 |
-
'''
|
248 |
-
|
249 |
-
# Read the video
|
250 |
-
try:
|
251 |
-
vr = VideoReader(video_file, ctx=cpu(0))
|
252 |
-
except:
|
253 |
-
msg = "Oops! Could not load the input video file"
|
254 |
-
return None, msg
|
255 |
-
|
256 |
-
|
257 |
-
# Extract the frames
|
258 |
-
frames = []
|
259 |
-
for k in range(len(vr)):
|
260 |
-
frames.append(vr[k].asnumpy())
|
261 |
-
|
262 |
-
frames = np.asarray(frames)
|
263 |
-
|
264 |
-
return frames, "success"
|
265 |
-
|
266 |
-
|
267 |
-
|
268 |
-
def get_keypoints(frames):
|
269 |
-
|
270 |
-
'''
|
271 |
-
This function extracts the keypoints from the frames using MediaPipe Holistic pipeline
|
272 |
-
|
273 |
-
Args:
|
274 |
-
- frames (list) : List of frames extracted from the video
|
275 |
-
Returns:
|
276 |
-
- kp_dict (dict) : Dictionary containing the keypoints and the resolution of the frames
|
277 |
-
- msg (string) : Message to be returned
|
278 |
-
'''
|
279 |
-
|
280 |
-
try:
|
281 |
-
holistic = mp_holistic.Holistic(min_detection_confidence=0.5, min_tracking_confidence=0.5)
|
282 |
-
|
283 |
-
resolution = frames[0].shape
|
284 |
-
all_frame_kps = []
|
285 |
-
|
286 |
-
for frame in frames:
|
287 |
-
|
288 |
-
results = holistic.process(frame)
|
289 |
-
|
290 |
-
pose, left_hand, right_hand, face = None, None, None, None
|
291 |
-
if results.pose_landmarks is not None:
|
292 |
-
pose = protobuf_to_dict(results.pose_landmarks)['landmark']
|
293 |
-
if results.left_hand_landmarks is not None:
|
294 |
-
left_hand = protobuf_to_dict(results.left_hand_landmarks)['landmark']
|
295 |
-
if results.right_hand_landmarks is not None:
|
296 |
-
right_hand = protobuf_to_dict(results.right_hand_landmarks)['landmark']
|
297 |
-
if results.face_landmarks is not None:
|
298 |
-
face = protobuf_to_dict(results.face_landmarks)['landmark']
|
299 |
-
|
300 |
-
frame_dict = {"pose":pose, "left_hand":left_hand, "right_hand":right_hand, "face":face}
|
301 |
-
|
302 |
-
all_frame_kps.append(frame_dict)
|
303 |
-
|
304 |
-
kp_dict = {"kps":all_frame_kps, "resolution":resolution}
|
305 |
-
except Exception as e:
|
306 |
-
print("Error: ", e)
|
307 |
-
return None, "Error: Could not extract keypoints from the frames"
|
308 |
-
|
309 |
-
return kp_dict, "success"
|
310 |
-
|
311 |
-
|
312 |
-
def check_visible_gestures(kp_dict):
|
313 |
-
|
314 |
-
'''
|
315 |
-
This function checks if the gestures in the video are visible
|
316 |
-
|
317 |
-
Args:
|
318 |
-
- kp_dict (dict) : Dictionary containing the keypoints and the resolution of the frames
|
319 |
-
Returns:
|
320 |
-
- msg (string) : Message to be returned
|
321 |
-
'''
|
322 |
-
|
323 |
-
keypoints = kp_dict['kps']
|
324 |
-
keypoints = np.array(keypoints)
|
325 |
-
|
326 |
-
if len(keypoints)<25:
|
327 |
-
msg = "Not enough keypoints to process! Please give a longer video as input"
|
328 |
-
return msg
|
329 |
-
|
330 |
-
pose_count, hand_count = 0, 0
|
331 |
-
for frame_kp_dict in keypoints:
|
332 |
-
|
333 |
-
pose = frame_kp_dict["pose"]
|
334 |
-
left_hand = frame_kp_dict["left_hand"]
|
335 |
-
right_hand = frame_kp_dict["right_hand"]
|
336 |
-
|
337 |
-
if pose is None:
|
338 |
-
pose_count += 1
|
339 |
-
|
340 |
-
if left_hand is None and right_hand is None:
|
341 |
-
hand_count += 1
|
342 |
-
|
343 |
-
|
344 |
-
if hand_count/len(keypoints) > 0.7 or pose_count/len(keypoints) > 0.7:
|
345 |
-
msg = "The gestures in the input video are not visible! Please give a video with visible gestures as input."
