Upload 96 files

.gitattributes

@@ -33,3 +33,33 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+input/audio_hindi_tony_stark.mp3 filter=lfs diff=lfs merge=lfs -text
+input/audio/audio_hindi_tony_stark.mp3 filter=lfs diff=lfs merge=lfs -text
+input/audio/harvard.wav filter=lfs diff=lfs merge=lfs -text
+input/audio/processed_tony_stark.mp3 filter=lfs diff=lfs merge=lfs -text
+input/audio/sample_male_audio.mp3 filter=lfs diff=lfs merge=lfs -text
+input/harvard.wav filter=lfs diff=lfs merge=lfs -text
+input/image/images.jpg filter=lfs diff=lfs merge=lfs -text
+input/image/portrait-young-confident-handsome-businessman-blue-red-light.jpg filter=lfs diff=lfs merge=lfs -text
+input/image/young-model-casual-fall-winter-outfits.jpg filter=lfs diff=lfs merge=lfs -text
+input/portrait-young-confident-handsome-businessman-blue-red-light.jpg filter=lfs diff=lfs merge=lfs -text
+input/processed_tony_stark.mp3 filter=lfs diff=lfs merge=lfs -text
+input/processed_tony_stark.mp4 filter=lfs diff=lfs merge=lfs -text
+input/tony_2.mp4 filter=lfs diff=lfs merge=lfs -text
+input/tony_stark.mp4 filter=lfs diff=lfs merge=lfs -text
+input/video/man_sample1_green_screen.mp4 filter=lfs diff=lfs merge=lfs -text
+input/video/tony_2.mp4 filter=lfs diff=lfs merge=lfs -text
+input/young-model-casual-fall-winter-outfits.jpg filter=lfs diff=lfs merge=lfs -text
+results/output_image_video.mp4 filter=lfs diff=lfs merge=lfs -text
+results/output_lipsynced.mp4 filter=lfs diff=lfs merge=lfs -text
+results/output_tony_howard.mp4 filter=lfs diff=lfs merge=lfs -text
+results/output_video_green_screen.mp4 filter=lfs diff=lfs merge=lfs -text
+results/output_video.mp4 filter=lfs diff=lfs merge=lfs -text
+results/result_14c740551950bab2.mp4 filter=lfs diff=lfs merge=lfs -text
+results/result_8d8eccddd8fe1694.mp4 filter=lfs diff=lfs merge=lfs -text
+results/result_be9149db044ae463.mp4 filter=lfs diff=lfs merge=lfs -text
+results/result_c5fbd07c81197985.mp4 filter=lfs diff=lfs merge=lfs -text
+results/result_ea00d0588539fef5.mp4 filter=lfs diff=lfs merge=lfs -text
+results/result_voice.mp4 filter=lfs diff=lfs merge=lfs -text
+temp/result.avi filter=lfs diff=lfs merge=lfs -text
+temp/temp_audio.wav filter=lfs diff=lfs merge=lfs -text

app.py

@@ -0,0 +1,173 @@
+from flask import Flask, render_template, request, session, redirect, url_for, flash, send_from_directory
+import os
+import secrets
+from werkzeug.utils import secure_filename
+import sys
+import shutil
+
+sys.path.append(os.path.dirname(__file__))
+import inference2 # Import your refactored inference script
+
+app = Flask(__name__)
+app.secret_key = os.urandom(24)
+app.config['UPLOAD_FOLDER'] = 'uploads'
+app.config['RESULTS_FOLDER'] = 'results' # This directory is NOT inside static
+app.config['CHECKPOINTS_FOLDER'] = 'checkpoints'
+app.config['TEMP_FOLDER'] = 'temp'
+
+ALLOWED_FACE_EXTENSIONS = {'png', 'jpg', 'jpeg', 'mp4', 'avi', 'mov'}
+ALLOWED_AUDIO_EXTENSIONS = {'wav', 'mp3', 'aac', 'flac'}
+ALLOWED_MODEL_EXTENSIONS = {'pth', 'pt'}
+
+os.makedirs(app.config['UPLOAD_FOLDER'], exist_ok=True)
+os.makedirs(app.config['RESULTS_FOLDER'], exist_ok=True)
+os.makedirs(app.config['CHECKPOINTS_FOLDER'], exist_ok=True)
+os.makedirs(app.config['TEMP_FOLDER'], exist_ok=True)
+
+
+def allowed_file(filename, allowed_extensions):
+    return '.' in filename and \
+           filename.rsplit('.', 1)[1].lower() in allowed_extensions
+
+@app.route('/')
+def index():
+    theme = session.get('theme', 'dark')
+    available_models = []
+    try:
+        model_files = [f for f in os.listdir(app.config['CHECKPOINTS_FOLDER'])
+                       if allowed_file(f, ALLOWED_MODEL_EXTENSIONS)]
+        available_models = sorted(model_files)
+    except FileNotFoundError:
+        # flash("Checkpoints folder not found. Please create a 'checkpoints' directory.", 'error') # Messages removed
+        pass
+    except Exception as e:
+        # flash(f"Error loading models: {e}", 'error') # Messages removed
+        pass
+    return render_template('index.html', theme=theme, models=available_models)
+
+@app.route('/toggle_theme')
+def toggle_theme():
+    current_theme = session.get('theme', 'dark')
+    if current_theme == 'dark':
+        session['theme'] = 'light'
+    else:
+        session['theme'] = 'dark'
+    return redirect(request.referrer or url_for('index'))
+
+@app.route('/infer', methods=['POST'])
+def infer():
+    if request.method == 'POST':
+        if 'face_file' not in request.files or 'audio_file' not in request.files:
+            # flash('Both face and audio files are required.', 'error') # Messages removed
+            return redirect(url_for('index'))
+
+        face_file = request.files['face_file']
+        audio_file = request.files['audio_file']
+        selected_model = request.form.get('model_select')
+
+        if face_file.filename == '' or audio_file.filename == '':
+            # flash('No selected file for face or audio.', 'error') # Messages removed
+            return redirect(url_for('index'))
+
+        if not selected_model:
+            # flash('No model selected.', 'error') # Messages removed
+            return redirect(url_for('index'))
+
+        if not allowed_file(face_file.filename, ALLOWED_FACE_EXTENSIONS):
+            # flash('Invalid face file type. Allowed: png, jpg, jpeg, mp4, avi, mov', 'error') # Messages removed
+            return redirect(url_for('index'))
+        if not allowed_file(audio_file.filename, ALLOWED_AUDIO_EXTENSIONS):
+            # flash('Invalid audio file type. Allowed: wav, mp3, aac, flac', 'error') # Messages removed
+            return redirect(url_for('index'))
+
+        face_filename = secure_filename(face_file.filename)
+        audio_filename = secure_filename(audio_file.filename)
+
+        face_uuid = secrets.token_hex(8)
+        audio_uuid = secrets.token_hex(8)
+
+        face_path = os.path.join(app.config['UPLOAD_FOLDER'], f"{face_uuid}_{face_filename}")
+        audio_path = os.path.join(app.config['UPLOAD_FOLDER'], f"{audio_uuid}_{audio_filename}")
+
+        try:
+            face_file.save(face_path)
+            audio_file.save(audio_path)
+        except Exception as e:
+            # flash(f"Error saving uploaded files: {e}", 'error') # Messages removed
+            return redirect(url_for('index'))
+
+        checkpoint_path = os.path.join(app.config['CHECKPOINTS_FOLDER'], selected_model)
+        output_video_name = f"result_{face_uuid}.mp4"
+
+        try:
+            # flash('Starting inference... This may take a while.', 'info') # Messages removed
+            generated_video_path = inference2.run_inference(
+                checkpoint_path=checkpoint_path,
+                face_path=face_path,
+                audio_path=audio_path,
+                output_filename=output_video_name,
+                static=request.form.get('static_input') == 'on',
+                fps=float(request.form.get('fps', 25.0)),
+                resize_factor=int(request.form.get('resize_factor', 1)),
+                rotate=request.form.get('rotate') == 'on',
+                nosmooth=request.form.get('nosmooth') == 'on',
+                pads=[0, 10, 0, 0],
+                crop=[0, -1, 0, -1],
+                box=[-1, -1, -1, -1],
+                face_det_batch_size=16,
+                wav2lip_batch_size=128,
+                img_size=96
+            )
+            # flash('Inference completed successfully!', 'success') # Messages removed
+            
+            # Redirect to the page that renders result.html
+            return redirect(url_for('render_result_page', filename=os.path.basename(generated_video_path)))
+
+        except ValueError as e:
+            # flash(f"Inference Error: {e}", 'error') # Messages removed
+            pass
+        except RuntimeError as e:
+            # flash(f"Runtime Error during inference: {e}", 'error') # Messages removed
+            pass
+        except Exception as e:
+            # flash(f"An unexpected error occurred: {e}", 'error') # Messages removed
+            pass
+        finally:
+            if os.path.exists(face_path):
+                os.remove(face_path)
+            if os.path.exists(audio_path):
+                os.remove(audio_path)
+
+    return redirect(url_for('index'))
+
+# Route to render the result.html template
+@app.route('/result_page/<filename>')
+def render_result_page(filename):
+    theme = session.get('theme', 'dark')
+    # Check if the file actually exists before rendering
+    if not os.path.exists(os.path.join(app.config['RESULTS_FOLDER'], filename)):
+        # If the video isn't found, redirect or show an error
+        # Consider a dedicated error page or a message within index.html if no flashes are used
+        return redirect(url_for('index'))
+    return render_template('result.html', theme=theme, video_filename=filename)
+
+
+# Route to serve the video file itself (used by <video src="...">)
+@app.route('/results/<path:filename>') # Use <path:filename> to handle potential subdirectories in filename (though not needed here)
+def serve_result_video(filename):
+    # This route is solely for serving the video file
+    return send_from_directory(app.config['RESULTS_FOLDER'], filename)
+
+# Route to download the video file
+@app.route('/download/<filename>') # Changed to /download/ for clarity
+def download_result(filename):
+    return send_from_directory(app.config['RESULTS_FOLDER'], filename, as_attachment=True)
+
+
+if __name__ == '__main__':
+    os.makedirs(app.config['UPLOAD_FOLDER'], exist_ok=True)
+    os.makedirs(app.config['RESULTS_FOLDER'], exist_ok=True)
+    os.makedirs(app.config['CHECKPOINTS_FOLDER'], exist_ok=True)
+    os.makedirs(app.config['TEMP_FOLDER'], exist_ok=True)
+
+    app.run(debug=True)

evaluation/README.md

@@ -0,0 +1,63 @@
+# Novel Evaluation Framework, new filelists, and using the LSE-D and LSE-C metric.
+
+Our paper also proposes a novel evaluation framework (Section 4). To evaluate on LRS2, LRS3, and LRW, the filelists are present in the `test_filelists` folder. Please use `gen_videos_from_filelist.py` script to generate the videos. After that, you can calculate the LSE-D and LSE-C scores using the instructions below. Please see [this thread](https://github.com/Rudrabha/Wav2Lip/issues/22#issuecomment-712825380) on how to calculate the FID scores. 
+
+The videos of the ReSyncED benchmark for real-world evaluation will be released soon. 
+
+### Steps to set-up the evaluation repository for LSE-D and LSE-C metric:
+We use the pre-trained syncnet model available in this [repository](https://github.com/joonson/syncnet_python). 
+
+* Clone the SyncNet repository.
+``` 
+git clone https://github.com/joonson/syncnet_python.git 
+```
+* Follow the procedure given in the above linked [repository](https://github.com/joonson/syncnet_python) to download the pretrained models and set up the dependencies. 
+    * **Note: Please install a separate virtual environment for the evaluation scripts. The versions used by Wav2Lip and the publicly released code of SyncNet is different and can cause version mis-match issues. To avoid this, we suggest the users to install a separate virtual environment for the evaluation scripts**
+```
+cd syncnet_python
+pip install -r requirements.txt
+sh download_model.sh
+```
+* The above step should ensure that all the dependencies required by the repository is installed and the pre-trained models are downloaded.
+
+### Running the evaluation scripts:
+* Copy our evaluation scripts given in this folder to the cloned repository.
+```  
+    cd Wav2Lip/evaluation/scores_LSE/
+    cp *.py syncnet_python/
+    cp *.sh syncnet_python/ 
+```
+**Note: We will release the test filelists for LRW, LRS2 and LRS3 shortly once we receive permission from the dataset creators. We will also release the Real World Dataset we have collected shortly.**
+
+* Our evaluation technique does not require ground-truth of any sorts. Given lip-synced videos we can directly calculate the scores from only the generated videos. Please store the generated videos (from our test sets or your own generated videos) in the following folder structure.
+```
+video data root (Folder containing all videos)
+├── All .mp4 files
+```
+* Change the folder back to the cloned repository. 
+```
+cd syncnet_python
+```
+* To run evaluation on the LRW, LRS2 and LRS3 test files, please run the following command:
+```
+python calculate_scores_LRS.py --data_root /path/to/video/data/root --tmp_dir tmp_dir/
+```
+
+* To run evaluation on the ReSynced dataset or your own generated videos, please run the following command:
+```
+sh calculate_scores_real_videos.sh /path/to/video/data/root
+```
+* The generated scores will be present in the all_scores.txt generated in the ```syncnet_python/``` folder
+
+# Evaluation of image quality using FID metric.
+We use the [pytorch-fid](https://github.com/mseitzer/pytorch-fid) repository for calculating the FID metrics. We dump all the frames in both ground-truth and generated videos and calculate the FID score. 
+
+
+# Opening issues related to evaluation scripts
+* Please open the issues with the "Evaluation" label if you face any issues in the evaluation scripts. 
+
+# Acknowledgements
+Our evaluation pipeline in based on two existing repositories. LSE metrics are based on the [syncnet_python](https://github.com/joonson/syncnet_python) repository and the FID score is based on [pytorch-fid](https://github.com/mseitzer/pytorch-fid) repository. We thank the authors of both the repositories for releasing their wonderful code.
+
+
+

