Spaces:
Running
Running
Update inference2.py
Browse files- inference2.py +345 -345
inference2.py
CHANGED
@@ -1,346 +1,346 @@
|
|
1 |
-
# inference.py (Updated)
|
2 |
-
|
3 |
-
from os import listdir, path
|
4 |
-
import numpy as np
|
5 |
-
import scipy, cv2, os, sys, argparse, audio
|
6 |
-
import json, subprocess, random, string
|
7 |
-
from tqdm import tqdm
|
8 |
-
from glob import glob
|
9 |
-
import torch # Ensure torch is imported
|
10 |
-
try:
|
11 |
-
import face_detection # Assuming this is installed or in a path accessible by your Flask app
|
12 |
-
except ImportError:
|
13 |
-
print("face_detection not found. Please ensure it's installed or available in your PYTHONPATH.")
|
14 |
-
# You might want to raise an error or handle this gracefully if face_detection is truly optional.
|
15 |
-
|
16 |
-
# Make sure you have a models/Wav2Lip.py or similar structure
|
17 |
-
try:
|
18 |
-
from models import Wav2Lip
|
19 |
-
except ImportError:
|
20 |
-
print("Wav2Lip model not found. Please ensure models/Wav2Lip.py exists and is correctly configured.")
|
21 |
-
# You might want to raise an error or handle this gracefully.
|
22 |
-
|
23 |
-
import platform
|
24 |
-
import shutil # For clearing temp directory
|
25 |
-
|
26 |
-
|
27 |
-
# These globals are still useful for shared configuration
|
28 |
-
mel_step_size = 16
|
29 |
-
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
30 |
-
print('Inference script using {} for inference.'.format(device))
|
31 |
-
|
32 |
-
|
33 |
-
def get_smoothened_boxes(boxes, T):
|
34 |
-
for i in range(len(boxes)):
|
35 |
-
if i + T > len(boxes):
|
36 |
-
window = boxes[len(boxes) - T:]
|
37 |
-
else:
|
38 |
-
window = boxes[i : i + T]
|
39 |
-
boxes[i] = np.mean(window, axis=0)
|
40 |
-
return boxes
|
41 |
-
|
42 |
-
def face_detect(images, pads, face_det_batch_size, nosmooth, img_size):
|
43 |
-
detector = face_detection.FaceAlignment(face_detection.LandmarksType._2D,
|
44 |
-
flip_input=False, device=device)
|
45 |
-
|
46 |
-
batch_size = face_det_batch_size
|
47 |
-
|
48 |
-
while 1:
|
49 |
-
predictions = []
|
50 |
-
try:
|
51 |
-
for i in tqdm(range(0, len(images), batch_size), desc="Face Detection"):
|
52 |
-
predictions.extend(detector.get_detections_for_batch(np.array(images[i:i + batch_size])))
|
53 |
-
except RuntimeError as e:
|
54 |
-
if batch_size == 1:
|
55 |
-
raise RuntimeError(f'Image too big to run face detection on GPU. Error: {e}')
|
56 |
-
batch_size //= 2
|
57 |
-
print('Recovering from OOM error; New face detection batch size: {}'.format(batch_size))
|
58 |
-
continue
|
59 |
-
break
|
60 |
-
|
61 |
-
results = []
|
62 |
-
pady1, pady2, padx1, padx2 = pads
|
63 |
-
for rect, image in zip(predictions, images):
|
64 |
-
if rect is None:
|
65 |
-
# Save the faulty frame for debugging
|
66 |
-
output_dir = 'temp' # Ensure this exists or create it
|
67 |
-
os.makedirs(output_dir, exist_ok=True)
|
68 |
-
cv2.imwrite(os.path.join(output_dir, 'faulty_frame.jpg'), image)
|
69 |
-
raise ValueError('Face not detected! Ensure the video/image contains a face in all the frames or try adjusting pads/box.')
|
70 |
-
|
71 |
-
y1 = max(0, rect[1] - pady1)
|
72 |
-
y2 = min(image.shape[0], rect[3] + pady2)
|
73 |
-
x1 = max(0, rect[0] - padx1)
|
74 |
-
x2 = min(image.shape[1], rect[2] + padx2)
|
75 |
-
|
76 |
-
results.append([x1, y1, x2, y2])
|
77 |
-
|
78 |
-
boxes = np.array(results)
|
79 |
-
if not nosmooth: boxes = get_smoothened_boxes(boxes, T=5)
|
80 |
-
results = [[image[y1: y2, x1:x2], (y1, y2, x1, x2)] for image, (x1, y1, x2, y2) in zip(images, boxes)]
|
81 |
-
|
82 |
-
del detector # Clean up detector
|
83 |
-
return results
|
84 |
-
|
85 |
-
def datagen(frames, mels, box, static, wav2lip_batch_size, img_size, pads, face_det_batch_size, nosmooth):
|
86 |
-
img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
|
87 |
-
|
88 |
-
if box[0] == -1:
|
89 |
-
if not static:
|
90 |
-
face_det_results = face_detect(frames, pads, face_det_batch_size, nosmooth, img_size) # BGR2RGB for CNN face detection
|
91 |
-
else:
|
92 |
-
face_det_results = face_detect([frames[0]], pads, face_det_batch_size, nosmooth, img_size)
|
93 |
-
else:
|
94 |
-
print('Using the specified bounding box instead of face detection...')