|
346 |
-
return msg
|
347 |
-
|
348 |
-
print("Successfully verified the input video - Gestures are visible!")
|
349 |
-
|
350 |
-
return "success"
|
351 |
-
|
352 |
-
def load_rgb_masked_frames(input_frames, kp_dict, stride=1, window_frames=25, width=480, height=270):
|
353 |
-
|
354 |
-
'''
|
355 |
-
This function masks the faces using the keypoints extracted from the frames
|
356 |
-
|
357 |
-
Args:
|
358 |
-
- input_frames (list) : List of frames extracted from the video
|
359 |
-
- kp_dict (dict) : Dictionary containing the keypoints and the resolution of the frames
|
360 |
-
- stride (int) : Stride to extract the frames
|
361 |
-
- window_frames (int) : Number of frames in each window that is given as input to the model
|
362 |
-
- width (int) : Width of the frames
|
363 |
-
- height (int) : Height of the frames
|
364 |
-
Returns:
|
365 |
-
- input_frames (array) : Frame window to be given as input to the model
|
366 |
-
- num_frames (int) : Number of frames to extract
|
367 |
-
- orig_masked_frames (array) : Masked frames extracted from the video
|
368 |
-
- msg (string) : Message to be returned
|
369 |
-
'''
|
370 |
-
|
371 |
-
# Face indices to extract the face-coordinates needed for masking
|
372 |
-
face_oval_idx = [10, 21, 54, 58, 67, 93, 103, 109, 127, 132, 136, 148, 149, 150, 152, 162, 172,
|
373 |
-
176, 234, 251, 284, 288, 297, 323, 332, 338, 356, 361, 365, 377, 378, 379, 389, 397, 400, 454]
|
374 |
-
|
375 |
-
|
376 |
-
input_keypoints, resolution = kp_dict['kps'], kp_dict['resolution']
|
377 |
-
print("Input keypoints: ", len(input_keypoints))
|
378 |
-
|
379 |
-
print("Creating masked input frames...")
|
380 |
-
input_frames_masked = []
|
381 |
-
for i, frame_kp_dict in tqdm(enumerate(input_keypoints)):
|
382 |
-
|
383 |
-
img = input_frames[i]
|
384 |
-
face = frame_kp_dict["face"]
|
385 |
-
|
386 |
-
if face is None:
|
387 |
-
img = cv2.resize(img, (width, height))
|
388 |
-
masked_img = cv2.rectangle(img, (0,0), (width,110), (0,0,0), -1)
|
389 |
-
else:
|
390 |
-
face_kps = []
|
391 |
-
for idx in range(len(face)):
|
392 |
-
if idx in face_oval_idx:
|
393 |
-
x, y = int(face[idx]["x"]*resolution[1]), int(face[idx]["y"]*resolution[0])
|
394 |
-
face_kps.append((x,y))
|
395 |
-
|
396 |
-
face_kps = np.array(face_kps)
|
397 |
-
x1, y1 = min(face_kps[:,0]), min(face_kps[:,1])
|
398 |
-
x2, y2 = max(face_kps[:,0]), max(face_kps[:,1])
|
399 |
-
masked_img = cv2.rectangle(img, (0,0), (resolution[1],y2+15), (0,0,0), -1)
|
400 |
-
|
401 |
-
if masked_img.shape[0] != width or masked_img.shape[1] != height:
|
402 |
-
masked_img = cv2.resize(masked_img, (width, height))
|
403 |
-
|
404 |
-
input_frames_masked.append(masked_img)
|
405 |
-
|
406 |
-
orig_masked_frames = np.array(input_frames_masked)
|
407 |
-
input_frames = np.array(input_frames_masked) / 255.
|
408 |
-
print("Input images full: ", input_frames.shape) # num_framesx270x480x3
|
409 |
-
|
410 |
-
input_frames = np.array([input_frames[i:i+window_frames, :, :] for i in range(0,input_frames.shape[0], stride) if (i+window_frames <= input_frames.shape[0])])
|
411 |
-
print("Input images window: ", input_frames.shape) # Tx25x270x480x3
|
412 |
-
|
413 |
-
num_frames = input_frames.shape[0]
|
414 |
-
|
415 |
-
if num_frames<10:
|
416 |
-
msg = "Not enough frames to process! Please give a longer video as input."