evaluation/gen_videos_from_filelist.py

@@ -0,0 +1,238 @@
+from os import listdir, path
+import numpy as np
+import scipy, cv2, os, sys, argparse
+import dlib, json, subprocess
+from tqdm import tqdm
+from glob import glob
+import torch
+
+sys.path.append('../')
+import audio
+import face_detection
+from models import Wav2Lip
+
+parser = argparse.ArgumentParser(description='Code to generate results for test filelists')
+
+parser.add_argument('--filelist', type=str, 
+					help='Filepath of filelist file to read', required=True)
+parser.add_argument('--results_dir', type=str, help='Folder to save all results into', 
+									required=True)
+parser.add_argument('--data_root', type=str, required=True)
+parser.add_argument('--checkpoint_path', type=str, 
+					help='Name of saved checkpoint to load weights from', required=True)
+
+parser.add_argument('--pads', nargs='+', type=int, default=[0, 0, 0, 0], 
+					help='Padding (top, bottom, left, right)')
+parser.add_argument('--face_det_batch_size', type=int, 
+					help='Single GPU batch size for face detection', default=64)
+parser.add_argument('--wav2lip_batch_size', type=int, help='Batch size for Wav2Lip', default=128)
+
+# parser.add_argument('--resize_factor', default=1, type=int)
+
+args = parser.parse_args()
+args.img_size = 96
+
+def get_smoothened_boxes(boxes, T):
+	for i in range(len(boxes)):
+		if i + T > len(boxes):
+			window = boxes[len(boxes) - T:]
+		else:
+			window = boxes[i : i + T]
+		boxes[i] = np.mean(window, axis=0)
+	return boxes
+
+def face_detect(images):
+	batch_size = args.face_det_batch_size
+	
+	while 1:
+		predictions = []
+		try:
+			for i in range(0, len(images), batch_size):
+				predictions.extend(detector.get_detections_for_batch(np.array(images[i:i + batch_size])))
+		except RuntimeError:
+			if batch_size == 1:
+				raise RuntimeError('Image too big to run face detection on GPU')
+			batch_size //= 2
+			args.face_det_batch_size = batch_size
+			print('Recovering from OOM error; New batch size: {}'.format(batch_size))
+			continue
+		break
+
+	results = []
+	pady1, pady2, padx1, padx2 = args.pads
+	for rect, image in zip(predictions, images):
+		if rect is None:
+			raise ValueError('Face not detected!')
+
+		y1 = max(0, rect[1] - pady1)
+		y2 = min(image.shape[0], rect[3] + pady2)
+		x1 = max(0, rect[0] - padx1)
+		x2 = min(image.shape[1], rect[2] + padx2)
+		
+		results.append([x1, y1, x2, y2])
+
+	boxes = get_smoothened_boxes(np.array(results), T=5)
+	results = [[image[y1: y2, x1:x2], (y1, y2, x1, x2), True] for image, (x1, y1, x2, y2) in zip(images, boxes)]
+
+	return results 
+
+def datagen(frames, face_det_results, mels):
+	img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
+
+	for i, m in enumerate(mels):
+		if i >= len(frames): raise ValueError('Equal or less lengths only')
+
+		frame_to_save = frames[i].copy()
+		face, coords, valid_frame = face_det_results[i].copy()
+		if not valid_frame:
+			continue
+
+		face = cv2.resize(face, (args.img_size, args.img_size))
+			
+		img_batch.append(face)
+		mel_batch.append(m)
+		frame_batch.append(frame_to_save)
+		coords_batch.append(coords)
+
+		if len(img_batch) >= args.wav2lip_batch_size:
+			img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
+
+			img_masked = img_batch.copy()
+			img_masked[:, args.img_size//2:] = 0
+
+			img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
+			mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
+
+			yield img_batch, mel_batch, frame_batch, coords_batch
+			img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
+
+	if len(img_batch) > 0:
+		img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
+
+		img_masked = img_batch.copy()
+		img_masked[:, args.img_size//2:] = 0
+
+		img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
+		mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
+
+		yield img_batch, mel_batch, frame_batch, coords_batch
+
+fps = 25
+mel_step_size = 16
+mel_idx_multiplier = 80./fps
+device = 'cuda' if torch.cuda.is_available() else 'cpu'
+print('Using {} for inference.'.format(device))
+
+detector = face_detection.FaceAlignment(face_detection.LandmarksType._2D, 
+											flip_input=False, device=device)
+
+def _load(checkpoint_path):
+	if device == 'cuda':
+		checkpoint = torch.load(checkpoint_path)
+	else:
+		checkpoint = torch.load(checkpoint_path,
+								map_location=lambda storage, loc: storage)
+	return checkpoint
+
+def load_model(path):
+	model = Wav2Lip()
+	print("Load checkpoint from: {}".format(path))
+	checkpoint = _load(path)
+	s = checkpoint["state_dict"]
+	new_s = {}
+	for k, v in s.items():
+		new_s[k.replace('module.', '')] = v
+	model.load_state_dict(new_s)
+
+	model = model.to(device)
+	return model.eval()
+
+model = load_model(args.checkpoint_path)
+
+def main():
+	assert args.data_root is not None
+	data_root = args.data_root
+
+	if not os.path.isdir(args.results_dir): os.makedirs(args.results_dir)
+
+	with open(args.filelist, 'r') as filelist:
+		lines = filelist.readlines()
+
+	for idx, line in enumerate(tqdm(lines)):
+		audio_src, video = line.strip().split()
+
+		audio_src = os.path.join(data_root, audio_src) + '.mp4'
+		video = os.path.join(data_root, video) + '.mp4'
+
+		command = 'ffmpeg -loglevel panic -y -i {} -strict -2 {}'.format(audio_src, '../temp/temp.wav')
+		subprocess.call(command, shell=True)
+		temp_audio = '../temp/temp.wav'
+
+		wav = audio.load_wav(temp_audio, 16000)
+		mel = audio.melspectrogram(wav)
+		if np.isnan(mel.reshape(-1)).sum() > 0:
+			continue
+
+		mel_chunks = []
+		i = 0
+		while 1:
+			start_idx = int(i * mel_idx_multiplier)
+			if start_idx + mel_step_size > len(mel[0]):
+				break
+			mel_chunks.append(mel[:, start_idx : start_idx + mel_step_size])
+			i += 1
+
+		video_stream = cv2.VideoCapture(video)
+			
+		full_frames = []
+		while 1:
+			still_reading, frame = video_stream.read()
+			if not still_reading or len(full_frames) > len(mel_chunks):
+				video_stream.release()
+				break
+			full_frames.append(frame)
+
+		if len(full_frames) < len(mel_chunks):
+			continue
+
+		full_frames = full_frames[:len(mel_chunks)]
+
+		try:
+			face_det_results = face_detect(full_frames.copy())
+		except ValueError as e:
+			continue
+
+		batch_size = args.wav2lip_batch_size
+		gen = datagen(full_frames.copy(), face_det_results, mel_chunks)
+
+		for i, (img_batch, mel_batch, frames, coords) in enumerate(gen):
+			if i == 0:
+				frame_h, frame_w = full_frames[0].shape[:-1]
+				out = cv2.VideoWriter('../temp/result.avi', 
+								cv2.VideoWriter_fourcc(*'DIVX'), fps, (frame_w, frame_h))
+
+			img_batch = torch.FloatTensor(np.transpose(img_batch, (0, 3, 1, 2))).to(device)
+			mel_batch = torch.FloatTensor(np.transpose(mel_batch, (0, 3, 1, 2))).to(device)
+
+			with torch.no_grad():
+				pred = model(mel_batch, img_batch)
+					
+
+			pred = pred.cpu().numpy().transpose(0, 2, 3, 1) * 255.
+			
+			for pl, f, c in zip(pred, frames, coords):
+				y1, y2, x1, x2 = c
+				pl = cv2.resize(pl.astype(np.uint8), (x2 - x1, y2 - y1))
+				f[y1:y2, x1:x2] = pl
+				out.write(f)
+
+		out.release()
+
+		vid = os.path.join(args.results_dir, '{}.mp4'.format(idx))
+
+		command = 'ffmpeg -loglevel panic -y -i {} -i {} -strict -2 -q:v 1 {}'.format(temp_audio, 
+								'../temp/result.avi', vid)
+		subprocess.call(command, shell=True)
+
+if __name__ == '__main__':
+	main()

evaluation/real_videos_inference.py

@@ -0,0 +1,305 @@
+from os import listdir, path
+import numpy as np
+import scipy, cv2, os, sys, argparse
+import dlib, json, subprocess
+from tqdm import tqdm
+from glob import glob
+import torch
+
+sys.path.append('../')
+import audio
+import face_detection
+from models import Wav2Lip
+
+parser = argparse.ArgumentParser(description='Code to generate results on ReSyncED evaluation set')
+
+parser.add_argument('--mode', type=str, 
+					help='random | dubbed | tts', required=True)
+
+parser.add_argument('--filelist', type=str, 
+					help='Filepath of filelist file to read', default=None)
+
+parser.add_argument('--results_dir', type=str, help='Folder to save all results into', 
+									required=True)
+parser.add_argument('--data_root', type=str, required=True)
+parser.add_argument('--checkpoint_path', type=str, 
+					help='Name of saved checkpoint to load weights from', required=True)
+parser.add_argument('--pads', nargs='+', type=int, default=[0, 10, 0, 0], 
+					help='Padding (top, bottom, left, right)')
+
+parser.add_argument('--face_det_batch_size', type=int, 
+					help='Single GPU batch size for face detection', default=16)
+
+parser.add_argument('--wav2lip_batch_size', type=int, help='Batch size for Wav2Lip', default=128)
+parser.add_argument('--face_res', help='Approximate resolution of the face at which to test', default=180)
+parser.add_argument('--min_frame_res', help='Do not downsample further below this frame resolution', default=480)
+parser.add_argument('--max_frame_res', help='Downsample to at least this frame resolution', default=720)
+# parser.add_argument('--resize_factor', default=1, type=int)
+
+args = parser.parse_args()
+args.img_size = 96
+
+def get_smoothened_boxes(boxes, T):
+	for i in range(len(boxes)):
+		if i + T > len(boxes):
+			window = boxes[len(boxes) - T:]
+		else:
+			window = boxes[i : i + T]
+		boxes[i] = np.mean(window, axis=0)
+	return boxes
+
+def rescale_frames(images):
+	rect = detector.get_detections_for_batch(np.array([images[0]]))[0]
+	if rect is None:
+		raise ValueError('Face not detected!')
+	h, w = images[0].shape[:-1]
+
+	x1, y1, x2, y2 = rect
+
+	face_size = max(np.abs(y1 - y2), np.abs(x1 - x2))
+
+	diff = np.abs(face_size - args.face_res)
+	for factor in range(2, 16):
+		downsampled_res = face_size // factor
+		if min(h//factor, w//factor) < args.min_frame_res: break 
+		if np.abs(downsampled_res - args.face_res) >= diff: break
+
+	factor -= 1
+	if factor == 1: return images
+
+	return [cv2.resize(im, (im.shape[1]//(factor), im.shape[0]//(factor))) for im in images]
+
+
+def face_detect(images):
+	batch_size = args.face_det_batch_size
+	images = rescale_frames(images)
+
+	while 1:
+		predictions = []
+		try:
+			for i in range(0, len(images), batch_size):
+				predictions.extend(detector.get_detections_for_batch(np.array(images[i:i + batch_size])))
+		except RuntimeError:
+			if batch_size == 1:
+				raise RuntimeError('Image too big to run face detection on GPU')
+			batch_size //= 2
+			print('Recovering from OOM error; New batch size: {}'.format(batch_size))
+			continue
+		break
+
+	results = []
+	pady1, pady2, padx1, padx2 = args.pads
+	for rect, image in zip(predictions, images):
+		if rect is None:
+			raise ValueError('Face not detected!')
+
+		y1 = max(0, rect[1] - pady1)
+		y2 = min(image.shape[0], rect[3] + pady2)
+		x1 = max(0, rect[0] - padx1)
+		x2 = min(image.shape[1], rect[2] + padx2)
+		
+		results.append([x1, y1, x2, y2])
+
+	boxes = get_smoothened_boxes(np.array(results), T=5)
+	results = [[image[y1: y2, x1:x2], (y1, y2, x1, x2), True] for image, (x1, y1, x2, y2) in zip(images, boxes)]
+
+	return results, images 
+
+def datagen(frames, face_det_results, mels):
+	img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
+
+	for i, m in enumerate(mels):
+		if i >= len(frames): raise ValueError('Equal or less lengths only')
+
+		frame_to_save = frames[i].copy()
+		face, coords, valid_frame = face_det_results[i].copy()
+		if not valid_frame:
+			continue
+
+		face = cv2.resize(face, (args.img_size, args.img_size))
+			
+		img_batch.append(face)
+		mel_batch.append(m)
+		frame_batch.append(frame_to_save)
+		coords_batch.append(coords)
+
+		if len(img_batch) >= args.wav2lip_batch_size:
+			img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
+
+			img_masked = img_batch.copy()
+			img_masked[:, args.img_size//2:] = 0
+
+			img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
+			mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
+
+			yield img_batch, mel_batch, frame_batch, coords_batch
+			img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
+
+	if len(img_batch) > 0:
+		img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
+
+		img_masked = img_batch.copy()
+		img_masked[:, args.img_size//2:] = 0
+
+		img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
+		mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
+
+		yield img_batch, mel_batch, frame_batch, coords_batch
+
+def increase_frames(frames, l):
+	## evenly duplicating frames to increase length of video
+	while len(frames) < l:
+		dup_every = float(l) / len(frames)
+
+		final_frames = []
+		next_duplicate = 0.
+
+		for i, f in enumerate(frames):
+			final_frames.append(f)
+
+			if int(np.ceil(next_duplicate)) == i:
+				final_frames.append(f)
+
+			next_duplicate += dup_every
+
+		frames = final_frames
+
+	return frames[:l]
+
+mel_step_size = 16
+device = 'cuda' if torch.cuda.is_available() else 'cpu'
+print('Using {} for inference.'.format(device))
+
+detector = face_detection.FaceAlignment(face_detection.LandmarksType._2D, 
+											flip_input=False, device=device)
+
+def _load(checkpoint_path):
+	if device == 'cuda':
+		checkpoint = torch.load(checkpoint_path)
+	else:
+		checkpoint = torch.load(checkpoint_path,
+								map_location=lambda storage, loc: storage)
+	return checkpoint
+
+def load_model(path):
+	model = Wav2Lip()
+	print("Load checkpoint from: {}".format(path))
+	checkpoint = _load(path)
+	s = checkpoint["state_dict"]
+	new_s = {}
+	for k, v in s.items():
+		new_s[k.replace('module.', '')] = v
+	model.load_state_dict(new_s)
+
+	model = model.to(device)
+	return model.eval()
+
+model = load_model(args.checkpoint_path)
+
+def main():
+	if not os.path.isdir(args.results_dir): os.makedirs(args.results_dir)
+
+	if args.mode == 'dubbed':
+		files = listdir(args.data_root)
+		lines = ['{} {}'.format(f, f) for f in files]
+
+	else:
+		assert args.filelist is not None
+		with open(args.filelist, 'r') as filelist:
+			lines = filelist.readlines()
+
+	for idx, line in enumerate(tqdm(lines)):
+		video, audio_src = line.strip().split()
+
+		audio_src = os.path.join(args.data_root, audio_src)
+		video = os.path.join(args.data_root, video)
+
+		command = 'ffmpeg -loglevel panic -y -i {} -strict -2 {}'.format(audio_src, '../temp/temp.wav')
+		subprocess.call(command, shell=True)
+		temp_audio = '../temp/temp.wav'
+
+		wav = audio.load_wav(temp_audio, 16000)
+		mel = audio.melspectrogram(wav)
+
+		if np.isnan(mel.reshape(-1)).sum() > 0:
+			raise ValueError('Mel contains nan!')
+
+		video_stream = cv2.VideoCapture(video)
+
+		fps = video_stream.get(cv2.CAP_PROP_FPS)
+		mel_idx_multiplier = 80./fps
+
+		full_frames = []
+		while 1:
+			still_reading, frame = video_stream.read()
+			if not still_reading:
+				video_stream.release()
+				break
+
+			if min(frame.shape[:-1]) > args.max_frame_res:
+				h, w = frame.shape[:-1]
+				scale_factor = min(h, w) / float(args.max_frame_res)
+				h = int(h/scale_factor)
+				w = int(w/scale_factor)
+
+				frame = cv2.resize(frame, (w, h))
+			full_frames.append(frame)
+
+		mel_chunks = []
+		i = 0
+		while 1:
+			start_idx = int(i * mel_idx_multiplier)
+			if start_idx + mel_step_size > len(mel[0]):
+				break
+			mel_chunks.append(mel[:, start_idx : start_idx + mel_step_size])
+			i += 1
+
+		if len(full_frames) < len(mel_chunks):
+			if args.mode == 'tts':
+				full_frames = increase_frames(full_frames, len(mel_chunks))
+			else:
+				raise ValueError('#Frames, audio length mismatch')
+
+		else:
+			full_frames = full_frames[:len(mel_chunks)]
+
+		try:
+			face_det_results, full_frames = face_detect(full_frames.copy())
+		except ValueError as e:
+			continue
+
+		batch_size = args.wav2lip_batch_size
+		gen = datagen(full_frames.copy(), face_det_results, mel_chunks)
+
+		for i, (img_batch, mel_batch, frames, coords) in enumerate(gen):
+			if i == 0:
+				frame_h, frame_w = full_frames[0].shape[:-1]
+
+				out = cv2.VideoWriter('../temp/result.avi', 
+								cv2.VideoWriter_fourcc(*'DIVX'), fps, (frame_w, frame_h))
+
+			img_batch = torch.FloatTensor(np.transpose(img_batch, (0, 3, 1, 2))).to(device)
+			mel_batch = torch.FloatTensor(np.transpose(mel_batch, (0, 3, 1, 2))).to(device)
+
+			with torch.no_grad():
+				pred = model(mel_batch, img_batch)
+					
+
+			pred = pred.cpu().numpy().transpose(0, 2, 3, 1) * 255.
+			
+			for pl, f, c in zip(pred, frames, coords):
+				y1, y2, x1, x2 = c
+				pl = cv2.resize(pl.astype(np.uint8), (x2 - x1, y2 - y1))
+				f[y1:y2, x1:x2] = pl
+				out.write(f)
+
+		out.release()
+
+		vid = os.path.join(args.results_dir, '{}.mp4'.format(idx))
+		command = 'ffmpeg -loglevel panic -y -i {} -i {} -strict -2 -q:v 1 {}'.format('../temp/temp.wav', 
+								'../temp/result.avi', vid)
+		subprocess.call(command, shell=True)
+
+
+if __name__ == '__main__':
+	main()