|
95 |
-
y1, y2, x1, x2 = box
|
96 |
-
face_det_results = [[f[y1: y2, x1:x2], (y1, y2, x1, x2)] for f in frames]
|
97 |
-
|
98 |
-
for i, m in enumerate(mels):
|
99 |
-
idx = 0 if static else i % len(frames)
|
100 |
-
frame_to_save = frames[idx].copy()
|
101 |
-
face, coords = face_det_results[idx].copy()
|
102 |
-
|
103 |
-
face = cv2.resize(face, (img_size, img_size))
|
104 |
-
|
105 |
-
img_batch.append(face)
|
106 |
-
mel_batch.append(m)
|
107 |
-
frame_batch.append(frame_to_save)
|
108 |
-
coords_batch.append(coords)
|
109 |
-
|
110 |
-
if len(img_batch) >= wav2lip_batch_size:
|
111 |
-
img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
|
112 |
-
|
113 |
-
img_masked = img_batch.copy()
|
114 |
-
img_masked[:, img_size//2:] = 0
|
115 |
-
|
116 |
-
img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
|
117 |
-
mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
|
118 |
-
|
119 |
-
yield img_batch, mel_batch, frame_batch, coords_batch
|
120 |
-
img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
|
121 |
-
|
122 |
-
if len(img_batch) > 0:
|
123 |
-
img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
|
124 |
-
|
125 |
-
img_masked = img_batch.copy()
|
126 |
-
img_masked[:, img_size//2:] = 0
|
127 |
-
|
128 |
-
img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
|
129 |
-
mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
|
130 |
-
|
131 |
-
yield img_batch, mel_batch, frame_batch, coords_batch
|
132 |
-
|
133 |
-
def _load(checkpoint_path):
|
134 |
-
# Use torch.jit.load for TorchScript archives
|
135 |
-
if device == 'cuda':
|
136 |
-
model = torch.jit.load(checkpoint_path)
|
137 |
-
else:
|
138 |
-
# Accepts string or torch.device, not a lambda
|
139 |
-
model = torch.jit.load(checkpoint_path, map_location='cpu')
|
140 |
-
return model
|
141 |
-
|
142 |
-
def load_model(path):
|
143 |
-
print("Loading scripted model from:", path)
|
144 |
-
model = _load(path) # returns the TorchScript Module
|
145 |
-
model = model.to(device) # move to CPU or GPU
|
146 |
-
return model.eval() # set to eval() mode
|
147 |
-
|
148 |
-
|
149 |
-
# New function to be called from Flask app
|
150 |
-
def run_inference(
|
151 |
-
checkpoint_path: str,
|
152 |
-
face_path: str,
|
153 |
-
audio_path: str,
|
154 |
-
output_filename: str,
|
155 |
-
static: bool = False,
|
156 |
-
fps: float = 25.,
|
157 |
-
pads: list = [0, 10, 0, 0],
|
158 |
-
face_det_batch_size: int = 16,
|
159 |
-
wav2lip_batch_size: int = 128,
|
160 |
-
resize_factor: int = 1,
|
161 |
-
crop: list = [0, -1, 0, -1],
|
162 |
-
box: list = [-1, -1, -1, -1],
|
163 |
-
rotate: bool = False,
|
164 |
-
nosmooth: bool = False,
|
165 |
-
img_size: int = 96 # Fixed for Wav2Lip
|
166 |
-
) -> str:
|
167 |
-
"""
|
168 |
-
Runs the Wav2Lip inference process.
|
169 |
-
|
170 |
-
Args:
|
171 |
-
checkpoint_path (str): Path to the Wav2Lip model checkpoint.
|
172 |
-
face_path (str): Path to the input video/image file with a face.
|
173 |
-
audio_path (str): Path to the input audio file.
|
174 |
-
output_filename (str): Name of the output video file (e.g., 'result.mp4').
|
175 |
-
static (bool): If True, use only the first video frame for inference.
|
176 |
-
fps (float): Frames per second for static image input.
|
177 |
-
pads (list): Padding for face detection (top, bottom, left, right).
|
178 |
-
face_det_batch_size (int): Batch size for face detection.
|
179 |
-
wav2lip_batch_size (int): Batch size for Wav2Lip model(s).
|
180 |
-
resize_factor (int): Reduce the resolution by this factor.