|
417 |
-
return None, None, None, msg
|
418 |
-
|
419 |
-
return input_frames, num_frames, orig_masked_frames, "success"
|
420 |
-
|
421 |
-
def load_spectrograms(wav_file, num_frames, window_frames=25, stride=4):
|
422 |
-
|
423 |
-
'''
|
424 |
-
This function extracts the spectrogram from the audio file
|
425 |
-
|
426 |
-
Args:
|
427 |
-
- wav_file (string) : Path of the extracted audio file
|
428 |
-
- num_frames (int) : Number of frames to extract
|
429 |
-
- window_frames (int) : Number of frames in each window that is given as input to the model
|
430 |
-
- stride (int) : Stride to extract the audio frames
|
431 |
-
Returns:
|
432 |
-
- spec (array) : Spectrogram array window to be used as input to the model
|
433 |
-
- orig_spec (array) : Spectrogram array extracted from the audio file
|
434 |
-
- msg (string) : Message to be returned
|
435 |
-
'''
|
436 |
-
|
437 |
-
# Extract the audio from the input video file using ffmpeg
|
438 |
-
try:
|
439 |
-
wav = librosa.load(wav_file, sr=16000)[0]
|
440 |
-
except:
|
441 |
-
msg = "Oops! Could extract the spectrograms from the audio file. Please check the input and try again."
|
442 |
-
return None, None, msg
|
443 |
-
|
444 |
-
# Convert to tensor
|
445 |
-
wav = torch.FloatTensor(wav).unsqueeze(0)
|
446 |
-
mel, _, _, _ = wav2filterbanks(wav.to(device))
|
447 |
-
spec = mel.squeeze(0).cpu().numpy()
|
448 |
-
orig_spec = spec
|
449 |
-
spec = np.array([spec[i:i+(window_frames*stride), :] for i in range(0, spec.shape[0], stride) if (i+(window_frames*stride) <= spec.shape[0])])
|
450 |
-
|
451 |
-
if len(spec) != num_frames:
|
452 |
-
spec = spec[:num_frames]
|
453 |
-
frame_diff = np.abs(len(spec) - num_frames)
|
454 |
-
if frame_diff > 60:
|
455 |
-
print("The input video and audio length do not match - The results can be unreliable! Please check the input video.")
|
456 |
-
|
457 |
-
return spec, orig_spec, "success"
|
458 |
-
|
459 |
-
|
460 |
-
def calc_optimal_av_offset(vid_emb, aud_emb, num_avg_frames, model):
|
461 |
-
'''
|
462 |
-
This function calculates the audio-visual offset between the video and audio
|
463 |
-
|
464 |
-
Args:
|
465 |
-
- vid_emb (array) : Video embedding array
|
466 |
-
- aud_emb (array) : Audio embedding array
|
467 |
-
- num_avg_frames (int) : Number of frames to average the scores
|
468 |
-
- model (object) : Model object
|
469 |
-
Returns:
|
470 |
-
- offset (int) : Optimal audio-visual offset
|
471 |
-
- msg (string) : Message to be returned
|
472 |
-
'''
|
473 |
-
|
474 |
-
pos_vid_emb, all_aud_emb, pos_idx, stride, status = create_online_sync_negatives(vid_emb, aud_emb, num_avg_frames)
|
475 |
-
if status != "success":
|
476 |
-
return None, status
|
477 |
-
scores, _ = calc_av_scores(pos_vid_emb, all_aud_emb, model)
|
478 |
-
offset = scores.argmax()*stride - pos_idx
|
479 |
-
|
480 |
-
return offset.item(), "success"
|
481 |
-
|
482 |
-
def create_online_sync_negatives(vid_emb, aud_emb, num_avg_frames, stride=5):
|
483 |
-
|
484 |
-
'''
|
485 |
-
This function creates all possible positive and negative audio embeddings to compare and obtain the sync offset
|
486 |
-
|
487 |
-
Args:
|
488 |
-
- vid_emb (array) : Video embedding array
|
489 |
-
- aud_emb (array) : Audio embedding array
|
490 |
-
- num_avg_frames (int) : Number of frames to average the scores
|
491 |
-
- stride (int) : Stride to extract the negative windows
|
492 |
-
Returns:
|
493 |
-
- vid_emb_pos (array) : Positive video embedding array
|
494 |
-
- aud_emb_posneg (array) : All possible combinations of audio embedding array
|
495 |
-
- pos_idx_frame (int) : Positive video embedding array frame
|
496 |
-
- stride (int) : Stride used to extract the negative windows
|
497 |
-
- msg (string) : Message to be returned
|
498 |
-
'''
|
499 |
-
|
500 |
-