evaluation/scores_LSE/SyncNetInstance_calc_scores.py

@@ -0,0 +1,210 @@
+#!/usr/bin/python
+#-*- coding: utf-8 -*-
+# Video 25 FPS, Audio 16000HZ
+
+import torch
+import numpy
+import time, pdb, argparse, subprocess, os, math, glob
+import cv2
+import python_speech_features
+
+from scipy import signal
+from scipy.io import wavfile
+from SyncNetModel import *
+from shutil import rmtree
+
+
+# ==================== Get OFFSET ====================
+
+def calc_pdist(feat1, feat2, vshift=10):
+    
+    win_size = vshift*2+1
+
+    feat2p = torch.nn.functional.pad(feat2,(0,0,vshift,vshift))
+
+    dists = []
+
+    for i in range(0,len(feat1)):
+
+        dists.append(torch.nn.functional.pairwise_distance(feat1[[i],:].repeat(win_size, 1), feat2p[i:i+win_size,:]))
+
+    return dists
+
+# ==================== MAIN DEF ====================
+
+class SyncNetInstance(torch.nn.Module):
+
+    def __init__(self, dropout = 0, num_layers_in_fc_layers = 1024):
+        super(SyncNetInstance, self).__init__();
+
+        self.__S__ = S(num_layers_in_fc_layers = num_layers_in_fc_layers).cuda();
+
+    def evaluate(self, opt, videofile):
+
+        self.__S__.eval();
+
+        # ========== ==========
+        # Convert files
+        # ========== ==========
+
+        if os.path.exists(os.path.join(opt.tmp_dir,opt.reference)):
+          rmtree(os.path.join(opt.tmp_dir,opt.reference))
+
+        os.makedirs(os.path.join(opt.tmp_dir,opt.reference))
+
+        command = ("ffmpeg -loglevel error -y -i %s -threads 1 -f image2 %s" % (videofile,os.path.join(opt.tmp_dir,opt.reference,'%06d.jpg'))) 
+        output = subprocess.call(command, shell=True, stdout=None)
+
+        command = ("ffmpeg -loglevel error -y -i %s -async 1 -ac 1 -vn -acodec pcm_s16le -ar 16000 %s" % (videofile,os.path.join(opt.tmp_dir,opt.reference,'audio.wav'))) 
+        output = subprocess.call(command, shell=True, stdout=None)
+        
+        # ========== ==========
+        # Load video 
+        # ========== ==========
+
+        images = []
+        
+        flist = glob.glob(os.path.join(opt.tmp_dir,opt.reference,'*.jpg'))
+        flist.sort()
+
+        for fname in flist:
+            img_input = cv2.imread(fname)
+            img_input = cv2.resize(img_input, (224,224)) #HARD CODED, CHANGE BEFORE RELEASE
+            images.append(img_input)
+
+        im = numpy.stack(images,axis=3)
+        im = numpy.expand_dims(im,axis=0)
+        im = numpy.transpose(im,(0,3,4,1,2))
+
+        imtv = torch.autograd.Variable(torch.from_numpy(im.astype(float)).float())
+
+        # ========== ==========
+        # Load audio
+        # ========== ==========
+
+        sample_rate, audio = wavfile.read(os.path.join(opt.tmp_dir,opt.reference,'audio.wav'))
+        mfcc = zip(*python_speech_features.mfcc(audio,sample_rate))
+        mfcc = numpy.stack([numpy.array(i) for i in mfcc])
+
+        cc = numpy.expand_dims(numpy.expand_dims(mfcc,axis=0),axis=0)
+        cct = torch.autograd.Variable(torch.from_numpy(cc.astype(float)).float())
+
+        # ========== ==========
+        # Check audio and video input length
+        # ========== ==========
+
+        #if (float(len(audio))/16000) != (float(len(images))/25) :
+        #    print("WARNING: Audio (%.4fs) and video (%.4fs) lengths are different."%(float(len(audio))/16000,float(len(images))/25))
+
+        min_length = min(len(images),math.floor(len(audio)/640))
+        
+        # ========== ==========
+        # Generate video and audio feats
+        # ========== ==========
+
+        lastframe = min_length-5
+        im_feat = []
+        cc_feat = []
+
+        tS = time.time()
+        for i in range(0,lastframe,opt.batch_size):
+            
+            im_batch = [ imtv[:,:,vframe:vframe+5,:,:] for vframe in range(i,min(lastframe,i+opt.batch_size)) ]
+            im_in = torch.cat(im_batch,0)
+            im_out  = self.__S__.forward_lip(im_in.cuda());
+            im_feat.append(im_out.data.cpu())
+
+            cc_batch = [ cct[:,:,:,vframe*4:vframe*4+20] for vframe in range(i,min(lastframe,i+opt.batch_size)) ]
+            cc_in = torch.cat(cc_batch,0)
+            cc_out  = self.__S__.forward_aud(cc_in.cuda())
+            cc_feat.append(cc_out.data.cpu())
+
+        im_feat = torch.cat(im_feat,0)
+        cc_feat = torch.cat(cc_feat,0)
+
+        # ========== ==========
+        # Compute offset
+        # ========== ==========
+            
+        #print('Compute time %.3f sec.' % (time.time()-tS))
+
+        dists = calc_pdist(im_feat,cc_feat,vshift=opt.vshift)
+        mdist = torch.mean(torch.stack(dists,1),1)
+
+        minval, minidx = torch.min(mdist,0)
+
+        offset = opt.vshift-minidx
+        conf   = torch.median(mdist) - minval
+
+        fdist   = numpy.stack([dist[minidx].numpy() for dist in dists])
+        # fdist   = numpy.pad(fdist, (3,3), 'constant', constant_values=15)
+        fconf   = torch.median(mdist).numpy() - fdist
+        fconfm  = signal.medfilt(fconf,kernel_size=9)
+        
+        numpy.set_printoptions(formatter={'float': '{: 0.3f}'.format})
+        #print('Framewise conf: ')
+        #print(fconfm)
+        #print('AV offset: \t%d \nMin dist: \t%.3f\nConfidence: \t%.3f' % (offset,minval,conf))
+
+        dists_npy = numpy.array([ dist.numpy() for dist in dists ])
+        return offset.numpy(), conf.numpy(), minval.numpy()
+
+    def extract_feature(self, opt, videofile):
+
+        self.__S__.eval();
+        
+        # ========== ==========
+        # Load video 
+        # ========== ==========
+        cap = cv2.VideoCapture(videofile)
+
+        frame_num = 1;
+        images = []
+        while frame_num:
+            frame_num += 1
+            ret, image = cap.read()
+            if ret == 0:
+                break
+
+            images.append(image)
+
+        im = numpy.stack(images,axis=3)
+        im = numpy.expand_dims(im,axis=0)
+        im = numpy.transpose(im,(0,3,4,1,2))
+
+        imtv = torch.autograd.Variable(torch.from_numpy(im.astype(float)).float())
+        
+        # ========== ==========
+        # Generate video feats
+        # ========== ==========
+
+        lastframe = len(images)-4
+        im_feat = []
+
+        tS = time.time()
+        for i in range(0,lastframe,opt.batch_size):
+            
+            im_batch = [ imtv[:,:,vframe:vframe+5,:,:] for vframe in range(i,min(lastframe,i+opt.batch_size)) ]
+            im_in = torch.cat(im_batch,0)
+            im_out  = self.__S__.forward_lipfeat(im_in.cuda());
+            im_feat.append(im_out.data.cpu())
+
+        im_feat = torch.cat(im_feat,0)
+
+        # ========== ==========
+        # Compute offset
+        # ========== ==========
+            
+        print('Compute time %.3f sec.' % (time.time()-tS))
+
+        return im_feat
+
+
+    def loadParameters(self, path):
+        loaded_state = torch.load(path, map_location=lambda storage, loc: storage);
+
+        self_state = self.__S__.state_dict();
+
+        for name, param in loaded_state.items():
+
+            self_state[name].copy_(param);

evaluation/scores_LSE/calculate_scores_LRS.py

@@ -0,0 +1,53 @@
+#!/usr/bin/python
+#-*- coding: utf-8 -*-
+
+import time, pdb, argparse, subprocess
+import glob
+import os
+from tqdm import tqdm
+
+from SyncNetInstance_calc_scores import *
+
+# ==================== LOAD PARAMS ====================
+
+
+parser = argparse.ArgumentParser(description = "SyncNet");
+
+parser.add_argument('--initial_model', type=str, default="data/syncnet_v2.model", help='');
+parser.add_argument('--batch_size', type=int, default='20', help='');
+parser.add_argument('--vshift', type=int, default='15', help='');
+parser.add_argument('--data_root', type=str, required=True, help='');
+parser.add_argument('--tmp_dir', type=str, default="data/work/pytmp", help='');
+parser.add_argument('--reference', type=str, default="demo", help='');
+
+opt = parser.parse_args();
+
+
+# ==================== RUN EVALUATION ====================
+
+s = SyncNetInstance();
+
+s.loadParameters(opt.initial_model);
+#print("Model %s loaded."%opt.initial_model);
+path = os.path.join(opt.data_root, "*.mp4")
+
+all_videos = glob.glob(path)
+
+prog_bar = tqdm(range(len(all_videos)))
+avg_confidence = 0.
+avg_min_distance = 0.
+
+
+for videofile_idx in prog_bar:
+	videofile = all_videos[videofile_idx]
+	offset, confidence, min_distance = s.evaluate(opt, videofile=videofile)
+	avg_confidence += confidence
+	avg_min_distance += min_distance
+	prog_bar.set_description('Avg Confidence: {}, Avg Minimum Dist: {}'.format(round(avg_confidence / (videofile_idx + 1), 3), round(avg_min_distance / (videofile_idx + 1), 3)))
+	prog_bar.refresh()
+
+print ('Average Confidence: {}'.format(avg_confidence/len(all_videos)))
+print ('Average Minimum Distance: {}'.format(avg_min_distance/len(all_videos)))
+
+
+

evaluation/scores_LSE/calculate_scores_real_videos.py

@@ -0,0 +1,45 @@
+#!/usr/bin/python
+#-*- coding: utf-8 -*-
+
+import time, pdb, argparse, subprocess, pickle, os, gzip, glob
+
+from SyncNetInstance_calc_scores import *
+
+# ==================== PARSE ARGUMENT ====================
+
+parser = argparse.ArgumentParser(description = "SyncNet");
+parser.add_argument('--initial_model', type=str, default="data/syncnet_v2.model", help='');
+parser.add_argument('--batch_size', type=int, default='20', help='');
+parser.add_argument('--vshift', type=int, default='15', help='');
+parser.add_argument('--data_dir', type=str, default='data/work', help='');
+parser.add_argument('--videofile', type=str, default='', help='');
+parser.add_argument('--reference', type=str, default='', help='');
+opt = parser.parse_args();
+
+setattr(opt,'avi_dir',os.path.join(opt.data_dir,'pyavi'))
+setattr(opt,'tmp_dir',os.path.join(opt.data_dir,'pytmp'))
+setattr(opt,'work_dir',os.path.join(opt.data_dir,'pywork'))
+setattr(opt,'crop_dir',os.path.join(opt.data_dir,'pycrop'))
+
+
+# ==================== LOAD MODEL AND FILE LIST ====================
+
+s = SyncNetInstance();
+
+s.loadParameters(opt.initial_model);
+#print("Model %s loaded."%opt.initial_model);
+
+flist = glob.glob(os.path.join(opt.crop_dir,opt.reference,'0*.avi'))
+flist.sort()
+
+# ==================== GET OFFSETS ====================
+
+dists = []
+for idx, fname in enumerate(flist):
+    offset, conf, dist = s.evaluate(opt,videofile=fname)
+    print (str(dist)+" "+str(conf))
+      
+# ==================== PRINT RESULTS TO FILE ====================
+
+#with open(os.path.join(opt.work_dir,opt.reference,'activesd.pckl'), 'wb') as fil:
+#    pickle.dump(dists, fil)

evaluation/scores_LSE/calculate_scores_real_videos.sh

@@ -0,0 +1,8 @@
+rm all_scores.txt
+yourfilenames=`ls $1`
+
+for eachfile in $yourfilenames
+do
+   python run_pipeline.py --videofile $1/$eachfile --reference wav2lip --data_dir tmp_dir
+   python calculate_scores_real_videos.py --videofile $1/$eachfile --reference wav2lip --data_dir tmp_dir >> all_scores.txt
+done

evaluation/test_filelists/README.md

@@ -0,0 +1,13 @@
+This folder contains the filelists for the new evaluation framework proposed in the paper. 
+
+## Test filelists for LRS2, LRS3, and LRW.
+
+This folder contains three filelists, each containing a list of names of audio-video pairs from the test sets of LRS2, LRS3, and LRW. The LRS2 and LRW filelists are strictly "Copyright BBC" and can only be used for “non-commercial research by applicants who have an agreement with the BBC to access the Lip Reading in the Wild and/or Lip Reading Sentences in the Wild datasets”. Please follow this link for more details: [https://www.bbc.co.uk/rd/projects/lip-reading-datasets](https://www.bbc.co.uk/rd/projects/lip-reading-datasets). 
+
+
+## ReSynCED benchmark
+
+The sub-folder `ReSynCED` contains filelists for our own Real-world lip-Sync Evaluation Dataset (ReSyncED).
+
+
+#### Instructions on how to use the above two filelists are available in the README of the parent folder. 