|
181 |
-
crop (list): Crop video to a smaller region (top, bottom, left, right).
|
182 |
-
box (list): Constant bounding box for the face.
|
183 |
-
rotate (bool): Rotate video right by 90deg.
|
184 |
-
nosmooth (bool): Prevent smoothing face detections.
|
185 |
-
img_size (int): Image size for the model.
|
186 |
-
|
187 |
-
Returns:
|
188 |
-
str: The path to the generated output video file.
|
189 |
-
"""
|
190 |
-
print(f"Starting inference with: face='{face_path}', audio='{audio_path}', checkpoint='{checkpoint_path}', outfile='{output_filename}'")
|
191 |
-
|
192 |
-
# Create necessary directories
|
193 |
-
output_dir = 'results'
|
194 |
-
temp_dir = 'temp'
|
195 |
-
os.makedirs(output_dir, exist_ok=True)
|
196 |
-
os.makedirs(temp_dir, exist_ok=True)
|
197 |
-
|
198 |
-
# Clear temp directory for fresh run
|
199 |
-
for item in os.listdir(temp_dir):
|
200 |
-
item_path = os.path.join(temp_dir, item)
|
201 |
-
if os.path.isfile(item_path):
|
202 |
-
os.remove(item_path)
|
203 |
-
elif os.path.isdir(item_path):
|
204 |
-
shutil.rmtree(item_path)
|
205 |
-
|
206 |
-
# Determine if input is static based on file extension
|
207 |
-
is_static_input = static or (os.path.isfile(face_path) and face_path.split('.')[-1].lower() in ['jpg', 'png', 'jpeg'])
|
208 |
-
|
209 |
-
full_frames = []
|
210 |
-
if is_static_input:
|
211 |
-
full_frames = [cv2.imread(face_path)]
|
212 |
-
if full_frames[0] is None:
|
213 |
-
raise ValueError(f"Could not read face image at: {face_path}")
|
214 |
-
else:
|
215 |
-
video_stream = cv2.VideoCapture(face_path)
|
216 |
-
if not video_stream.isOpened():
|
217 |
-
raise ValueError(f"Could not open video file at: {face_path}")
|
218 |
-
fps = video_stream.get(cv2.CAP_PROP_FPS)
|
219 |
-
|
220 |
-
print('Reading video frames...')
|
221 |
-
while 1:
|
222 |
-
still_reading, frame = video_stream.read()
|
223 |
-
if not still_reading:
|
224 |
-
video_stream.release()
|
225 |
-
break
|
226 |
-
if resize_factor > 1:
|
227 |
-
frame = cv2.resize(frame, (frame.shape[1]//resize_factor, frame.shape[0]//resize_factor))
|
228 |
-
|
229 |
-
if rotate:
|
230 |
-
frame = cv2.rotate(frame, cv2.ROTATE_90_CLOCKWISE)
|
231 |
-
|
232 |
-
y1, y2, x1, x2 = crop
|
233 |
-
if x2 == -1: x2 = frame.shape[1]
|
234 |
-
if y2 == -1: y2 = frame.shape[0]
|
235 |
-
|
236 |
-
frame = frame[y1:y2, x1:x2]
|
237 |
-
full_frames.append(frame)
|
238 |
-
|
239 |
-
print ("Number of frames available for inference: "+str(len(full_frames)))
|
240 |
-
if not full_frames:
|
241 |
-
raise ValueError("No frames could be read from the input face file.")
|
242 |
-
|
243 |
-
temp_audio_path = os.path.join(temp_dir, 'temp_audio.wav')
|
244 |
-
if not audio_path.endswith('.wav'):
|
245 |
-
print('Extracting raw audio...')