slice_size = num_avg_frames
|
501 |
-
aud_emb_posneg = aud_emb.squeeze(1).unfold(-1, slice_size, stride)
|
502 |
-
aud_emb_posneg = aud_emb_posneg.permute([0, 2, 1, 3])
|
503 |
-
aud_emb_posneg = aud_emb_posneg[:, :int(n_negative_samples/stride)+1]
|
504 |
-
|
505 |
-
pos_idx = (aud_emb_posneg.shape[1]//2)
|
506 |
-
pos_idx_frame = pos_idx*stride
|
507 |
-
|
508 |
-
min_offset_frames = -(pos_idx)*stride
|
509 |
-
max_offset_frames = (aud_emb_posneg.shape[1] - pos_idx - 1)*stride
|
510 |
-
print("With the current video length and the number of average frames, the model can predict the offsets in the range: [{}, {}]".format(min_offset_frames, max_offset_frames))
|
511 |
-
|
512 |
-
vid_emb_pos = vid_emb[:, :, pos_idx_frame:pos_idx_frame+slice_size]
|
513 |
-
if vid_emb_pos.shape[2] != slice_size:
|
514 |
-
msg = "Video is too short to use {} frames to average the scores. Please use a longer input video or reduce the number of average frames".format(slice_size)
|
515 |
-
return None, None, None, None, msg
|
516 |
-
|
517 |
-
return vid_emb_pos, aud_emb_posneg, pos_idx_frame, stride, "success"
|
518 |
-
|
519 |
-
def calc_av_scores(vid_emb, aud_emb, model):
|
520 |
-
|
521 |
-
'''
|
522 |
-
This function calls functions to calculate the audio-visual similarity and attention map between the video and audio embeddings
|
523 |
-
|
524 |
-
Args:
|
525 |
-
- vid_emb (array) : Video embedding array
|
526 |
-
- aud_emb (array) : Audio embedding array
|
527 |
-
- model (object) : Model object
|
528 |
-
Returns:
|
529 |
-
- scores (array) : Audio-visual similarity scores
|
530 |
-
- att_map (array) : Attention map
|
531 |
-
'''
|
532 |
-
|
533 |
-
scores = calc_att_map(vid_emb, aud_emb, model)
|
534 |
-
att_map = logsoftmax_2d(scores)
|
535 |
-
scores = scores.mean(-1)
|
536 |
-
|
537 |
-
return scores, att_map
|
538 |
-
|
539 |
-
def calc_att_map(vid_emb, aud_emb, model):
|
540 |
-
|
541 |
-
'''
|
542 |
-
This function calculates the similarity between the video and audio embeddings
|
543 |
-
|
544 |
-
Args:
|
545 |
-
- vid_emb (array) : Video embedding array
|
546 |
-
- aud_emb (array) : Audio embedding array
|
547 |
-
- model (object) : Model object
|
548 |
-
Returns:
|
549 |
-
- scores (array) : Audio-visual similarity scores
|
550 |
-
'''
|
551 |
-
|
552 |
-
vid_emb = vid_emb[:, :, None]
|
553 |
-
aud_emb = aud_emb.transpose(1, 2)
|
554 |
-
|
555 |
-
scores = run_func_in_parts(lambda x, y: (x * y).sum(1),
|
556 |
-
vid_emb,
|
557 |
-
aud_emb,
|
558 |
-
part_len=10,
|
559 |
-
dim=3,
|
560 |
-
device=device)
|
561 |
-
|
562 |
-
scores = model.logits_scale(scores[..., None]).squeeze(-1)
|
563 |
-
|
564 |
-
return scores
|
565 |
-
|
566 |
-
def generate_video(frames, audio_file, video_fname):
|
567 |
-
|
568 |
-
'''
|
569 |
-
This function generates the video from the frames and audio file
|
570 |
-
|
571 |
-
Args:
|
572 |
-
- frames (array) : Frames to be used to generate the video
|
573 |
-
- audio_file (string) : Path of the audio file
|
574 |
-
- video_fname (string) : Path of the video file
|
575 |
-
Returns:
|
576 |
-
- video_output (string) : Path of the video file
|
577 |
-
'''
|
578 |
-
|
579 |
-
fname = 'inference.avi'
|
580 |
-
video = cv2.VideoWriter(fname, cv2.VideoWriter_fourcc(*'DIVX'), 25, (frames[0].shape[1], frames[0].shape[0]))
|
581 |
-
|
582 |
-
for i in range(len(frames)):
|
583 |
-
video.write(cv2.cvtColor(frames[i], cv2.COLOR_BGR2RGB))
|
584 |
-
video.release()
|
585 |
-
|
586 |
-
no_sound_video = video_fname + '_nosound.mp4'
|
587 |
-
status = subprocess.call('ffmpeg -hide_banner -loglevel panic -y -i %s -c copy -an -strict -2 %s' % (fname, no_sound_video), shell=True)
|
588 |
-
if status != 0:
|
589 |
-
msg = "Oops! Could not generate the video. Please check the input video and try again."