evaluation/test_filelists/ReSyncED/random_pairs.txt

@@ -0,0 +1,160 @@
+sachin.mp4 emma_cropped.mp4
+sachin.mp4 mourinho.mp4
+sachin.mp4 elon.mp4
+sachin.mp4 messi2.mp4
+sachin.mp4 cr1.mp4
+sachin.mp4 sachin.mp4
+sachin.mp4 sg.mp4
+sachin.mp4 fergi.mp4
+sachin.mp4 spanish_lec1.mp4
+sachin.mp4 bush_small.mp4
+sachin.mp4 macca_cut.mp4
+sachin.mp4 ca_cropped.mp4
+sachin.mp4 lecun.mp4
+sachin.mp4 spanish_lec0.mp4
+srk.mp4 emma_cropped.mp4
+srk.mp4 mourinho.mp4
+srk.mp4 elon.mp4
+srk.mp4 messi2.mp4
+srk.mp4 cr1.mp4
+srk.mp4 srk.mp4
+srk.mp4 sachin.mp4
+srk.mp4 sg.mp4
+srk.mp4 fergi.mp4
+srk.mp4 spanish_lec1.mp4
+srk.mp4 bush_small.mp4
+srk.mp4 macca_cut.mp4
+srk.mp4 ca_cropped.mp4
+srk.mp4 guardiola.mp4
+srk.mp4 lecun.mp4
+srk.mp4 spanish_lec0.mp4
+cr1.mp4 emma_cropped.mp4
+cr1.mp4 elon.mp4
+cr1.mp4 messi2.mp4
+cr1.mp4 cr1.mp4
+cr1.mp4 spanish_lec1.mp4
+cr1.mp4 bush_small.mp4
+cr1.mp4 macca_cut.mp4
+cr1.mp4 ca_cropped.mp4
+cr1.mp4 lecun.mp4
+cr1.mp4 spanish_lec0.mp4
+macca_cut.mp4 emma_cropped.mp4
+macca_cut.mp4 elon.mp4
+macca_cut.mp4 messi2.mp4
+macca_cut.mp4 spanish_lec1.mp4
+macca_cut.mp4 macca_cut.mp4
+macca_cut.mp4 ca_cropped.mp4
+macca_cut.mp4 spanish_lec0.mp4
+lecun.mp4 emma_cropped.mp4
+lecun.mp4 elon.mp4
+lecun.mp4 messi2.mp4
+lecun.mp4 spanish_lec1.mp4
+lecun.mp4 macca_cut.mp4
+lecun.mp4 ca_cropped.mp4
+lecun.mp4 lecun.mp4
+lecun.mp4 spanish_lec0.mp4
+messi2.mp4 emma_cropped.mp4
+messi2.mp4 elon.mp4
+messi2.mp4 messi2.mp4
+messi2.mp4 spanish_lec1.mp4
+messi2.mp4 macca_cut.mp4
+messi2.mp4 ca_cropped.mp4
+messi2.mp4 spanish_lec0.mp4
+ca_cropped.mp4 emma_cropped.mp4
+ca_cropped.mp4 elon.mp4
+ca_cropped.mp4 spanish_lec1.mp4
+ca_cropped.mp4 ca_cropped.mp4
+ca_cropped.mp4 spanish_lec0.mp4
+spanish_lec1.mp4 spanish_lec1.mp4
+spanish_lec1.mp4 spanish_lec0.mp4
+elon.mp4 elon.mp4
+elon.mp4 spanish_lec1.mp4
+elon.mp4 spanish_lec0.mp4
+guardiola.mp4 emma_cropped.mp4
+guardiola.mp4 mourinho.mp4
+guardiola.mp4 elon.mp4
+guardiola.mp4 messi2.mp4
+guardiola.mp4 cr1.mp4
+guardiola.mp4 sachin.mp4
+guardiola.mp4 sg.mp4
+guardiola.mp4 fergi.mp4
+guardiola.mp4 spanish_lec1.mp4
+guardiola.mp4 bush_small.mp4
+guardiola.mp4 macca_cut.mp4
+guardiola.mp4 ca_cropped.mp4
+guardiola.mp4 guardiola.mp4
+guardiola.mp4 lecun.mp4
+guardiola.mp4 spanish_lec0.mp4
+fergi.mp4 emma_cropped.mp4
+fergi.mp4 mourinho.mp4
+fergi.mp4 elon.mp4
+fergi.mp4 messi2.mp4
+fergi.mp4 cr1.mp4
+fergi.mp4 sachin.mp4
+fergi.mp4 sg.mp4
+fergi.mp4 fergi.mp4
+fergi.mp4 spanish_lec1.mp4
+fergi.mp4 bush_small.mp4
+fergi.mp4 macca_cut.mp4
+fergi.mp4 ca_cropped.mp4
+fergi.mp4 lecun.mp4
+fergi.mp4 spanish_lec0.mp4
+spanish.mp4 emma_cropped.mp4
+spanish.mp4 spanish.mp4
+spanish.mp4 mourinho.mp4
+spanish.mp4 elon.mp4
+spanish.mp4 messi2.mp4
+spanish.mp4 cr1.mp4
+spanish.mp4 srk.mp4
+spanish.mp4 sachin.mp4
+spanish.mp4 sg.mp4
+spanish.mp4 fergi.mp4
+spanish.mp4 spanish_lec1.mp4
+spanish.mp4 bush_small.mp4
+spanish.mp4 macca_cut.mp4
+spanish.mp4 ca_cropped.mp4
+spanish.mp4 guardiola.mp4
+spanish.mp4 lecun.mp4
+spanish.mp4 spanish_lec0.mp4
+bush_small.mp4 emma_cropped.mp4
+bush_small.mp4 elon.mp4
+bush_small.mp4 messi2.mp4
+bush_small.mp4 spanish_lec1.mp4
+bush_small.mp4 bush_small.mp4
+bush_small.mp4 macca_cut.mp4
+bush_small.mp4 ca_cropped.mp4
+bush_small.mp4 lecun.mp4
+bush_small.mp4 spanish_lec0.mp4
+emma_cropped.mp4 emma_cropped.mp4
+emma_cropped.mp4 elon.mp4
+emma_cropped.mp4 spanish_lec1.mp4
+emma_cropped.mp4 spanish_lec0.mp4
+sg.mp4 emma_cropped.mp4
+sg.mp4 mourinho.mp4
+sg.mp4 elon.mp4
+sg.mp4 messi2.mp4
+sg.mp4 cr1.mp4
+sg.mp4 sachin.mp4
+sg.mp4 sg.mp4
+sg.mp4 fergi.mp4
+sg.mp4 spanish_lec1.mp4
+sg.mp4 bush_small.mp4
+sg.mp4 macca_cut.mp4
+sg.mp4 ca_cropped.mp4
+sg.mp4 lecun.mp4
+sg.mp4 spanish_lec0.mp4
+spanish_lec0.mp4 spanish_lec0.mp4
+mourinho.mp4 emma_cropped.mp4
+mourinho.mp4 mourinho.mp4
+mourinho.mp4 elon.mp4
+mourinho.mp4 messi2.mp4
+mourinho.mp4 cr1.mp4
+mourinho.mp4 sachin.mp4
+mourinho.mp4 sg.mp4
+mourinho.mp4 fergi.mp4
+mourinho.mp4 spanish_lec1.mp4
+mourinho.mp4 bush_small.mp4
+mourinho.mp4 macca_cut.mp4
+mourinho.mp4 ca_cropped.mp4
+mourinho.mp4 lecun.mp4
+mourinho.mp4 spanish_lec0.mp4

evaluation/test_filelists/ReSyncED/tts_pairs.txt

@@ -0,0 +1,18 @@
+adam_1.mp4 andreng_optimization.wav
+agad_2.mp4 agad_2.wav
+agad_1.mp4 agad_1.wav
+agad_3.mp4 agad_3.wav
+rms_prop_1.mp4 rms_prop_tts.wav
+tf_1.mp4 tf_1.wav
+tf_2.mp4 tf_2.wav
+andrew_ng_ai_business.mp4 andrewng_business_tts.wav
+covid_autopsy_1.mp4 autopsy_tts.wav
+news_1.mp4 news_tts.wav
+andrew_ng_fund_1.mp4 andrewng_ai_fund.wav
+covid_treatments_1.mp4 covid_tts.wav
+pytorch_v_tf.mp4 pytorch_vs_tf_eng.wav
+pytorch_1.mp4 pytorch.wav
+pkb_1.mp4 pkb_1.wav
+ss_1.mp4 ss_1.wav
+carlsen_1.mp4 carlsen_eng.wav
+french.mp4 french.wav

face_detection/README.md

@@ -0,0 +1 @@
+The code for Face Detection in this folder has been taken from the wonderful [face_alignment](https://github.com/1adrianb/face-alignment) repository. This has been modified to take batches of faces at a time. 

face_detection/__init__.py

@@ -0,0 +1,7 @@
+# -*- coding: utf-8 -*-
+
+__author__ = """Adrian Bulat"""
+__email__ = '[email protected]'
+__version__ = '1.0.1'
+
+from .api import FaceAlignment, LandmarksType, NetworkSize

face_detection/api.py

@@ -0,0 +1,79 @@
+from __future__ import print_function
+import os
+import torch
+from torch.utils.model_zoo import load_url
+from enum import Enum
+import numpy as np
+import cv2
+try:
+    import urllib.request as request_file
+except BaseException:
+    import urllib as request_file
+
+from .models import FAN, ResNetDepth
+from .utils import *
+
+
+class LandmarksType(Enum):
+    """Enum class defining the type of landmarks to detect.
+
+    ``_2D`` - the detected points ``(x,y)`` are detected in a 2D space and follow the visible contour of the face
+    ``_2halfD`` - this points represent the projection of the 3D points into 3D
+    ``_3D`` - detect the points ``(x,y,z)``` in a 3D space
+
+    """
+    _2D = 1
+    _2halfD = 2
+    _3D = 3
+
+
+class NetworkSize(Enum):
+    # TINY = 1
+    # SMALL = 2
+    # MEDIUM = 3
+    LARGE = 4
+
+    def __new__(cls, value):
+        member = object.__new__(cls)
+        member._value_ = value
+        return member
+
+    def __int__(self):
+        return self.value
+
+ROOT = os.path.dirname(os.path.abspath(__file__))
+
+class FaceAlignment:
+    def __init__(self, landmarks_type, network_size=NetworkSize.LARGE,
+                 device='cuda', flip_input=False, face_detector='sfd', verbose=False):
+        self.device = device
+        self.flip_input = flip_input
+        self.landmarks_type = landmarks_type
+        self.verbose = verbose
+
+        network_size = int(network_size)
+
+        if 'cuda' in device:
+            torch.backends.cudnn.benchmark = True
+
+        # Get the face detector
+        face_detector_module = __import__('face_detection.detection.' + face_detector,
+                                          globals(), locals(), [face_detector], 0)
+        self.face_detector = face_detector_module.FaceDetector(device=device, verbose=verbose)
+
+    def get_detections_for_batch(self, images):
+        images = images[..., ::-1]
+        detected_faces = self.face_detector.detect_from_batch(images.copy())
+        results = []
+
+        for i, d in enumerate(detected_faces):
+            if len(d) == 0:
+                results.append(None)
+                continue
+            d = d[0]
+            d = np.clip(d, 0, None)
+            
+            x1, y1, x2, y2 = map(int, d[:-1])
+            results.append((x1, y1, x2, y2))
+
+        return results

face_detection/detection/__init__.py

@@ -0,0 +1 @@
+from .core import FaceDetector

face_detection/detection/core.py

@@ -0,0 +1,130 @@
+import logging
+import glob
+from tqdm import tqdm
+import numpy as np
+import torch
+import cv2
+
+
+class FaceDetector(object):
+    """An abstract class representing a face detector.
+
+    Any other face detection implementation must subclass it. All subclasses
+    must implement ``detect_from_image``, that return a list of detected
+    bounding boxes. Optionally, for speed considerations detect from path is
+    recommended.
+    """
+
+    def __init__(self, device, verbose):
+        self.device = device
+        self.verbose = verbose
+
+        if verbose:
+            if 'cpu' in device:
+                logger = logging.getLogger(__name__)
+                logger.warning("Detection running on CPU, this may be potentially slow.")
+
+        if 'cpu' not in device and 'cuda' not in device:
+            if verbose:
+                logger.error("Expected values for device are: {cpu, cuda} but got: %s", device)
+            raise ValueError
+
+    def detect_from_image(self, tensor_or_path):
+        """Detects faces in a given image.
+
+        This function detects the faces present in a provided BGR(usually)
+        image. The input can be either the image itself or the path to it.
+
+        Arguments:
+            tensor_or_path {numpy.ndarray, torch.tensor or string} -- the path
+            to an image or the image itself.
+
+        Example::
+
+            >>> path_to_image = 'data/image_01.jpg'
+            ...   detected_faces = detect_from_image(path_to_image)
+            [A list of bounding boxes (x1, y1, x2, y2)]
+            >>> image = cv2.imread(path_to_image)
+            ...   detected_faces = detect_from_image(image)
+            [A list of bounding boxes (x1, y1, x2, y2)]
+
+        """
+        raise NotImplementedError
+
+    def detect_from_directory(self, path, extensions=['.jpg', '.png'], recursive=False, show_progress_bar=True):
+        """Detects faces from all the images present in a given directory.
+
+        Arguments:
+            path {string} -- a string containing a path that points to the folder containing the images
+
+        Keyword Arguments:
+            extensions {list} -- list of string containing the extensions to be
+            consider in the following format: ``.extension_name`` (default:
+            {['.jpg', '.png']}) recursive {bool} -- option wherever to scan the
+            folder recursively (default: {False}) show_progress_bar {bool} --
+            display a progressbar (default: {True})
+
+        Example:
+        >>> directory = 'data'
+        ...   detected_faces = detect_from_directory(directory)
+        {A dictionary of [lists containing bounding boxes(x1, y1, x2, y2)]}
+
+        """
+        if self.verbose:
+            logger = logging.getLogger(__name__)
+
+        if len(extensions) == 0:
+            if self.verbose:
+                logger.error("Expected at list one extension, but none was received.")
+            raise ValueError
+
+        if self.verbose:
+            logger.info("Constructing the list of images.")
+        additional_pattern = '/**/*' if recursive else '/*'
+        files = []
+        for extension in extensions:
+            files.extend(glob.glob(path + additional_pattern + extension, recursive=recursive))
+
+        if self.verbose:
+            logger.info("Finished searching for images. %s images found", len(files))
+            logger.info("Preparing to run the detection.")
+
+        predictions = {}
+        for image_path in tqdm(files, disable=not show_progress_bar):
+            if self.verbose:
+                logger.info("Running the face detector on image: %s", image_path)
+            predictions[image_path] = self.detect_from_image(image_path)
+
+        if self.verbose:
+            logger.info("The detector was successfully run on all %s images", len(files))
+
+        return predictions
+
+    @property
+    def reference_scale(self):
+        raise NotImplementedError
+
+    @property
+    def reference_x_shift(self):
+        raise NotImplementedError
+
+    @property
+    def reference_y_shift(self):
+        raise NotImplementedError
+
+    @staticmethod
+    def tensor_or_path_to_ndarray(tensor_or_path, rgb=True):
+        """Convert path (represented as a string) or torch.tensor to a numpy.ndarray
+
+        Arguments:
+            tensor_or_path {numpy.ndarray, torch.tensor or string} -- path to the image, or the image itself
+        """
+        if isinstance(tensor_or_path, str):
+            return cv2.imread(tensor_or_path) if not rgb else cv2.imread(tensor_or_path)[..., ::-1]
+        elif torch.is_tensor(tensor_or_path):
+            # Call cpu in case its coming from cuda
+            return tensor_or_path.cpu().numpy()[..., ::-1].copy() if not rgb else tensor_or_path.cpu().numpy()
+        elif isinstance(tensor_or_path, np.ndarray):
+            return tensor_or_path[..., ::-1].copy() if not rgb else tensor_or_path
+        else:
+            raise TypeError