|
246 |
-
command = f'ffmpeg -y -i "{audio_path}" -strict -2 "{temp_audio_path}"'
|
247 |
-
try:
|
248 |
-
subprocess.run(command, shell=True, check=True, capture_output=True)
|
249 |
-
audio_path = temp_audio_path
|
250 |
-
except subprocess.CalledProcessError as e:
|
251 |
-
print(f"FFmpeg error: {e.stderr.decode()}")
|
252 |
-
raise RuntimeError(f"Failed to extract audio from {audio_path}. Error: {e.stderr.decode()}")
|
253 |
-
else:
|
254 |
-
# Copy the wav file to temp if it's already wav to maintain consistency in naming
|
255 |
-
shutil.copy(audio_path, temp_audio_path)
|
256 |
-
audio_path = temp_audio_path
|
257 |
-
|
258 |
-
|
259 |
-
wav = audio.load_wav(audio_path, 16000)
|
260 |
-
mel = audio.melspectrogram(wav)
|
261 |
-
print("Mel spectrogram shape:", mel.shape)
|
262 |
-
|
263 |
-
if np.isnan(mel.reshape(-1)).sum() > 0:
|
264 |
-
raise ValueError('Mel contains nan! Using a TTS voice? Add a small epsilon noise to the wav file and try again')
|
265 |
-
|
266 |
-
mel_chunks = []
|
267 |
-
mel_idx_multiplier = 80./fps
|
268 |
-
i = 0
|
269 |
-
while 1:
|
270 |
-
start_idx = int(i * mel_idx_multiplier)
|
271 |
-
if start_idx + mel_step_size > len(mel[0]):
|
272 |
-
mel_chunks.append(mel[:, len(mel[0]) - mel_step_size:])
|
273 |
-
break
|
274 |
-
mel_chunks.append(mel[:, start_idx : start_idx + mel_step_size])
|
275 |
-
i += 1
|
276 |
-
|
277 |
-
print("Length of mel chunks: {}".format(len(mel_chunks)))
|
278 |
-
|
279 |
-
# Ensure full_frames matches mel_chunks length, or loop if static
|
280 |
-
if not is_static_input:
|
281 |
-
full_frames = full_frames[:len(mel_chunks)]
|
282 |
-
else:
|
283 |
-
# If static, replicate the first frame for the duration of the audio
|
284 |
-
full_frames = [full_frames[0]] * len(mel_chunks)
|
285 |
-
|
286 |
-
|
287 |
-
gen = datagen(full_frames.copy(), mel_chunks, box, is_static_input, wav2lip_batch_size, img_size, pads, face_det_batch_size, nosmooth)
|
288 |
-
|
289 |
-
output_avi_path = os.path.join(temp_dir, 'result.avi')
|
290 |
-
|
291 |
-
model_loaded = False
|
292 |
-
model = None
|
293 |
-
frame_h, frame_w = 0, 0
|
294 |
-
out = None
|
295 |
-
|
296 |
-
for i, (img_batch, mel_batch, frames, coords) in enumerate(tqdm(gen, desc="Wav2Lip Inference",
|
297 |
-
total=int(np.ceil(float(len(mel_chunks))/wav2lip_batch_size)))):
|
298 |
-
if not model_loaded:
|
299 |
-
model = load_model(checkpoint_path)
|
300 |
-
model_loaded = True
|
301 |
-
print ("Model loaded successfully")
|
302 |
-
|
303 |
-
frame_h, frame_w = full_frames[0].shape[:-1]
|
304 |
-
out = cv2.VideoWriter(output_avi_path,
|
305 |
-
cv2.VideoWriter_fourcc(*'DIVX'), fps, (frame_w, frame_h))
|
306 |
-
if out is None: # In case no frames were generated for some reason
|
307 |
-
raise RuntimeError("Video writer could not be initialized.")
|
308 |
-
|
309 |
-
|
310 |
-
img_batch = torch.FloatTensor(np.transpose(img_batch, (0, 3, 1, 2))).to(device)
|
311 |
-
mel_batch = torch.FloatTensor(np.transpose(mel_batch, (0, 3, 1, 2))).to(device)
|
312 |
-
|
313 |
-
with torch.no_grad():
|
314 |
-
pred = model(mel_batch, img_batch)
|
315 |
-
|
316 |
-
pred = pred.cpu().numpy().transpose(0, 2, 3, 1) * 255.
|
317 |
-
|
318 |
-
for p, f, c in zip(pred, frames, coords):
|
319 |
-
y1, y2, x1, x2 = c
|
320 |
-
p = cv2.resize(p.astype(np.uint8), (x2 - x1, y2 - y1))
|
321 |
-
|
322 |
-
f[y1:y2, x1:x2] = p
|
323 |
-
out.write(f)
|
324 |
-
|
325 |
-
if out:
|
326 |
-
out.release()
|
327 |
-
else:
|
328 |
-
print("Warning: Video writer was not initialized or no frames were processed.")
|
329 |
-
|
330 |
-
|
331 |
-
final_output_path = os.path.join(output_dir, output_filename)
|
332 |
-
command = f'ffmpeg -y -i "{audio_path}" -i "{output_avi_path}" -strict -2 -q:v 1 "{final_output_path}"'
|
333 |
-
|
334 |
-
try:
|
335 |
-
subprocess.run(command, shell=True, check=True, capture_output=True)
|
336 |
-
print(f"Output saved to: {final_output_path}")
|
337 |
-
except subprocess.CalledProcessError as e:
|
338 |
-
print(f"FFmpeg final merge error: {e.stderr.decode()}")
|
339 |
-
raise RuntimeError(f"Failed to merge audio and video. Error: {e.stderr.decode()}")
|
340 |
-
|
341 |
-
# Clean up temporary files (optional, but good practice)
|
342 |
-
# shutil.rmtree(temp_dir) # Be careful with this if you want to inspect temp files
|
343 |
-
|
344 |
-
return final_output_path
|
345 |
-
|
346 |
# No `if __name__ == '__main__':` block here, as it's meant to be imported
|
|
|
1 |
+
# inference.py (Updated)
|
2 |
+
import audio
|
3 |
+
from os import listdir, path
|
4 |
+
import numpy as np
|
5 |
+
import scipy, cv2, os, sys, argparse, audio
|
6 |
+
import json, subprocess, random, string
|
7 |
+
from tqdm import tqdm
|
8 |
+
from glob import glob
|
9 |
+
import torch # Ensure torch is imported
|
10 |
+
try:
|
11 |
+
import face_detection # Assuming this is installed or in a path accessible by your Flask app
|
12 |
+
except ImportError:
|
13 |
+
print("face_detection not found. Please ensure it's installed or available in your PYTHONPATH.")