|
590 |
-
return None, msg
|
591 |
-
|
592 |
-
video_output = video_fname + '.mp4'
|
593 |
-
status = subprocess.call('ffmpeg -hide_banner -loglevel panic -y -i %s -i %s -strict -2 -q:v 1 -shortest %s' %
|
594 |
-
(audio_file, no_sound_video, video_output), shell=True)
|
595 |
-
if status != 0:
|
596 |
-
msg = "Oops! Could not generate the video. Please check the input video and try again."
|
597 |
-
return None, msg
|
598 |
-
|
599 |
-
os.remove(fname)
|
600 |
-
os.remove(no_sound_video)
|
601 |
-
|
602 |
-
return video_output
|
603 |
-
|
604 |
-
def sync_correct_video(video_path, frames, wav_file, offset, result_folder, sample_rate=16000, fps=25):
|
605 |
-
|
606 |
-
'''
|
607 |
-
This function corrects the video and audio to sync with each other
|
608 |
-
|
609 |
-
Args:
|
610 |
-
- video_path (string) : Path of the video file
|
611 |
-
- frames (array) : Frames to be used to generate the video
|
612 |
-
- wav_file (string) : Path of the audio file
|
613 |
-
- offset (int) : Predicted sync-offset to be used to correct the video
|
614 |
-
- result_folder (string) : Path of the result folder to save the output sync-corrected video
|
615 |
-
- sample_rate (int) : Sample rate of the audio
|
616 |
-
- fps (int) : Frames per second of the video
|
617 |
-
Returns:
|
618 |
-
- video_output (string) : Path of the video file
|
619 |
-
'''
|
620 |
-
|
621 |
-
if offset == 0:
|
622 |
-
print("The input audio and video are in-sync! No need to perform sync correction.")
|
623 |
-
return video_path
|
624 |
-
|
625 |
-
print("Performing Sync Correction...")
|
626 |
-
corrected_frames = np.zeros_like(frames)
|
627 |
-
if offset > 0:
|
628 |
-
audio_offset = int(offset*(sample_rate/fps))
|
629 |
-
wav = librosa.core.load(wav_file, sr=sample_rate)[0]
|
630 |
-
corrected_wav = wav[audio_offset:]
|
631 |
-
corrected_wav_file = os.path.join(result_folder, "audio_sync_corrected.wav")
|
632 |
-
write(corrected_wav_file, sample_rate, corrected_wav)
|
633 |
-
wav_file = corrected_wav_file
|
634 |
-
corrected_frames = frames
|
635 |
-
elif offset < 0:
|
636 |
-
corrected_frames[0:len(frames)+offset] = frames[np.abs(offset):]
|
637 |
-
corrected_frames = corrected_frames[:len(frames)-np.abs(offset)]
|
638 |
-
|
639 |
-
corrected_video_path = os.path.join(result_folder, "result_sync_corrected")
|
640 |
-
video_output = generate_video(corrected_frames, wav_file, corrected_video_path)
|
641 |
-
|
642 |
-
return video_output
|
643 |
-
|
644 |
-
class Logger:
|
645 |
-
def __init__(self, filename):
|
646 |
-
self.terminal = sys.stdout
|
647 |
-
self.log = open(filename, "w")
|
648 |
-
|
649 |
-
def write(self, message):
|
650 |
-
self.terminal.write(message)
|
651 |
-
self.log.write(message)
|
652 |
-
|
653 |
-
def flush(self):
|
654 |
-
self.terminal.flush()
|
655 |
-
self.log.flush()
|
656 |
-
|
657 |
-
def isatty(self):
|
658 |
-
return False
|
659 |
-
|
660 |
-
|
661 |
-
def process_video(video_path, num_avg_frames, apply_preprocess):
|
662 |
-
try:
|
663 |
-
# Extract the video filename
|
664 |
-
video_fname = os.path.basename(video_path.split(".")[0])
|
665 |
-
|
666 |
-
# Create folders to save the inputs and results
|
667 |
-
result_folder = os.path.join("results", video_fname)
|
668 |
-
result_folder_input = os.path.join(result_folder, "input")
|
669 |
-
result_folder_output = os.path.join(result_folder, "output")
|
670 |
-
|
671 |
-
if os.path.exists(result_folder):
|
672 |
-
rmtree(result_folder)
|
673 |
-
|
674 |
-
os.makedirs(result_folder)
|