face_detection/detection/sfd/__init__.py

@@ -0,0 +1 @@
+from .sfd_detector import SFDDetector as FaceDetector

face_detection/detection/sfd/bbox.py

@@ -0,0 +1,129 @@
+from __future__ import print_function
+import os
+import sys
+import cv2
+import random
+import datetime
+import time
+import math
+import argparse
+import numpy as np
+import torch
+
+try:
+    from iou import IOU
+except BaseException:
+    # IOU cython speedup 10x
+    def IOU(ax1, ay1, ax2, ay2, bx1, by1, bx2, by2):
+        sa = abs((ax2 - ax1) * (ay2 - ay1))
+        sb = abs((bx2 - bx1) * (by2 - by1))
+        x1, y1 = max(ax1, bx1), max(ay1, by1)
+        x2, y2 = min(ax2, bx2), min(ay2, by2)
+        w = x2 - x1
+        h = y2 - y1
+        if w < 0 or h < 0:
+            return 0.0
+        else:
+            return 1.0 * w * h / (sa + sb - w * h)
+
+
+def bboxlog(x1, y1, x2, y2, axc, ayc, aww, ahh):
+    xc, yc, ww, hh = (x2 + x1) / 2, (y2 + y1) / 2, x2 - x1, y2 - y1
+    dx, dy = (xc - axc) / aww, (yc - ayc) / ahh
+    dw, dh = math.log(ww / aww), math.log(hh / ahh)
+    return dx, dy, dw, dh
+
+
+def bboxloginv(dx, dy, dw, dh, axc, ayc, aww, ahh):
+    xc, yc = dx * aww + axc, dy * ahh + ayc
+    ww, hh = math.exp(dw) * aww, math.exp(dh) * ahh
+    x1, x2, y1, y2 = xc - ww / 2, xc + ww / 2, yc - hh / 2, yc + hh / 2
+    return x1, y1, x2, y2
+
+
+def nms(dets, thresh):
+    if 0 == len(dets):
+        return []
+    x1, y1, x2, y2, scores = dets[:, 0], dets[:, 1], dets[:, 2], dets[:, 3], dets[:, 4]
+    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
+    order = scores.argsort()[::-1]
+
+    keep = []
+    while order.size > 0:
+        i = order[0]
+        keep.append(i)
+        xx1, yy1 = np.maximum(x1[i], x1[order[1:]]), np.maximum(y1[i], y1[order[1:]])
+        xx2, yy2 = np.minimum(x2[i], x2[order[1:]]), np.minimum(y2[i], y2[order[1:]])
+
+        w, h = np.maximum(0.0, xx2 - xx1 + 1), np.maximum(0.0, yy2 - yy1 + 1)
+        ovr = w * h / (areas[i] + areas[order[1:]] - w * h)
+
+        inds = np.where(ovr <= thresh)[0]
+        order = order[inds + 1]
+
+    return keep
+
+
+def encode(matched, priors, variances):
+    """Encode the variances from the priorbox layers into the ground truth boxes
+    we have matched (based on jaccard overlap) with the prior boxes.
+    Args:
+        matched: (tensor) Coords of ground truth for each prior in point-form
+            Shape: [num_priors, 4].
+        priors: (tensor) Prior boxes in center-offset form
+            Shape: [num_priors,4].
+        variances: (list[float]) Variances of priorboxes
+    Return:
+        encoded boxes (tensor), Shape: [num_priors, 4]
+    """
+
+    # dist b/t match center and prior's center
+    g_cxcy = (matched[:, :2] + matched[:, 2:]) / 2 - priors[:, :2]
+    # encode variance
+    g_cxcy /= (variances[0] * priors[:, 2:])
+    # match wh / prior wh
+    g_wh = (matched[:, 2:] - matched[:, :2]) / priors[:, 2:]
+    g_wh = torch.log(g_wh) / variances[1]
+    # return target for smooth_l1_loss
+    return torch.cat([g_cxcy, g_wh], 1)  # [num_priors,4]
+
+
+def decode(loc, priors, variances):
+    """Decode locations from predictions using priors to undo
+    the encoding we did for offset regression at train time.
+    Args:
+        loc (tensor): location predictions for loc layers,
+            Shape: [num_priors,4]
+        priors (tensor): Prior boxes in center-offset form.
+            Shape: [num_priors,4].
+        variances: (list[float]) Variances of priorboxes
+    Return:
+        decoded bounding box predictions
+    """
+
+    boxes = torch.cat((
+        priors[:, :2] + loc[:, :2] * variances[0] * priors[:, 2:],
+        priors[:, 2:] * torch.exp(loc[:, 2:] * variances[1])), 1)
+    boxes[:, :2] -= boxes[:, 2:] / 2
+    boxes[:, 2:] += boxes[:, :2]
+    return boxes
+
+def batch_decode(loc, priors, variances):
+    """Decode locations from predictions using priors to undo
+    the encoding we did for offset regression at train time.
+    Args:
+        loc (tensor): location predictions for loc layers,
+            Shape: [num_priors,4]
+        priors (tensor): Prior boxes in center-offset form.
+            Shape: [num_priors,4].
+        variances: (list[float]) Variances of priorboxes
+    Return:
+        decoded bounding box predictions
+    """
+
+    boxes = torch.cat((
+        priors[:, :, :2] + loc[:, :, :2] * variances[0] * priors[:, :, 2:],
+        priors[:, :, 2:] * torch.exp(loc[:, :, 2:] * variances[1])), 2)
+    boxes[:, :, :2] -= boxes[:, :, 2:] / 2
+    boxes[:, :, 2:] += boxes[:, :, :2]
+    return boxes

face_detection/detection/sfd/detect.py

@@ -0,0 +1,112 @@
+import torch
+import torch.nn.functional as F
+
+import os
+import sys
+import cv2
+import random
+import datetime
+import math
+import argparse
+import numpy as np
+
+import scipy.io as sio
+import zipfile
+from .net_s3fd import s3fd
+from .bbox import *
+
+
+def detect(net, img, device):
+    img = img - np.array([104, 117, 123])
+    img = img.transpose(2, 0, 1)
+    img = img.reshape((1,) + img.shape)
+
+    if 'cuda' in device:
+        torch.backends.cudnn.benchmark = True
+
+    img = torch.from_numpy(img).float().to(device)
+    BB, CC, HH, WW = img.size()
+    with torch.no_grad():
+        olist = net(img)
+
+    bboxlist = []
+    for i in range(len(olist) // 2):
+        olist[i * 2] = F.softmax(olist[i * 2], dim=1)
+    olist = [oelem.data.cpu() for oelem in olist]
+    for i in range(len(olist) // 2):
+        ocls, oreg = olist[i * 2], olist[i * 2 + 1]
+        FB, FC, FH, FW = ocls.size()  # feature map size
+        stride = 2**(i + 2)    # 4,8,16,32,64,128
+        anchor = stride * 4
+        poss = zip(*np.where(ocls[:, 1, :, :] > 0.05))
+        for Iindex, hindex, windex in poss:
+            axc, ayc = stride / 2 + windex * stride, stride / 2 + hindex * stride
+            score = ocls[0, 1, hindex, windex]
+            loc = oreg[0, :, hindex, windex].contiguous().view(1, 4)
+            priors = torch.Tensor([[axc / 1.0, ayc / 1.0, stride * 4 / 1.0, stride * 4 / 1.0]])
+            variances = [0.1, 0.2]
+            box = decode(loc, priors, variances)
+            x1, y1, x2, y2 = box[0] * 1.0
+            # cv2.rectangle(imgshow,(int(x1),int(y1)),(int(x2),int(y2)),(0,0,255),1)
+            bboxlist.append([x1, y1, x2, y2, score])
+    bboxlist = np.array(bboxlist)
+    if 0 == len(bboxlist):
+        bboxlist = np.zeros((1, 5))
+
+    return bboxlist
+
+def batch_detect(net, imgs, device):
+    imgs = imgs - np.array([104, 117, 123])
+    imgs = imgs.transpose(0, 3, 1, 2)
+
+    if 'cuda' in device:
+        torch.backends.cudnn.benchmark = True
+
+    imgs = torch.from_numpy(imgs).float().to(device)
+    BB, CC, HH, WW = imgs.size()
+    with torch.no_grad():
+        olist = net(imgs)
+
+    bboxlist = []
+    for i in range(len(olist) // 2):
+        olist[i * 2] = F.softmax(olist[i * 2], dim=1)
+    olist = [oelem.data.cpu() for oelem in olist]
+    for i in range(len(olist) // 2):
+        ocls, oreg = olist[i * 2], olist[i * 2 + 1]
+        FB, FC, FH, FW = ocls.size()  # feature map size
+        stride = 2**(i + 2)    # 4,8,16,32,64,128
+        anchor = stride * 4
+        poss = zip(*np.where(ocls[:, 1, :, :] > 0.05))
+        for Iindex, hindex, windex in poss:
+            axc, ayc = stride / 2 + windex * stride, stride / 2 + hindex * stride
+            score = ocls[:, 1, hindex, windex]
+            loc = oreg[:, :, hindex, windex].contiguous().view(BB, 1, 4)
+            priors = torch.Tensor([[axc / 1.0, ayc / 1.0, stride * 4 / 1.0, stride * 4 / 1.0]]).view(1, 1, 4)
+            variances = [0.1, 0.2]
+            box = batch_decode(loc, priors, variances)
+            box = box[:, 0] * 1.0
+            # cv2.rectangle(imgshow,(int(x1),int(y1)),(int(x2),int(y2)),(0,0,255),1)
+            bboxlist.append(torch.cat([box, score.unsqueeze(1)], 1).cpu().numpy())
+    bboxlist = np.array(bboxlist)
+    if 0 == len(bboxlist):
+        bboxlist = np.zeros((1, BB, 5))
+
+    return bboxlist
+
+def flip_detect(net, img, device):
+    img = cv2.flip(img, 1)
+    b = detect(net, img, device)
+
+    bboxlist = np.zeros(b.shape)
+    bboxlist[:, 0] = img.shape[1] - b[:, 2]
+    bboxlist[:, 1] = b[:, 1]
+    bboxlist[:, 2] = img.shape[1] - b[:, 0]
+    bboxlist[:, 3] = b[:, 3]
+    bboxlist[:, 4] = b[:, 4]
+    return bboxlist
+
+
+def pts_to_bb(pts):
+    min_x, min_y = np.min(pts, axis=0)
+    max_x, max_y = np.max(pts, axis=0)
+    return np.array([min_x, min_y, max_x, max_y])

face_detection/detection/sfd/net_s3fd.py

@@ -0,0 +1,129 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class L2Norm(nn.Module):
+    def __init__(self, n_channels, scale=1.0):
+        super(L2Norm, self).__init__()
+        self.n_channels = n_channels
+        self.scale = scale
+        self.eps = 1e-10
+        self.weight = nn.Parameter(torch.Tensor(self.n_channels))
+        self.weight.data *= 0.0
+        self.weight.data += self.scale
+
+    def forward(self, x):
+        norm = x.pow(2).sum(dim=1, keepdim=True).sqrt() + self.eps
+        x = x / norm * self.weight.view(1, -1, 1, 1)
+        return x
+
+
+class s3fd(nn.Module):
+    def __init__(self):
+        super(s3fd, self).__init__()
+        self.conv1_1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1)
+        self.conv1_2 = nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1)
+
+        self.conv2_1 = nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1)
+        self.conv2_2 = nn.Conv2d(128, 128, kernel_size=3, stride=1, padding=1)
+
+        self.conv3_1 = nn.Conv2d(128, 256, kernel_size=3, stride=1, padding=1)
+        self.conv3_2 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
+        self.conv3_3 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
+
+        self.conv4_1 = nn.Conv2d(256, 512, kernel_size=3, stride=1, padding=1)
+        self.conv4_2 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
+        self.conv4_3 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
+
+        self.conv5_1 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
+        self.conv5_2 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
+        self.conv5_3 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
+
+        self.fc6 = nn.Conv2d(512, 1024, kernel_size=3, stride=1, padding=3)
+        self.fc7 = nn.Conv2d(1024, 1024, kernel_size=1, stride=1, padding=0)
+
+        self.conv6_1 = nn.Conv2d(1024, 256, kernel_size=1, stride=1, padding=0)
+        self.conv6_2 = nn.Conv2d(256, 512, kernel_size=3, stride=2, padding=1)
+
+        self.conv7_1 = nn.Conv2d(512, 128, kernel_size=1, stride=1, padding=0)
+        self.conv7_2 = nn.Conv2d(128, 256, kernel_size=3, stride=2, padding=1)
+
+        self.conv3_3_norm = L2Norm(256, scale=10)
+        self.conv4_3_norm = L2Norm(512, scale=8)
+        self.conv5_3_norm = L2Norm(512, scale=5)
+
+        self.conv3_3_norm_mbox_conf = nn.Conv2d(256, 4, kernel_size=3, stride=1, padding=1)
+        self.conv3_3_norm_mbox_loc = nn.Conv2d(256, 4, kernel_size=3, stride=1, padding=1)
+        self.conv4_3_norm_mbox_conf = nn.Conv2d(512, 2, kernel_size=3, stride=1, padding=1)
+        self.conv4_3_norm_mbox_loc = nn.Conv2d(512, 4, kernel_size=3, stride=1, padding=1)
+        self.conv5_3_norm_mbox_conf = nn.Conv2d(512, 2, kernel_size=3, stride=1, padding=1)
+        self.conv5_3_norm_mbox_loc = nn.Conv2d(512, 4, kernel_size=3, stride=1, padding=1)
+
+        self.fc7_mbox_conf = nn.Conv2d(1024, 2, kernel_size=3, stride=1, padding=1)
+        self.fc7_mbox_loc = nn.Conv2d(1024, 4, kernel_size=3, stride=1, padding=1)
+        self.conv6_2_mbox_conf = nn.Conv2d(512, 2, kernel_size=3, stride=1, padding=1)
+        self.conv6_2_mbox_loc = nn.Conv2d(512, 4, kernel_size=3, stride=1, padding=1)
+        self.conv7_2_mbox_conf = nn.Conv2d(256, 2, kernel_size=3, stride=1, padding=1)
+        self.conv7_2_mbox_loc = nn.Conv2d(256, 4, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, x):
+        h = F.relu(self.conv1_1(x))
+        h = F.relu(self.conv1_2(h))
+        h = F.max_pool2d(h, 2, 2)
+
+        h = F.relu(self.conv2_1(h))
+        h = F.relu(self.conv2_2(h))
+        h = F.max_pool2d(h, 2, 2)
+
+        h = F.relu(self.conv3_1(h))
+        h = F.relu(self.conv3_2(h))
+        h = F.relu(self.conv3_3(h))
+        f3_3 = h
+        h = F.max_pool2d(h, 2, 2)
+
+        h = F.relu(self.conv4_1(h))
+        h = F.relu(self.conv4_2(h))
+        h = F.relu(self.conv4_3(h))
+        f4_3 = h
+        h = F.max_pool2d(h, 2, 2)
+
+        h = F.relu(self.conv5_1(h))
+        h = F.relu(self.conv5_2(h))
+        h = F.relu(self.conv5_3(h))
+        f5_3 = h
+        h = F.max_pool2d(h, 2, 2)
+
+        h = F.relu(self.fc6(h))
+        h = F.relu(self.fc7(h))
+        ffc7 = h
+        h = F.relu(self.conv6_1(h))
+        h = F.relu(self.conv6_2(h))
+        f6_2 = h
+        h = F.relu(self.conv7_1(h))
+        h = F.relu(self.conv7_2(h))
+        f7_2 = h
+
+        f3_3 = self.conv3_3_norm(f3_3)
+        f4_3 = self.conv4_3_norm(f4_3)
+        f5_3 = self.conv5_3_norm(f5_3)
+
+        cls1 = self.conv3_3_norm_mbox_conf(f3_3)
+        reg1 = self.conv3_3_norm_mbox_loc(f3_3)
+        cls2 = self.conv4_3_norm_mbox_conf(f4_3)
+        reg2 = self.conv4_3_norm_mbox_loc(f4_3)
+        cls3 = self.conv5_3_norm_mbox_conf(f5_3)
+        reg3 = self.conv5_3_norm_mbox_loc(f5_3)
+        cls4 = self.fc7_mbox_conf(ffc7)
+        reg4 = self.fc7_mbox_loc(ffc7)
+        cls5 = self.conv6_2_mbox_conf(f6_2)
+        reg5 = self.conv6_2_mbox_loc(f6_2)
+        cls6 = self.conv7_2_mbox_conf(f7_2)
+        reg6 = self.conv7_2_mbox_loc(f7_2)
+
+        # max-out background label
+        chunk = torch.chunk(cls1, 4, 1)
+        bmax = torch.max(torch.max(chunk[0], chunk[1]), chunk[2])
+        cls1 = torch.cat([bmax, chunk[3]], dim=1)
+
+        return [cls1, reg1, cls2, reg2, cls3, reg3, cls4, reg4, cls5, reg5, cls6, reg6]