|
14 |
+
# You might want to raise an error or handle this gracefully if face_detection is truly optional.
|
15 |
+
|
16 |
+
# Make sure you have a models/Wav2Lip.py or similar structure
|
17 |
+
try:
|
18 |
+
from models import Wav2Lip
|
19 |
+
except ImportError:
|
20 |
+
print("Wav2Lip model not found. Please ensure models/Wav2Lip.py exists and is correctly configured.")
|
21 |
+
# You might want to raise an error or handle this gracefully.
|
22 |
+
|
23 |
+
import platform
|
24 |
+
import shutil # For clearing temp directory
|
25 |
+
|
26 |
+
|
27 |
+
# These globals are still useful for shared configuration
|
28 |
+
mel_step_size = 16
|
29 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
30 |
+
print('Inference script using {} for inference.'.format(device))
|
31 |
+
|
32 |
+
|
33 |
+
def get_smoothened_boxes(boxes, T):
|
34 |
+
for i in range(len(boxes)):
|
35 |
+
if i + T > len(boxes):
|
36 |
+
window = boxes[len(boxes) - T:]
|
37 |
+
else:
|
38 |
+
window = boxes[i : i + T]
|
39 |
+
boxes[i] = np.mean(window, axis=0)
|
40 |
+
return boxes
|
41 |
+
|
42 |
+
def face_detect(images, pads, face_det_batch_size, nosmooth, img_size):
|
43 |
+
detector = face_detection.FaceAlignment(face_detection.LandmarksType._2D,
|
44 |
+
flip_input=False, device=device)
|
45 |
+
|
46 |
+
batch_size = face_det_batch_size
|
47 |
+
|
48 |
+
while 1:
|
49 |
+
predictions = []
|
50 |
+
try:
|
51 |
+
for i in tqdm(range(0, len(images), batch_size), desc="Face Detection"):
|
52 |
+
predictions.extend(detector.get_detections_for_batch(np.array(images[i:i + batch_size])))
|
53 |
+
except RuntimeError as e:
|
54 |
+
if batch_size == 1:
|
55 |
+
raise RuntimeError(f'Image too big to run face detection on GPU. Error: {e}')
|
56 |
+
batch_size //= 2
|
57 |
+
print('Recovering from OOM error; New face detection batch size: {}'.format(batch_size))
|
58 |
+
continue
|
59 |
+
break
|
60 |
+
|
61 |
+
results = []
|
62 |
+
pady1, pady2, padx1, padx2 = pads
|
63 |
+
for rect, image in zip(predictions, images):
|
64 |
+
if rect is None:
|
65 |
+
# Save the faulty frame for debugging
|
66 |
+
output_dir = 'temp' # Ensure this exists or create it
|
67 |
+
os.makedirs(output_dir, exist_ok=True)
|
68 |
+
cv2.imwrite(os.path.join(output_dir, 'faulty_frame.jpg'), image)
|
69 |
+
raise ValueError('Face not detected! Ensure the video/image contains a face in all the frames or try adjusting pads/box.')
|
70 |
+
|
71 |
+
y1 = max(0, rect[1] - pady1)
|
72 |
+
y2 = min(image.shape[0], rect[3] + pady2)
|
73 |
+
x1 = max(0, rect[0] - padx1)
|
74 |
+
x2 = min(image.shape[1], rect[2] + padx2)
|
75 |
+
|
76 |
+
results.append([x1, y1, x2, y2])
|
77 |
+
|
78 |
+
boxes = np.array(results)
|
79 |
+
if not nosmooth: boxes = get_smoothened_boxes(boxes, T=5)
|
80 |
+
results = [[image[y1: y2, x1:x2], (y1, y2, x1, x2)] for image, (x1, y1, x2, y2) in zip(images, boxes)]
|
81 |
+
|
82 |
+
del detector # Clean up detector
|
83 |
+
return results
|
84 |
+
|
85 |
+
def datagen(frames, mels, box, static, wav2lip_batch_size, img_size, pads, face_det_batch_size, nosmooth):
|
86 |
+
img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
|
87 |
+
|
88 |
+
if box[0] == -1:
|
89 |
+
if not static:
|
90 |
+
face_det_results = face_detect(frames, pads, face_det_batch_size, nosmooth, img_size) # BGR2RGB for CNN face detection
|
91 |
+
else:
|
92 |
+
face_det_results = face_detect([frames[0]], pads, face_det_batch_size, nosmooth, img_size)
|
93 |
+
else:
|
94 |
+
print('Using the specified bounding box instead of face detection...')