675 |
-
os.makedirs(result_folder_input)
|
676 |
-
os.makedirs(result_folder_output)
|
677 |
-
|
678 |
-
|
679 |
-
# Preprocess the video
|
680 |
-
print("Applying preprocessing: ", apply_preprocess)
|
681 |
-
wav_file, fps, vid_path_processed, status = preprocess_video(video_path, result_folder_input, apply_preprocess)
|
682 |
-
if status != "success":
|
683 |
-
return status, None
|
684 |
-
print("Successfully preprocessed the video")
|
685 |
-
|
686 |
-
# Resample the video to 25 fps if it is not already 25 fps
|
687 |
-
print("FPS of video: ", fps)
|
688 |
-
if fps!=25:
|
689 |
-
vid_path = resample_video(vid_path_processed, "preprocessed_video_25fps", result_folder_input)
|
690 |
-
orig_vid_path_25fps = resample_video(video_path, "input_video_25fps", result_folder_input)
|
691 |
-
else:
|
692 |
-
vid_path = vid_path_processed
|
693 |
-
orig_vid_path_25fps = video_path
|
694 |
-
|
695 |
-
# Load the original video frames (before pre-processing) - Needed for the final sync-correction
|
696 |
-
orig_frames, status = load_video_frames(orig_vid_path_25fps)
|
697 |
-
if status != "success":
|
698 |
-
return status, None
|
699 |
-
|
700 |
-
# Load the pre-processed video frames
|
701 |
-
frames, status = load_video_frames(vid_path)
|
702 |
-
if status != "success":
|
703 |
-
return status, None
|
704 |
-
print("Successfully extracted the video frames")
|
705 |
-
|
706 |
-
if len(frames) < num_avg_frames:
|
707 |
-
return "Error: The input video is too short. Please use a longer input video.", None
|
708 |
-
|
709 |
-
# Load keypoints and check if gestures are visible
|
710 |
-
kp_dict, status = get_keypoints(frames)
|
711 |
-
if status != "success":
|
712 |
-
return status, None
|
713 |
-
print("Successfully extracted the keypoints: ", len(kp_dict), len(kp_dict["kps"]))
|
714 |
-
|
715 |
-
status = check_visible_gestures(kp_dict)
|
716 |
-
if status != "success":
|
717 |
-
return status, None
|
718 |
-
|
719 |
-
# Load RGB frames
|
720 |
-
rgb_frames, num_frames, orig_masked_frames, status = load_rgb_masked_frames(frames, kp_dict, window_frames=25, width=480, height=270)
|
721 |
-
if status != "success":
|
722 |
-
return status, None
|
723 |
-
print("Successfully loaded the RGB frames")
|
724 |
-
|
725 |
-
# Convert frames to tensor
|
726 |
-
rgb_frames = np.transpose(rgb_frames, (4, 0, 1, 2, 3))
|
727 |
-
rgb_frames = torch.FloatTensor(rgb_frames).unsqueeze(0)
|
728 |
-
B = rgb_frames.size(0)
|
729 |
-
print("Successfully converted the frames to tensor")
|
730 |
-
|
731 |
-
# Load spectrograms
|
732 |
-
spec, orig_spec, status = load_spectrograms(wav_file, num_frames, window_frames=25)
|
733 |
-
if status != "success":
|
734 |
-
return status, None
|
735 |
-
spec = torch.FloatTensor(spec).unsqueeze(0).unsqueeze(0).permute(0, 1, 2, 4, 3)
|
736 |
-
print("Successfully loaded the spectrograms")
|
737 |
-
|
738 |
-
# Create input windows
|
739 |
-
video_sequences = torch.cat([rgb_frames[:, :, i] for i in range(rgb_frames.size(2))], dim=0)
|
740 |
-
audio_sequences = torch.cat([spec[:, :, i] for i in range(spec.size(2))], dim=0)
|
741 |
-
|
742 |
-
# Load the trained model
|
743 |
-
model = Transformer_RGB()
|
744 |
-
model = load_checkpoint(CHECKPOINT_PATH, model)
|
745 |
-
print("Successfully loaded the model")
|
746 |
-
|
747 |
-
# Process in batches
|
748 |
-
batch_size = 12
|
749 |
-
video_emb = []
|
750 |
-
audio_emb = []
|
751 |
-
|
752 |
-
for i in tqdm(range(0, len(video_sequences), batch_size)):
|
753 |
-