face_detection/detection/sfd/s3fd-619a316812.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:619a31681264d3f7f7fc7a16a42cbbe8b23f31a256f75a366e5a1bcd59b33543
+size 89843225

face_detection/detection/sfd/sfd_detector.py

@@ -0,0 +1,59 @@
+import os
+import cv2
+from torch.utils.model_zoo import load_url
+
+from ..core import FaceDetector
+
+from .net_s3fd import s3fd
+from .bbox import *
+from .detect import *
+
+models_urls = {
+    's3fd': 'https://www.adrianbulat.com/downloads/python-fan/s3fd-619a316812.pth',
+}
+
+
+class SFDDetector(FaceDetector):
+    def __init__(self, device, path_to_detector=os.path.join(os.path.dirname(os.path.abspath(__file__)), 's3fd.pth'), verbose=False):
+        super(SFDDetector, self).__init__(device, verbose)
+
+        # Initialise the face detector
+        if not os.path.isfile(path_to_detector):
+            model_weights = load_url(models_urls['s3fd'])
+        else:
+            model_weights = torch.load(path_to_detector)
+
+        self.face_detector = s3fd()
+        self.face_detector.load_state_dict(model_weights)
+        self.face_detector.to(device)
+        self.face_detector.eval()
+
+    def detect_from_image(self, tensor_or_path):
+        image = self.tensor_or_path_to_ndarray(tensor_or_path)
+
+        bboxlist = detect(self.face_detector, image, device=self.device)
+        keep = nms(bboxlist, 0.3)
+        bboxlist = bboxlist[keep, :]
+        bboxlist = [x for x in bboxlist if x[-1] > 0.5]
+
+        return bboxlist
+
+    def detect_from_batch(self, images):
+        bboxlists = batch_detect(self.face_detector, images, device=self.device)
+        keeps = [nms(bboxlists[:, i, :], 0.3) for i in range(bboxlists.shape[1])]
+        bboxlists = [bboxlists[keep, i, :] for i, keep in enumerate(keeps)]
+        bboxlists = [[x for x in bboxlist if x[-1] > 0.5] for bboxlist in bboxlists]
+
+        return bboxlists
+
+    @property
+    def reference_scale(self):
+        return 195
+
+    @property
+    def reference_x_shift(self):
+        return 0
+
+    @property
+    def reference_y_shift(self):
+        return 0

face_detection/models.py

@@ -0,0 +1,261 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import math
+
+
+def conv3x3(in_planes, out_planes, strd=1, padding=1, bias=False):
+    "3x3 convolution with padding"
+    return nn.Conv2d(in_planes, out_planes, kernel_size=3,
+                     stride=strd, padding=padding, bias=bias)
+
+
+class ConvBlock(nn.Module):
+    def __init__(self, in_planes, out_planes):
+        super(ConvBlock, self).__init__()
+        self.bn1 = nn.BatchNorm2d(in_planes)
+        self.conv1 = conv3x3(in_planes, int(out_planes / 2))
+        self.bn2 = nn.BatchNorm2d(int(out_planes / 2))
+        self.conv2 = conv3x3(int(out_planes / 2), int(out_planes / 4))
+        self.bn3 = nn.BatchNorm2d(int(out_planes / 4))
+        self.conv3 = conv3x3(int(out_planes / 4), int(out_planes / 4))
+
+        if in_planes != out_planes:
+            self.downsample = nn.Sequential(
+                nn.BatchNorm2d(in_planes),
+                nn.ReLU(True),
+                nn.Conv2d(in_planes, out_planes,
+                          kernel_size=1, stride=1, bias=False),
+            )
+        else:
+            self.downsample = None
+
+    def forward(self, x):
+        residual = x
+
+        out1 = self.bn1(x)
+        out1 = F.relu(out1, True)
+        out1 = self.conv1(out1)
+
+        out2 = self.bn2(out1)
+        out2 = F.relu(out2, True)
+        out2 = self.conv2(out2)
+
+        out3 = self.bn3(out2)
+        out3 = F.relu(out3, True)
+        out3 = self.conv3(out3)
+
+        out3 = torch.cat((out1, out2, out3), 1)
+
+        if self.downsample is not None:
+            residual = self.downsample(residual)
+
+        out3 += residual
+
+        return out3
+
+
+class Bottleneck(nn.Module):
+
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None):
+        super(Bottleneck, self).__init__()
+        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride,
+                               padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(planes * 4)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class HourGlass(nn.Module):
+    def __init__(self, num_modules, depth, num_features):
+        super(HourGlass, self).__init__()
+        self.num_modules = num_modules
+        self.depth = depth
+        self.features = num_features
+
+        self._generate_network(self.depth)
+
+    def _generate_network(self, level):
+        self.add_module('b1_' + str(level), ConvBlock(self.features, self.features))
+
+        self.add_module('b2_' + str(level), ConvBlock(self.features, self.features))
+
+        if level > 1:
+            self._generate_network(level - 1)
+        else:
+            self.add_module('b2_plus_' + str(level), ConvBlock(self.features, self.features))
+
+        self.add_module('b3_' + str(level), ConvBlock(self.features, self.features))
+
+    def _forward(self, level, inp):
+        # Upper branch
+        up1 = inp
+        up1 = self._modules['b1_' + str(level)](up1)
+
+        # Lower branch
+        low1 = F.avg_pool2d(inp, 2, stride=2)
+        low1 = self._modules['b2_' + str(level)](low1)
+
+        if level > 1:
+            low2 = self._forward(level - 1, low1)
+        else:
+            low2 = low1
+            low2 = self._modules['b2_plus_' + str(level)](low2)
+
+        low3 = low2
+        low3 = self._modules['b3_' + str(level)](low3)
+
+        up2 = F.interpolate(low3, scale_factor=2, mode='nearest')
+
+        return up1 + up2
+
+    def forward(self, x):
+        return self._forward(self.depth, x)
+
+
+class FAN(nn.Module):
+
+    def __init__(self, num_modules=1):
+        super(FAN, self).__init__()
+        self.num_modules = num_modules
+
+        # Base part
+        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3)
+        self.bn1 = nn.BatchNorm2d(64)
+        self.conv2 = ConvBlock(64, 128)
+        self.conv3 = ConvBlock(128, 128)
+        self.conv4 = ConvBlock(128, 256)
+
+        # Stacking part
+        for hg_module in range(self.num_modules):
+            self.add_module('m' + str(hg_module), HourGlass(1, 4, 256))
+            self.add_module('top_m_' + str(hg_module), ConvBlock(256, 256))
+            self.add_module('conv_last' + str(hg_module),
+                            nn.Conv2d(256, 256, kernel_size=1, stride=1, padding=0))
+            self.add_module('bn_end' + str(hg_module), nn.BatchNorm2d(256))
+            self.add_module('l' + str(hg_module), nn.Conv2d(256,
+                                                            68, kernel_size=1, stride=1, padding=0))
+
+            if hg_module < self.num_modules - 1:
+                self.add_module(
+                    'bl' + str(hg_module), nn.Conv2d(256, 256, kernel_size=1, stride=1, padding=0))
+                self.add_module('al' + str(hg_module), nn.Conv2d(68,
+                                                                 256, kernel_size=1, stride=1, padding=0))
+
+    def forward(self, x):
+        x = F.relu(self.bn1(self.conv1(x)), True)
+        x = F.avg_pool2d(self.conv2(x), 2, stride=2)
+        x = self.conv3(x)
+        x = self.conv4(x)
+
+        previous = x
+
+        outputs = []
+        for i in range(self.num_modules):
+            hg = self._modules['m' + str(i)](previous)
+
+            ll = hg
+            ll = self._modules['top_m_' + str(i)](ll)
+
+            ll = F.relu(self._modules['bn_end' + str(i)]
+                        (self._modules['conv_last' + str(i)](ll)), True)
+
+            # Predict heatmaps
+            tmp_out = self._modules['l' + str(i)](ll)
+            outputs.append(tmp_out)
+
+            if i < self.num_modules - 1:
+                ll = self._modules['bl' + str(i)](ll)
+                tmp_out_ = self._modules['al' + str(i)](tmp_out)
+                previous = previous + ll + tmp_out_
+
+        return outputs
+
+
+class ResNetDepth(nn.Module):
+
+    def __init__(self, block=Bottleneck, layers=[3, 8, 36, 3], num_classes=68):
+        self.inplanes = 64
+        super(ResNetDepth, self).__init__()
+        self.conv1 = nn.Conv2d(3 + 68, 64, kernel_size=7, stride=2, padding=3,
+                               bias=False)
+        self.bn1 = nn.BatchNorm2d(64)
+        self.relu = nn.ReLU(inplace=True)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
+        self.avgpool = nn.AvgPool2d(7)
+        self.fc = nn.Linear(512 * block.expansion, num_classes)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+                m.weight.data.normal_(0, math.sqrt(2. / n))
+            elif isinstance(m, nn.BatchNorm2d):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+
+    def _make_layer(self, block, planes, blocks, stride=1):
+        downsample = None
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                nn.Conv2d(self.inplanes, planes * block.expansion,
+                          kernel_size=1, stride=stride, bias=False),
+                nn.BatchNorm2d(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(block(self.inplanes, planes, stride, downsample))
+        self.inplanes = planes * block.expansion
+        for i in range(1, blocks):
+            layers.append(block(self.inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = self.avgpool(x)
+        x = x.view(x.size(0), -1)
+        x = self.fc(x)
+
+        return x