|
95 |
+
y1, y2, x1, x2 = box
|
96 |
+
face_det_results = [[f[y1: y2, x1:x2], (y1, y2, x1, x2)] for f in frames]
|
97 |
+
|
98 |
+
for i, m in enumerate(mels):
|
99 |
+
idx = 0 if static else i % len(frames)
|
100 |
+
frame_to_save = frames[idx].copy()
|
101 |
+
face, coords = face_det_results[idx].copy()
|
102 |
+
|
103 |
+
face = cv2.resize(face, (img_size, img_size))
|
104 |
+
|
105 |
+
img_batch.append(face)
|
106 |
+
mel_batch.append(m)
|
107 |
+
frame_batch.append(frame_to_save)
|
108 |
+
coords_batch.append(coords)
|
109 |
+
|
110 |
+
if len(img_batch) >= wav2lip_batch_size:
|
111 |
+
img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
|
112 |
+
|
113 |
+
img_masked = img_batch.copy()
|
114 |
+
img_masked[:, img_size//2:] = 0
|
115 |
+
|
116 |
+
img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
|
117 |
+
mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
|
118 |
+
|
119 |
+
yield img_batch, mel_batch, frame_batch, coords_batch
|
120 |
+
img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
|
121 |
+
|
122 |
+
if len(img_batch) > 0:
|
123 |
+
img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
|
124 |
+
|
125 |
+
img_masked = img_batch.copy()
|
126 |
+
img_masked[:, img_size//2:] = 0
|
127 |
+
|
128 |
+
img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
|
129 |
+
mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
|
130 |
+
|
131 |
+
yield img_batch, mel_batch, frame_batch, coords_batch
|
132 |
+
|
133 |
+
def _load(checkpoint_path):
|
134 |
+
# Use torch.jit.load for TorchScript archives
|
135 |
+
if device == 'cuda':
|
136 |
+
model = torch.jit.load(checkpoint_path)
|
137 |
+
else:
|
138 |
+
# Accepts string or torch.device, not a lambda
|
139 |
+
model = torch.jit.load(checkpoint_path, map_location='cpu')
|
140 |
+
return model
|
141 |
+
|
142 |
+
def load_model(path):
|
143 |
+
print("Loading scripted model from:", path)
|
144 |
+
model = _load(path) # returns the TorchScript Module
|
145 |
+
model = model.to(device) # move to CPU or GPU
|
146 |
+
return model.eval() # set to eval() mode
|
147 |
+
|
148 |
+
|
149 |
+
# New function to be called from Flask app
|
150 |
+
def run_inference(
|
151 |
+
checkpoint_path: str,
|
152 |
+
face_path: str,
|
153 |
+
audio_path: str,
|
154 |
+
output_filename: str,
|
155 |
+
static: bool = False,
|
156 |
+
fps: float = 25.,
|
157 |
+
pads: list = [0, 10, 0, 0],
|
158 |
+
face_det_batch_size: int = 16,
|
159 |
+
wav2lip_batch_size: int = 128,
|
160 |
+
resize_factor: int = 1,
|
161 |
+
crop: list = [0, -1, 0, -1],
|
162 |
+
box: list = [-1, -1, -1, -1],
|
163 |
+
rotate: bool = False,
|
164 |
+
nosmooth: bool = False,
|
165 |
+
img_size: int = 96 # Fixed for Wav2Lip
|
166 |
+
) -> str:
|
167 |
+
"""
|
168 |
+
Runs the Wav2Lip inference process.
|
169 |
+
|
170 |
+
Args:
|
171 |
+
checkpoint_path (str): Path to the Wav2Lip model checkpoint.
|
172 |
+
face_path (str): Path to the input video/image file with a face.
|
173 |
+
audio_path (str): Path to the input audio file.
|
174 |
+
output_filename (str): Name of the output video file (e.g., 'result.mp4').
|
175 |
+
static (bool): If True, use only the first video frame for inference.
|
176 |
+
fps (float): Frames per second for static image input.
|
177 |
+
pads (list): Padding for face detection (top, bottom, left, right).
|
178 |
+
face_det_batch_size (int): Batch size for face detection.
|
179 |
+
wav2lip_batch_size (int): Batch size for Wav2Lip model(s).
|
180 |
+
resize_factor (int): Reduce the resolution by this factor.
|
181 |
+
crop (list): Crop video to a smaller region (top, bottom, left, right).