video_inp = video_sequences[i:i+batch_size, ]
|
754 |
-
audio_inp = audio_sequences[i:i+batch_size, ]
|
755 |
-
|
756 |
-
vid_emb = model.forward_vid(video_inp.to(device))
|
757 |
-
vid_emb = torch.mean(vid_emb, axis=-1).unsqueeze(-1)
|
758 |
-
aud_emb = model.forward_aud(audio_inp.to(device))
|
759 |
-
|
760 |
-
video_emb.append(vid_emb.detach())
|
761 |
-
audio_emb.append(aud_emb.detach())
|
762 |
-
|
763 |
-
torch.cuda.empty_cache()
|
764 |
-
|
765 |
-
audio_emb = torch.cat(audio_emb, dim=0)
|
766 |
-
video_emb = torch.cat(video_emb, dim=0)
|
767 |
-
|
768 |
-
# L2 normalize embeddings
|
769 |
-
video_emb = torch.nn.functional.normalize(video_emb, p=2, dim=1)
|
770 |
-
audio_emb = torch.nn.functional.normalize(audio_emb, p=2, dim=1)
|
771 |
-
|
772 |
-
audio_emb = torch.split(audio_emb, B, dim=0)
|
773 |
-
audio_emb = torch.stack(audio_emb, dim=2)
|
774 |
-
audio_emb = audio_emb.squeeze(3)
|
775 |
-
audio_emb = audio_emb[:, None]
|
776 |
-
|
777 |
-
video_emb = torch.split(video_emb, B, dim=0)
|
778 |
-
video_emb = torch.stack(video_emb, dim=2)
|
779 |
-
video_emb = video_emb.squeeze(3)
|
780 |
-
print("Successfully extracted GestSync embeddings")
|
781 |
-
|
782 |
-
# Calculate sync offset
|
783 |
-
pred_offset, status = calc_optimal_av_offset(video_emb, audio_emb, num_avg_frames, model)
|
784 |
-
if status != "success":
|
785 |
-
return status, None
|
786 |
-
print("Predicted offset: ", pred_offset)
|
787 |
-
|
788 |
-
# Generate sync-corrected video
|
789 |
-
video_output = sync_correct_video(video_path, orig_frames, wav_file, pred_offset, result_folder_output, sample_rate=16000, fps=fps)
|
790 |
-
print("Successfully generated the video:", video_output)
|
791 |
-
|
792 |
-
return f"Predicted offset: {pred_offset}", video_output
|
793 |
-
|
794 |
-
except Exception as e:
|
795 |
-
return f"Error: {str(e)}", None
|
796 |
-
|
797 |
-
def read_logs():
|
798 |
-
sys.stdout.flush()
|
799 |
-
with open("output.log", "r") as f:
|
800 |
-
return f.read()
|
801 |
-
|
802 |
-
|
803 |
-
if __name__ == "__main__":
|
804 |
-
|
805 |
-
sys.stdout = Logger("output.log")
|
806 |
-
|
807 |
-
|
808 |
-
# Define the custom HTML for the header
|
809 |
-
custom_css = """
|
810 |
-
<style>
|
811 |
-
body {
|
812 |
-
background-color: #ffffff;
|
813 |
-
color: #333333; /* Default text color */
|
814 |
-
}
|
815 |
-
.container {
|
816 |
-
max-width: 100% !important;
|
817 |
-
padding-left: 0 !important;
|
818 |
-
padding-right: 0 !important;
|
819 |
-
}
|
820 |
-
.header {
|
821 |
-
background-color: #f0f0f0;
|
822 |
-
color: #333333;
|
823 |
-
padding: 30px;
|
824 |
-
margin-bottom: 30px;
|
825 |
-
text-align: center;
|
826 |
-
font-family: 'Helvetica Neue', Arial, sans-serif;
|
827 |
-
box-shadow: 0 2px 4px rgba(0,0,0,0.1);
|
828 |
-
}
|
829 |
-
.header h1 {
|
830 |
-
font-size: 36px;
|
831 |
-
margin-bottom: 15px;
|
832 |
-
font-weight: bold;
|
833 |
-
color: #333333; /* Explicitly set heading color */
|
834 |
-
}
|
835 |
-
.header h2 {
|
836 |
-
font-size: 24px;
|
837 |
-
margin-bottom: 10px;
|
838 |
-
color: #333333; /* Explicitly set subheading color */
|
839 |
-
}
|
840 |
-
.header p {
|
841 |
-
font-size: 18px;
|
842 |
-
margin: 5px 0;
|
843 |
-
color: #666666;
|
844 |
-
}
|
845 |
-
.blue-text {
|
846 |
-
color: #4a90e2;
|
847 |
-
}
|
848 |
-
/* Custom styles for slider container */
|
849 |
-
.slider-container {
|
850 |
-
background-color: white !important;
|
851 |
-
padding-top: 0.9em;
|
852 |
-
padding-bottom: 0.9em;
|
853 |
-
}
|
854 |
-
/* Add gap before examples */