face_detection/utils.py

@@ -0,0 +1,313 @@
+from __future__ import print_function
+import os
+import sys
+import time
+import torch
+import math
+import numpy as np
+import cv2
+
+
+def _gaussian(
+        size=3, sigma=0.25, amplitude=1, normalize=False, width=None,
+        height=None, sigma_horz=None, sigma_vert=None, mean_horz=0.5,
+        mean_vert=0.5):
+    # handle some defaults
+    if width is None:
+        width = size
+    if height is None:
+        height = size
+    if sigma_horz is None:
+        sigma_horz = sigma
+    if sigma_vert is None:
+        sigma_vert = sigma
+    center_x = mean_horz * width + 0.5
+    center_y = mean_vert * height + 0.5
+    gauss = np.empty((height, width), dtype=np.float32)
+    # generate kernel
+    for i in range(height):
+        for j in range(width):
+            gauss[i][j] = amplitude * math.exp(-(math.pow((j + 1 - center_x) / (
+                sigma_horz * width), 2) / 2.0 + math.pow((i + 1 - center_y) / (sigma_vert * height), 2) / 2.0))
+    if normalize:
+        gauss = gauss / np.sum(gauss)
+    return gauss
+
+
+def draw_gaussian(image, point, sigma):
+    # Check if the gaussian is inside
+    ul = [math.floor(point[0] - 3 * sigma), math.floor(point[1] - 3 * sigma)]
+    br = [math.floor(point[0] + 3 * sigma), math.floor(point[1] + 3 * sigma)]
+    if (ul[0] > image.shape[1] or ul[1] > image.shape[0] or br[0] < 1 or br[1] < 1):
+        return image
+    size = 6 * sigma + 1
+    g = _gaussian(size)
+    g_x = [int(max(1, -ul[0])), int(min(br[0], image.shape[1])) - int(max(1, ul[0])) + int(max(1, -ul[0]))]
+    g_y = [int(max(1, -ul[1])), int(min(br[1], image.shape[0])) - int(max(1, ul[1])) + int(max(1, -ul[1]))]
+    img_x = [int(max(1, ul[0])), int(min(br[0], image.shape[1]))]
+    img_y = [int(max(1, ul[1])), int(min(br[1], image.shape[0]))]
+    assert (g_x[0] > 0 and g_y[1] > 0)
+    image[img_y[0] - 1:img_y[1], img_x[0] - 1:img_x[1]
+          ] = image[img_y[0] - 1:img_y[1], img_x[0] - 1:img_x[1]] + g[g_y[0] - 1:g_y[1], g_x[0] - 1:g_x[1]]
+    image[image > 1] = 1
+    return image
+
+
+def transform(point, center, scale, resolution, invert=False):
+    """Generate and affine transformation matrix.
+
+    Given a set of points, a center, a scale and a targer resolution, the
+    function generates and affine transformation matrix. If invert is ``True``
+    it will produce the inverse transformation.
+
+    Arguments:
+        point {torch.tensor} -- the input 2D point
+        center {torch.tensor or numpy.array} -- the center around which to perform the transformations
+        scale {float} -- the scale of the face/object
+        resolution {float} -- the output resolution
+
+    Keyword Arguments:
+        invert {bool} -- define wherever the function should produce the direct or the
+        inverse transformation matrix (default: {False})
+    """
+    _pt = torch.ones(3)
+    _pt[0] = point[0]
+    _pt[1] = point[1]
+
+    h = 200.0 * scale
+    t = torch.eye(3)
+    t[0, 0] = resolution / h
+    t[1, 1] = resolution / h
+    t[0, 2] = resolution * (-center[0] / h + 0.5)
+    t[1, 2] = resolution * (-center[1] / h + 0.5)
+
+    if invert:
+        t = torch.inverse(t)
+
+    new_point = (torch.matmul(t, _pt))[0:2]
+
+    return new_point.int()
+
+
+def crop(image, center, scale, resolution=256.0):
+    """Center crops an image or set of heatmaps
+
+    Arguments:
+        image {numpy.array} -- an rgb image
+        center {numpy.array} -- the center of the object, usually the same as of the bounding box
+        scale {float} -- scale of the face
+
+    Keyword Arguments:
+        resolution {float} -- the size of the output cropped image (default: {256.0})
+
+    Returns:
+        [type] -- [description]
+    """  # Crop around the center point
+    """ Crops the image around the center. Input is expected to be an np.ndarray """
+    ul = transform([1, 1], center, scale, resolution, True)
+    br = transform([resolution, resolution], center, scale, resolution, True)
+    # pad = math.ceil(torch.norm((ul - br).float()) / 2.0 - (br[0] - ul[0]) / 2.0)
+    if image.ndim > 2:
+        newDim = np.array([br[1] - ul[1], br[0] - ul[0],
+                           image.shape[2]], dtype=np.int32)
+        newImg = np.zeros(newDim, dtype=np.uint8)
+    else:
+        newDim = np.array([br[1] - ul[1], br[0] - ul[0]], dtype=np.int)
+        newImg = np.zeros(newDim, dtype=np.uint8)
+    ht = image.shape[0]
+    wd = image.shape[1]
+    newX = np.array(
+        [max(1, -ul[0] + 1), min(br[0], wd) - ul[0]], dtype=np.int32)
+    newY = np.array(
+        [max(1, -ul[1] + 1), min(br[1], ht) - ul[1]], dtype=np.int32)
+    oldX = np.array([max(1, ul[0] + 1), min(br[0], wd)], dtype=np.int32)
+    oldY = np.array([max(1, ul[1] + 1), min(br[1], ht)], dtype=np.int32)
+    newImg[newY[0] - 1:newY[1], newX[0] - 1:newX[1]
+           ] = image[oldY[0] - 1:oldY[1], oldX[0] - 1:oldX[1], :]
+    newImg = cv2.resize(newImg, dsize=(int(resolution), int(resolution)),
+                        interpolation=cv2.INTER_LINEAR)
+    return newImg
+
+
+def get_preds_fromhm(hm, center=None, scale=None):
+    """Obtain (x,y) coordinates given a set of N heatmaps. If the center
+    and the scale is provided the function will return the points also in
+    the original coordinate frame.
+
+    Arguments:
+        hm {torch.tensor} -- the predicted heatmaps, of shape [B, N, W, H]
+
+    Keyword Arguments:
+        center {torch.tensor} -- the center of the bounding box (default: {None})
+        scale {float} -- face scale (default: {None})
+    """
+    max, idx = torch.max(
+        hm.view(hm.size(0), hm.size(1), hm.size(2) * hm.size(3)), 2)
+    idx += 1
+    preds = idx.view(idx.size(0), idx.size(1), 1).repeat(1, 1, 2).float()
+    preds[..., 0].apply_(lambda x: (x - 1) % hm.size(3) + 1)
+    preds[..., 1].add_(-1).div_(hm.size(2)).floor_().add_(1)
+
+    for i in range(preds.size(0)):
+        for j in range(preds.size(1)):
+            hm_ = hm[i, j, :]
+            pX, pY = int(preds[i, j, 0]) - 1, int(preds[i, j, 1]) - 1
+            if pX > 0 and pX < 63 and pY > 0 and pY < 63:
+                diff = torch.FloatTensor(
+                    [hm_[pY, pX + 1] - hm_[pY, pX - 1],
+                     hm_[pY + 1, pX] - hm_[pY - 1, pX]])
+                preds[i, j].add_(diff.sign_().mul_(.25))
+
+    preds.add_(-.5)
+
+    preds_orig = torch.zeros(preds.size())
+    if center is not None and scale is not None:
+        for i in range(hm.size(0)):
+            for j in range(hm.size(1)):
+                preds_orig[i, j] = transform(
+                    preds[i, j], center, scale, hm.size(2), True)
+
+    return preds, preds_orig
+
+def get_preds_fromhm_batch(hm, centers=None, scales=None):
+    """Obtain (x,y) coordinates given a set of N heatmaps. If the centers
+    and the scales is provided the function will return the points also in
+    the original coordinate frame.
+
+    Arguments:
+        hm {torch.tensor} -- the predicted heatmaps, of shape [B, N, W, H]
+
+    Keyword Arguments:
+        centers {torch.tensor} -- the centers of the bounding box (default: {None})
+        scales {float} -- face scales (default: {None})
+    """
+    max, idx = torch.max(
+        hm.view(hm.size(0), hm.size(1), hm.size(2) * hm.size(3)), 2)
+    idx += 1
+    preds = idx.view(idx.size(0), idx.size(1), 1).repeat(1, 1, 2).float()
+    preds[..., 0].apply_(lambda x: (x - 1) % hm.size(3) + 1)
+    preds[..., 1].add_(-1).div_(hm.size(2)).floor_().add_(1)
+
+    for i in range(preds.size(0)):
+        for j in range(preds.size(1)):
+            hm_ = hm[i, j, :]
+            pX, pY = int(preds[i, j, 0]) - 1, int(preds[i, j, 1]) - 1
+            if pX > 0 and pX < 63 and pY > 0 and pY < 63:
+                diff = torch.FloatTensor(
+                    [hm_[pY, pX + 1] - hm_[pY, pX - 1],
+                     hm_[pY + 1, pX] - hm_[pY - 1, pX]])
+                preds[i, j].add_(diff.sign_().mul_(.25))
+
+    preds.add_(-.5)
+
+    preds_orig = torch.zeros(preds.size())
+    if centers is not None and scales is not None:
+        for i in range(hm.size(0)):
+            for j in range(hm.size(1)):
+                preds_orig[i, j] = transform(
+                    preds[i, j], centers[i], scales[i], hm.size(2), True)
+
+    return preds, preds_orig
+
+def shuffle_lr(parts, pairs=None):
+    """Shuffle the points left-right according to the axis of symmetry
+    of the object.
+
+    Arguments:
+        parts {torch.tensor} -- a 3D or 4D object containing the
+        heatmaps.
+
+    Keyword Arguments:
+        pairs {list of integers} -- [order of the flipped points] (default: {None})
+    """
+    if pairs is None:
+        pairs = [16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0,
+                 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 27, 28, 29, 30, 35,
+                 34, 33, 32, 31, 45, 44, 43, 42, 47, 46, 39, 38, 37, 36, 41,
+                 40, 54, 53, 52, 51, 50, 49, 48, 59, 58, 57, 56, 55, 64, 63,
+                 62, 61, 60, 67, 66, 65]
+    if parts.ndimension() == 3:
+        parts = parts[pairs, ...]
+    else:
+        parts = parts[:, pairs, ...]
+
+    return parts
+
+
+def flip(tensor, is_label=False):
+    """Flip an image or a set of heatmaps left-right
+
+    Arguments:
+        tensor {numpy.array or torch.tensor} -- [the input image or heatmaps]
+
+    Keyword Arguments:
+        is_label {bool} -- [denote wherever the input is an image or a set of heatmaps ] (default: {False})
+    """
+    if not torch.is_tensor(tensor):
+        tensor = torch.from_numpy(tensor)
+
+    if is_label:
+        tensor = shuffle_lr(tensor).flip(tensor.ndimension() - 1)
+    else:
+        tensor = tensor.flip(tensor.ndimension() - 1)
+
+    return tensor
+
+# From pyzolib/paths.py (https://bitbucket.org/pyzo/pyzolib/src/tip/paths.py)
+
+
+def appdata_dir(appname=None, roaming=False):
+    """ appdata_dir(appname=None, roaming=False)
+
+    Get the path to the application directory, where applications are allowed
+    to write user specific files (e.g. configurations). For non-user specific
+    data, consider using common_appdata_dir().
+    If appname is given, a subdir is appended (and created if necessary).
+    If roaming is True, will prefer a roaming directory (Windows Vista/7).
+    """
+
+    # Define default user directory
+    userDir = os.getenv('FACEALIGNMENT_USERDIR', None)
+    if userDir is None:
+        userDir = os.path.expanduser('~')
+        if not os.path.isdir(userDir):  # pragma: no cover
+            userDir = '/var/tmp'  # issue #54
+
+    # Get system app data dir
+    path = None
+    if sys.platform.startswith('win'):
+        path1, path2 = os.getenv('LOCALAPPDATA'), os.getenv('APPDATA')
+        path = (path2 or path1) if roaming else (path1 or path2)
+    elif sys.platform.startswith('darwin'):
+        path = os.path.join(userDir, 'Library', 'Application Support')
+    # On Linux and as fallback
+    if not (path and os.path.isdir(path)):
+        path = userDir
+
+    # Maybe we should store things local to the executable (in case of a
+    # portable distro or a frozen application that wants to be portable)
+    prefix = sys.prefix
+    if getattr(sys, 'frozen', None):
+        prefix = os.path.abspath(os.path.dirname(sys.executable))
+    for reldir in ('settings', '../settings'):
+        localpath = os.path.abspath(os.path.join(prefix, reldir))
+        if os.path.isdir(localpath):  # pragma: no cover
+            try:
+                open(os.path.join(localpath, 'test.write'), 'wb').close()
+                os.remove(os.path.join(localpath, 'test.write'))
+            except IOError:
+                pass  # We cannot write in this directory
+            else:
+                path = localpath
+                break
+
+    # Get path specific for this app
+    if appname:
+        if path == userDir:
+            appname = '.' + appname.lstrip('.')  # Make it a hidden directory
+        path = os.path.join(path, appname)
+        if not os.path.isdir(path):  # pragma: no cover
+            os.mkdir(path)
+
+    # Done
+    return path

filelists/README.md

@@ -0,0 +1 @@
+Place LRS2 (and any other) filelists here for training.