|
182 |
+
box (list): Constant bounding box for the face.
|
183 |
+
rotate (bool): Rotate video right by 90deg.
|
184 |
+
nosmooth (bool): Prevent smoothing face detections.
|
185 |
+
img_size (int): Image size for the model.
|
186 |
+
|
187 |
+
Returns:
|
188 |
+
str: The path to the generated output video file.
|
189 |
+
"""
|
190 |
+
print(f"Starting inference with: face='{face_path}', audio='{audio_path}', checkpoint='{checkpoint_path}', outfile='{output_filename}'")
|
191 |
+
|
192 |
+
# Create necessary directories
|
193 |
+
output_dir = 'results'
|
194 |
+
temp_dir = 'temp'
|
195 |
+
os.makedirs(output_dir, exist_ok=True)
|
196 |
+
os.makedirs(temp_dir, exist_ok=True)
|
197 |
+
|
198 |
+
# Clear temp directory for fresh run
|
199 |
+
for item in os.listdir(temp_dir):
|
200 |
+
item_path = os.path.join(temp_dir, item)
|
201 |
+
if os.path.isfile(item_path):
|
202 |
+
os.remove(item_path)
|
203 |
+
elif os.path.isdir(item_path):
|
204 |
+
shutil.rmtree(item_path)
|
205 |
+
|
206 |
+
# Determine if input is static based on file extension
|
207 |
+
is_static_input = static or (os.path.isfile(face_path) and face_path.split('.')[-1].lower() in ['jpg', 'png', 'jpeg'])
|
208 |
+
|
209 |
+
full_frames = []
|
210 |
+
if is_static_input:
|
211 |
+
full_frames = [cv2.imread(face_path)]
|
212 |
+
if full_frames[0] is None:
|
213 |
+
raise ValueError(f"Could not read face image at: {face_path}")
|
214 |
+
else:
|
215 |
+
video_stream = cv2.VideoCapture(face_path)
|
216 |
+
if not video_stream.isOpened():
|
217 |
+
raise ValueError(f"Could not open video file at: {face_path}")
|
218 |
+
fps = video_stream.get(cv2.CAP_PROP_FPS)
|
219 |
+
|
220 |
+
print('Reading video frames...')
|
221 |
+
while 1:
|
222 |
+
still_reading, frame = video_stream.read()
|
223 |
+
if not still_reading:
|
224 |
+
video_stream.release()
|
225 |
+
break
|
226 |
+
if resize_factor > 1:
|
227 |
+
frame = cv2.resize(frame, (frame.shape[1]//resize_factor, frame.shape[0]//resize_factor))
|
228 |
+
|
229 |
+
if rotate:
|
230 |
+
frame = cv2.rotate(frame, cv2.ROTATE_90_CLOCKWISE)
|
231 |
+
|
232 |
+
y1, y2, x1, x2 = crop
|
233 |
+
if x2 == -1: x2 = frame.shape[1]
|
234 |
+
if y2 == -1: y2 = frame.shape[0]
|
235 |
+
|
236 |
+
frame = frame[y1:y2, x1:x2]
|
237 |
+
full_frames.append(frame)
|
238 |
+
|
239 |
+
print ("Number of frames available for inference: "+str(len(full_frames)))
|
240 |
+
if not full_frames:
|
241 |
+
raise ValueError("No frames could be read from the input face file.")
|
242 |
+
|
243 |
+
temp_audio_path = os.path.join(temp_dir, 'temp_audio.wav')
|
244 |
+
if not audio_path.endswith('.wav'):
|
245 |
+
print('Extracting raw audio...')