|
855 |
-
.examples-holder {
|
856 |
-
margin-top: 2em;
|
857 |
-
}
|
858 |
-
/* Set fixed size for example videos */
|
859 |
-
.gradio-container .gradio-examples .gr-sample {
|
860 |
-
width: 240px !important;
|
861 |
-
height: 135px !important;
|
862 |
-
object-fit: cover;
|
863 |
-
display: inline-block;
|
864 |
-
margin-right: 10px;
|
865 |
-
}
|
866 |
-
|
867 |
-
.gradio-container .gradio-examples {
|
868 |
-
display: flex;
|
869 |
-
flex-wrap: wrap;
|
870 |
-
gap: 10px;
|
871 |
-
}
|
872 |
-
|
873 |
-
/* Ensure the parent container does not stretch */
|
874 |
-
.gradio-container .gradio-examples {
|
875 |
-
max-width: 100%;
|
876 |
-
overflow: hidden;
|
877 |
-
}
|
878 |
-
|
879 |
-
/* Additional styles to ensure proper sizing in Safari */
|
880 |
-
.gradio-container .gradio-examples .gr-sample img {
|
881 |
-
width: 240px !important;
|
882 |
-
height: 135px !important;
|
883 |
-
object-fit: cover;
|
884 |
-
}
|
885 |
-
</style>
|
886 |
-
"""
|
887 |
-
|
888 |
-
custom_html = custom_css + """
|
889 |
-
<div class="header">
|
890 |
-
<h1><span class="blue-text">GestSync:</span> Determining who is speaking without a talking head</h1>
|
891 |
-
<h2>Upload any video to predict the synchronization offset and generate a sync-corrected video</h2>
|
892 |
-
<p>Sindhu Hegde and Andrew Zisserman</p>
|
893 |
-
<p>VGG, University of Oxford</p>
|
894 |
-
</div>
|
895 |
-
"""
|
896 |
-
|
897 |
-
# Define paths to sample videos
|
898 |
-
sample_videos = [
|
899 |
-
"samples/sync_sample_1.mp4",
|
900 |
-
"samples/sync_sample_2.mp4",
|
901 |
-
]
|
902 |
-
|
903 |
-
# Define Gradio interface
|
904 |
-
with gr.Blocks(css=custom_css, theme=gr.themes.Default(primary_hue=gr.themes.colors.red, secondary_hue=gr.themes.colors.pink)) as demo:
|
905 |
-
gr.HTML(custom_html)
|
906 |
-
with gr.Row():
|
907 |
-
with gr.Column():
|
908 |
-
with gr.Group(elem_classes="slider-container"):
|
909 |
-
num_avg_frames = gr.Slider(
|
910 |
-
minimum=50,
|
911 |
-
maximum=150,
|
912 |
-
step=5,
|
913 |
-
value=75,
|
914 |
-
label="Number of Average Frames",
|
915 |
-
)
|
916 |
-
apply_preprocess = gr.Checkbox(label="Apply Preprocessing", value=False)
|
917 |
-
video_input = gr.Video(label="Upload Video", height=400)
|
918 |
-
|
919 |
-
with gr.Column():
|
920 |
-
result_text = gr.Textbox(label="Result")
|
921 |
-
output_video = gr.Video(label="Sync Corrected Video", height=400)
|
922 |
-
|
923 |
-
with gr.Row():
|
924 |
-
submit_button = gr.Button("Submit", variant="primary")
|
925 |
-
clear_button = gr.Button("Clear")
|
926 |
-
|
927 |
-
submit_button.click(
|
928 |
-
fn=process_video,
|
929 |
-
inputs=[video_input, num_avg_frames, apply_preprocess],
|
930 |
-
outputs=[result_text, output_video]
|
931 |
-
)
|
932 |
-
|
933 |
-
clear_button.click(
|
934 |
-
fn=lambda: (None, 75, False, "", None),
|
935 |
-
inputs=[],
|
936 |
-
outputs=[video_input, num_avg_frames, apply_preprocess, result_text, output_video]
|
937 |
-
)
|
938 |
-
|
939 |
-
gr.HTML('<div class="examples-holder"></div>')
|
940 |
-
|
941 |
-
# Add examples
|
942 |
-
gr.Examples(
|
943 |
-
examples=sample_videos,
|
944 |
-
inputs=video_input,
|
945 |
-
outputs=None,
|
946 |
-
fn=None,
|
947 |
-
cache_examples=False,
|
948 |
-
)
|
949 |
-
|
950 |
-
logs = gr.Textbox(label="Logs")
|
951 |
-
demo.load(read_logs, None, logs, every=1)
|
952 |
-
|
953 |
-
# Launch the interface
|
954 |
-
demo.queue().launch(allowed_paths=["."], show_error=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|