|
246 |
+
command = f'ffmpeg -y -i "{audio_path}" -strict -2 "{temp_audio_path}"'
|
247 |
+
try:
|
248 |
+
subprocess.run(command, shell=True, check=True, capture_output=True)
|
249 |
+
audio_path = temp_audio_path
|
250 |
+
except subprocess.CalledProcessError as e:
|
251 |
+
print(f"FFmpeg error: {e.stderr.decode()}")
|
252 |
+
raise RuntimeError(f"Failed to extract audio from {audio_path}. Error: {e.stderr.decode()}")
|
253 |
+
else:
|
254 |
+
# Copy the wav file to temp if it's already wav to maintain consistency in naming
|
255 |
+
shutil.copy(audio_path, temp_audio_path)
|
256 |
+
audio_path = temp_audio_path
|
257 |
+
|
258 |
+
|
259 |
+
wav = audio.load_wav(audio_path, 16000)
|
260 |
+
mel = audio.melspectrogram(wav)
|
261 |
+
print("Mel spectrogram shape:", mel.shape)
|
262 |
+
|
263 |
+
if np.isnan(mel.reshape(-1)).sum() > 0:
|
264 |
+
raise ValueError('Mel contains nan! Using a TTS voice? Add a small epsilon noise to the wav file and try again')
|
265 |
+
|
266 |
+
mel_chunks = []
|
267 |
+
mel_idx_multiplier = 80./fps
|
268 |
+
i = 0
|
269 |
+
while 1:
|
270 |
+
start_idx = int(i * mel_idx_multiplier)
|
271 |
+
if start_idx + mel_step_size > len(mel[0]):
|
272 |
+
mel_chunks.append(mel[:, len(mel[0]) - mel_step_size:])
|
273 |
+
break
|
274 |
+
mel_chunks.append(mel[:, start_idx : start_idx + mel_step_size])
|
275 |
+
i += 1
|
276 |
+
|
277 |
+
print("Length of mel chunks: {}".format(len(mel_chunks)))
|
278 |
+
|
279 |
+
# Ensure full_frames matches mel_chunks length, or loop if static
|
280 |
+
if not is_static_input:
|
281 |
+
full_frames = full_frames[:len(mel_chunks)]
|
282 |
+
else:
|
283 |
+
# If static, replicate the first frame for the duration of the audio
|
284 |
+
full_frames = [full_frames[0]] * len(mel_chunks)
|
285 |
+
|
286 |
+
|
287 |
+
gen = datagen(full_frames.copy(), mel_chunks, box, is_static_input, wav2lip_batch_size, img_size, pads, face_det_batch_size, nosmooth)
|
288 |
+
|
289 |
+
output_avi_path = os.path.join(temp_dir, 'result.avi')
|
290 |
+
|
291 |
+
model_loaded = False
|
292 |
+
model = None
|
293 |
+
frame_h, frame_w = 0, 0
|
294 |
+
out = None
|
295 |
+
|
296 |
+
for i, (img_batch, mel_batch, frames, coords) in enumerate(tqdm(gen, desc="Wav2Lip Inference",
|
297 |
+
total=int(np.ceil(float(len(mel_chunks))/wav2lip_batch_size)))):
|
298 |
+
if not model_loaded:
|
299 |
+
model = load_model(checkpoint_path)
|
300 |
+
model_loaded = True
|
301 |
+
print ("Model loaded successfully")
|
302 |
+
|
303 |
+
frame_h, frame_w = full_frames[0].shape[:-1]
|
304 |
+
out = cv2.VideoWriter(output_avi_path,
|
305 |
+
cv2.VideoWriter_fourcc(*'DIVX'), fps, (frame_w, frame_h))
|
306 |
+
if out is None: # In case no frames were generated for some reason
|
307 |
+
raise RuntimeError("Video writer could not be initialized.")
|
308 |
+
|
309 |
+
|
310 |
+
img_batch = torch.FloatTensor(np.transpose(img_batch, (0, 3, 1, 2))).to(device)
|
311 |
+
mel_batch = torch.FloatTensor(np.transpose(mel_batch, (0, 3, 1, 2))).to(device)
|
312 |
+
|
313 |
+
with torch.no_grad():
|
314 |
+
pred = model(mel_batch, img_batch)
|
315 |
+
|
316 |
+
pred = pred.cpu().numpy().transpose(0, 2, 3, 1) * 255.
|
317 |
+
|
318 |
+
for p, f, c in zip(pred, frames, coords):
|
319 |
+
y1, y2, x1, x2 = c
|
320 |
+
p = cv2.resize(p.astype(np.uint8), (x2 - x1, y2 - y1))
|
321 |
+
|
322 |
+
f[y1:y2, x1:x2] = p
|
323 |
+
out.write(f)
|
324 |
+
|
325 |
+
if out:
|
326 |
+
out.release()
|
327 |
+
else:
|
328 |
+
print("Warning: Video writer was not initialized or no frames were processed.")
|
329 |
+
|
330 |
+
|
331 |
+
final_output_path = os.path.join(output_dir, output_filename)
|
332 |
+
command = f'ffmpeg -y -i "{audio_path}" -i "{output_avi_path}" -strict -2 -q:v 1 "{final_output_path}"'
|
333 |
+
|
334 |
+
try:
|
335 |
+
subprocess.run(command, shell=True, check=True, capture_output=True)
|
336 |
+
print(f"Output saved to: {final_output_path}")
|
337 |
+
except subprocess.CalledProcessError as e:
|
338 |
+
print(f"FFmpeg final merge error: {e.stderr.decode()}")
|
339 |
+
raise RuntimeError(f"Failed to merge audio and video. Error: {e.stderr.decode()}")
|
340 |
+
|
341 |
+
# Clean up temporary files (optional, but good practice)
|
342 |
+
# shutil.rmtree(temp_dir) # Be careful with this if you want to inspect temp files
|
343 |
+
|
344 |
+
return final_output_path
|
345 |
+
|
346 |
# No `if __name__ == '__main__':` block here, as it's meant to be imported
|