inference.py

@@ -0,0 +1,294 @@
+from os import listdir, path
+import numpy as np
+import scipy, cv2, os, sys, argparse, audio
+import json, subprocess, random, string
+from tqdm import tqdm
+from glob import glob
+import torch, face_detection
+from models import Wav2Lip
+import platform
+
+parser = argparse.ArgumentParser(description='Inference code to lip-sync videos in the wild using Wav2Lip models')
+
+parser.add_argument('--checkpoint_path', type=str, 
+					help='Name of saved checkpoint to load weights from', required=True)
+
+parser.add_argument('--face', type=str, 
+					help='Filepath of video/image that contains faces to use', required=True)
+parser.add_argument('--audio', type=str, 
+					help='Filepath of video/audio file to use as raw audio source', required=True)
+parser.add_argument('--outfile', type=str, help='Video path to save result. See default for an e.g.', 
+								default='results/result_voice.mp4')
+
+parser.add_argument('--static', type=bool, 
+					help='If True, then use only first video frame for inference', default=False)
+parser.add_argument('--fps', type=float, help='Can be specified only if input is a static image (default: 25)', 
+					default=25., required=False)
+
+parser.add_argument('--pads', nargs='+', type=int, default=[0, 10, 0, 0], 
+					help='Padding (top, bottom, left, right). Please adjust to include chin at least')
+
+parser.add_argument('--face_det_batch_size', type=int, 
+					help='Batch size for face detection', default=16)
+parser.add_argument('--wav2lip_batch_size', type=int, help='Batch size for Wav2Lip model(s)', default=128)
+
+parser.add_argument('--resize_factor', default=1, type=int, 
+			help='Reduce the resolution by this factor. Sometimes, best results are obtained at 480p or 720p')
+
+parser.add_argument('--crop', nargs='+', type=int, default=[0, -1, 0, -1], 
+					help='Crop video to a smaller region (top, bottom, left, right). Applied after resize_factor and rotate arg. ' 
+					'Useful if multiple face present. -1 implies the value will be auto-inferred based on height, width')
+
+parser.add_argument('--box', nargs='+', type=int, default=[-1, -1, -1, -1], 
+					help='Specify a constant bounding box for the face. Use only as a last resort if the face is not detected.'
+					'Also, might work only if the face is not moving around much. Syntax: (top, bottom, left, right).')
+
+parser.add_argument('--rotate', default=False, action='store_true',
+					help='Sometimes videos taken from a phone can be flipped 90deg. If true, will flip video right by 90deg.'
+					'Use if you get a flipped result, despite feeding a normal looking video')
+
+parser.add_argument('--nosmooth', default=False, action='store_true',
+					help='Prevent smoothing face detections over a short temporal window')
+
+args = parser.parse_args()
+args.img_size = 96
+
+if os.path.isfile(args.face) and args.face.split('.')[1] in ['jpg', 'png', 'jpeg']:
+	args.static = True
+
+def get_smoothened_boxes(boxes, T):
+	for i in range(len(boxes)):
+		if i + T > len(boxes):
+			window = boxes[len(boxes) - T:]
+		else:
+			window = boxes[i : i + T]
+		boxes[i] = np.mean(window, axis=0)
+	return boxes
+
+def face_detect(images):
+	detector = face_detection.FaceAlignment(face_detection.LandmarksType._2D, 
+											flip_input=False, device=device)
+
+	batch_size = args.face_det_batch_size
+	
+	while 1:
+		predictions = []
+		try:
+			for i in tqdm(range(0, len(images), batch_size)):
+				predictions.extend(detector.get_detections_for_batch(np.array(images[i:i + batch_size])))
+		except RuntimeError:
+			if batch_size == 1: 
+				raise RuntimeError('Image too big to run face detection on GPU. Please use the --resize_factor argument')
+			batch_size //= 2
+			print('Recovering from OOM error; New batch size: {}'.format(batch_size))
+			continue
+		break
+
+	results = []
+	pady1, pady2, padx1, padx2 = args.pads
+	for rect, image in zip(predictions, images):
+		if rect is None:
+			cv2.imwrite('temp/faulty_frame.jpg', image) # check this frame where the face was not detected.
+			raise ValueError('Face not detected! Ensure the video contains a face in all the frames.')
+
+		y1 = max(0, rect[1] - pady1)
+		y2 = min(image.shape[0], rect[3] + pady2)
+		x1 = max(0, rect[0] - padx1)
+		x2 = min(image.shape[1], rect[2] + padx2)
+		
+		results.append([x1, y1, x2, y2])
+
+	boxes = np.array(results)
+	if not args.nosmooth: boxes = get_smoothened_boxes(boxes, T=5)
+	results = [[image[y1: y2, x1:x2], (y1, y2, x1, x2)] for image, (x1, y1, x2, y2) in zip(images, boxes)]
+
+	del detector
+	return results 
+
+def datagen(frames, mels):
+	img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
+
+	if args.box[0] == -1:
+		if not args.static:
+			face_det_results = face_detect(frames) # BGR2RGB for CNN face detection
+		else:
+			face_det_results = face_detect([frames[0]])
+	else:
+		print('Using the specified bounding box instead of face detection...')
+		y1, y2, x1, x2 = args.box
+		face_det_results = [[f[y1: y2, x1:x2], (y1, y2, x1, x2)] for f in frames]
+
+	for i, m in enumerate(mels):
+		idx = 0 if args.static else i%len(frames)
+		frame_to_save = frames[idx].copy()
+		face, coords = face_det_results[idx].copy()
+
+		face = cv2.resize(face, (args.img_size, args.img_size))
+			
+		img_batch.append(face)
+		mel_batch.append(m)
+		frame_batch.append(frame_to_save)
+		coords_batch.append(coords)
+
+		if len(img_batch) >= args.wav2lip_batch_size:
+			img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
+
+			img_masked = img_batch.copy()
+			img_masked[:, args.img_size//2:] = 0
+
+			img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
+			mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
+
+			yield img_batch, mel_batch, frame_batch, coords_batch
+			img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
+
+	if len(img_batch) > 0:
+		img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
+
+		img_masked = img_batch.copy()
+		img_masked[:, args.img_size//2:] = 0
+
+		img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
+		mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
+
+		yield img_batch, mel_batch, frame_batch, coords_batch
+
+mel_step_size = 16
+device = 'cuda' if torch.cuda.is_available() else 'cpu'
+print('Using {} for inference.'.format(device))
+
+# def _load(checkpoint_path):
+# 	if device == 'cuda':
+# 		checkpoint = torch.load(checkpoint_path)
+# 	else:
+# 		checkpoint = torch.load(checkpoint_path,
+# 								map_location=lambda storage, loc: storage)
+# 	return checkpoint
+
+def _load(checkpoint_path):
+    # Use torch.jit.load for TorchScript archives
+    if device == 'cuda':
+        model = torch.jit.load(checkpoint_path)
+    else:
+        # Accepts string or torch.device, not a lambda
+        model = torch.jit.load(checkpoint_path, map_location='cpu')
+    return model
+
+# def load_model(path):
+# 	model = Wav2Lip()
+# 	print("Load checkpoint from: {}".format(path))
+# 	checkpoint = _load(path)
+# 	s = checkpoint["state_dict"]
+# 	new_s = {}
+# 	for k, v in s.items():
+# 		new_s[k.replace('module.', '')] = v
+# 	model.load_state_dict(new_s)
+
+# 	model = model.to(device)
+# 	return model.eval()
+def load_model(path):
+    print("Loading scripted model from:", path)
+    model = _load(path)            # returns the TorchScript Module
+    model = model.to(device)       # move to CPU or GPU
+    return model.eval()            # set to eval() mode
+
+def main():
+	if not os.path.isfile(args.face):
+		raise ValueError('--face argument must be a valid path to video/image file')
+
+	elif args.face.split('.')[1] in ['jpg', 'png', 'jpeg']:
+		full_frames = [cv2.imread(args.face)]
+		fps = args.fps
+
+	else:
+		video_stream = cv2.VideoCapture(args.face)
+		fps = video_stream.get(cv2.CAP_PROP_FPS)
+
+		print('Reading video frames...')
+
+		full_frames = []
+		while 1:
+			still_reading, frame = video_stream.read()
+			if not still_reading:
+				video_stream.release()
+				break
+			if args.resize_factor > 1:
+				frame = cv2.resize(frame, (frame.shape[1]//args.resize_factor, frame.shape[0]//args.resize_factor))
+
+			if args.rotate:
+				frame = cv2.rotate(frame, cv2.cv2.ROTATE_90_CLOCKWISE)
+
+			y1, y2, x1, x2 = args.crop
+			if x2 == -1: x2 = frame.shape[1]
+			if y2 == -1: y2 = frame.shape[0]
+
+			frame = frame[y1:y2, x1:x2]
+
+			full_frames.append(frame)
+
+	print ("Number of frames available for inference: "+str(len(full_frames)))
+
+	if not args.audio.endswith('.wav'):
+		print('Extracting raw audio...')
+		command = 'ffmpeg -y -i {} -strict -2 {}'.format(args.audio, 'temp/temp.wav')
+
+		subprocess.call(command, shell=True)
+		args.audio = 'temp/temp.wav'
+
+	wav = audio.load_wav(args.audio, 16000)
+	mel = audio.melspectrogram(wav)
+	print(mel.shape)
+
+	if np.isnan(mel.reshape(-1)).sum() > 0:
+		raise ValueError('Mel contains nan! Using a TTS voice? Add a small epsilon noise to the wav file and try again')
+
+	mel_chunks = []
+	mel_idx_multiplier = 80./fps 
+	i = 0
+	while 1:
+		start_idx = int(i * mel_idx_multiplier)
+		if start_idx + mel_step_size > len(mel[0]):
+			mel_chunks.append(mel[:, len(mel[0]) - mel_step_size:])
+			break
+		mel_chunks.append(mel[:, start_idx : start_idx + mel_step_size])
+		i += 1
+
+	print("Length of mel chunks: {}".format(len(mel_chunks)))
+
+	full_frames = full_frames[:len(mel_chunks)]
+
+	batch_size = args.wav2lip_batch_size
+	gen = datagen(full_frames.copy(), mel_chunks)
+
+	for i, (img_batch, mel_batch, frames, coords) in enumerate(tqdm(gen, 
+											total=int(np.ceil(float(len(mel_chunks))/batch_size)))):
+		if i == 0:
+			model = load_model(args.checkpoint_path)
+			print ("Model loaded")
+
+			frame_h, frame_w = full_frames[0].shape[:-1]
+			out = cv2.VideoWriter('temp/result.avi', 
+									cv2.VideoWriter_fourcc(*'DIVX'), fps, (frame_w, frame_h))
+
+		img_batch = torch.FloatTensor(np.transpose(img_batch, (0, 3, 1, 2))).to(device)
+		mel_batch = torch.FloatTensor(np.transpose(mel_batch, (0, 3, 1, 2))).to(device)
+
+		with torch.no_grad():
+			pred = model(mel_batch, img_batch)
+
+		pred = pred.cpu().numpy().transpose(0, 2, 3, 1) * 255.
+		
+		for p, f, c in zip(pred, frames, coords):
+			y1, y2, x1, x2 = c
+			p = cv2.resize(p.astype(np.uint8), (x2 - x1, y2 - y1))
+
+			f[y1:y2, x1:x2] = p
+			out.write(f)
+
+	out.release()
+
+	command = 'ffmpeg -y -i {} -i {} -strict -2 -q:v 1 {}'.format(args.audio, 'temp/result.avi', args.outfile)
+	subprocess.call(command, shell=platform.system() != 'Windows')
+
+if __name__ == '__main__':
+	main()

inference2.py

@@ -0,0 +1,346 @@
+# inference.py (Updated)
+
+from os import listdir, path
+import numpy as np
+import scipy, cv2, os, sys, argparse, audio
+import json, subprocess, random, string
+from tqdm import tqdm
+from glob import glob
+import torch # Ensure torch is imported
+try:
+    import face_detection # Assuming this is installed or in a path accessible by your Flask app
+except ImportError:
+    print("face_detection not found. Please ensure it's installed or available in your PYTHONPATH.")
+    # You might want to raise an error or handle this gracefully if face_detection is truly optional.
+
+# Make sure you have a models/Wav2Lip.py or similar structure
+try:
+    from models import Wav2Lip
+except ImportError:
+    print("Wav2Lip model not found. Please ensure models/Wav2Lip.py exists and is correctly configured.")
+    # You might want to raise an error or handle this gracefully.
+
+import platform
+import shutil # For clearing temp directory
+
+
+# These globals are still useful for shared configuration
+mel_step_size = 16
+device = 'cuda' if torch.cuda.is_available() else 'cpu'
+print('Inference script using {} for inference.'.format(device))
+
+
+def get_smoothened_boxes(boxes, T):
+    for i in range(len(boxes)):
+        if i + T > len(boxes):
+            window = boxes[len(boxes) - T:]
+        else:
+            window = boxes[i : i + T]
+        boxes[i] = np.mean(window, axis=0)
+    return boxes
+
+def face_detect(images, pads, face_det_batch_size, nosmooth, img_size):
+    detector = face_detection.FaceAlignment(face_detection.LandmarksType._2D,
+                                            flip_input=False, device=device)
+
+    batch_size = face_det_batch_size
+
+    while 1:
+        predictions = []
+        try:
+            for i in tqdm(range(0, len(images), batch_size), desc="Face Detection"):
+                predictions.extend(detector.get_detections_for_batch(np.array(images[i:i + batch_size])))
+        except RuntimeError as e:
+            if batch_size == 1:
+                raise RuntimeError(f'Image too big to run face detection on GPU. Error: {e}')
+            batch_size //= 2
+            print('Recovering from OOM error; New face detection batch size: {}'.format(batch_size))
+            continue
+        break
+
+    results = []
+    pady1, pady2, padx1, padx2 = pads
+    for rect, image in zip(predictions, images):
+        if rect is None:
+            # Save the faulty frame for debugging
+            output_dir = 'temp' # Ensure this exists or create it
+            os.makedirs(output_dir, exist_ok=True)
+            cv2.imwrite(os.path.join(output_dir, 'faulty_frame.jpg'), image)
+            raise ValueError('Face not detected! Ensure the video/image contains a face in all the frames or try adjusting pads/box.')
+
+        y1 = max(0, rect[1] - pady1)
+        y2 = min(image.shape[0], rect[3] + pady2)
+        x1 = max(0, rect[0] - padx1)
+        x2 = min(image.shape[1], rect[2] + padx2)
+
+        results.append([x1, y1, x2, y2])
+
+    boxes = np.array(results)
+    if not nosmooth: boxes = get_smoothened_boxes(boxes, T=5)
+    results = [[image[y1: y2, x1:x2], (y1, y2, x1, x2)] for image, (x1, y1, x2, y2) in zip(images, boxes)]
+
+    del detector # Clean up detector
+    return results
+
+def datagen(frames, mels, box, static, wav2lip_batch_size, img_size, pads, face_det_batch_size, nosmooth):
+    img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
+
+    if box[0] == -1:
+        if not static:
+            face_det_results = face_detect(frames, pads, face_det_batch_size, nosmooth, img_size) # BGR2RGB for CNN face detection
+        else:
+            face_det_results = face_detect([frames[0]], pads, face_det_batch_size, nosmooth, img_size)
+    else:
+        print('Using the specified bounding box instead of face detection...')
+        y1, y2, x1, x2 = box
+        face_det_results = [[f[y1: y2, x1:x2], (y1, y2, x1, x2)] for f in frames]
+
+    for i, m in enumerate(mels):
+        idx = 0 if static else i % len(frames)
+        frame_to_save = frames[idx].copy()
+        face, coords = face_det_results[idx].copy()
+
+        face = cv2.resize(face, (img_size, img_size))
+
+        img_batch.append(face)
+        mel_batch.append(m)
+        frame_batch.append(frame_to_save)
+        coords_batch.append(coords)
+
+        if len(img_batch) >= wav2lip_batch_size:
+            img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
+
+            img_masked = img_batch.copy()
+            img_masked[:, img_size//2:] = 0
+
+            img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
+            mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
+
+            yield img_batch, mel_batch, frame_batch, coords_batch
+            img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []
+
+    if len(img_batch) > 0:
+        img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)
+
+        img_masked = img_batch.copy()
+        img_masked[:, img_size//2:] = 0
+
+        img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
+        mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
+
+        yield img_batch, mel_batch, frame_batch, coords_batch
+
+def _load(checkpoint_path):
+    # Use torch.jit.load for TorchScript archives
+    if device == 'cuda':
+        model = torch.jit.load(checkpoint_path)
+    else:
+        # Accepts string or torch.device, not a lambda
+        model = torch.jit.load(checkpoint_path, map_location='cpu')
+    return model
+
+def load_model(path):
+    print("Loading scripted model from:", path)
+    model = _load(path) # returns the TorchScript Module
+    model = model.to(device) # move to CPU or GPU
+    return model.eval() # set to eval() mode
+
+
+# New function to be called from Flask app
+def run_inference(
+    checkpoint_path: str,
+    face_path: str,
+    audio_path: str,
+    output_filename: str,
+    static: bool = False,
+    fps: float = 25.,
+    pads: list = [0, 10, 0, 0],
+    face_det_batch_size: int = 16,
+    wav2lip_batch_size: int = 128,
+    resize_factor: int = 1,
+    crop: list = [0, -1, 0, -1],
+    box: list = [-1, -1, -1, -1],
+    rotate: bool = False,
+    nosmooth: bool = False,
+    img_size: int = 96 # Fixed for Wav2Lip
+) -> str:
+    """
+    Runs the Wav2Lip inference process.
+
+    Args:
+        checkpoint_path (str): Path to the Wav2Lip model checkpoint.
+        face_path (str): Path to the input video/image file with a face.
+        audio_path (str): Path to the input audio file.
+        output_filename (str): Name of the output video file (e.g., 'result.mp4').
+        static (bool): If True, use only the first video frame for inference.
+        fps (float): Frames per second for static image input.
+        pads (list): Padding for face detection (top, bottom, left, right).
+        face_det_batch_size (int): Batch size for face detection.
+        wav2lip_batch_size (int): Batch size for Wav2Lip model(s).
+        resize_factor (int): Reduce the resolution by this factor.
+        crop (list): Crop video to a smaller region (top, bottom, left, right).
+        box (list): Constant bounding box for the face.
+        rotate (bool): Rotate video right by 90deg.
+        nosmooth (bool): Prevent smoothing face detections.
+        img_size (int): Image size for the model.
+
+    Returns:
+        str: The path to the generated output video file.
+    """
+    print(f"Starting inference with: face='{face_path}', audio='{audio_path}', checkpoint='{checkpoint_path}', outfile='{output_filename}'")
+
+    # Create necessary directories
+    output_dir = 'results'
+    temp_dir = 'temp'
+    os.makedirs(output_dir, exist_ok=True)
+    os.makedirs(temp_dir, exist_ok=True)
+
+    # Clear temp directory for fresh run
+    for item in os.listdir(temp_dir):
+        item_path = os.path.join(temp_dir, item)
+        if os.path.isfile(item_path):
+            os.remove(item_path)
+        elif os.path.isdir(item_path):
+            shutil.rmtree(item_path)
+
+    # Determine if input is static based on file extension
+    is_static_input = static or (os.path.isfile(face_path) and face_path.split('.')[-1].lower() in ['jpg', 'png', 'jpeg'])
+
+    full_frames = []
+    if is_static_input:
+        full_frames = [cv2.imread(face_path)]
+        if full_frames[0] is None:
+            raise ValueError(f"Could not read face image at: {face_path}")
+    else:
+        video_stream = cv2.VideoCapture(face_path)
+        if not video_stream.isOpened():
+            raise ValueError(f"Could not open video file at: {face_path}")
+        fps = video_stream.get(cv2.CAP_PROP_FPS)
+
+        print('Reading video frames...')
+        while 1:
+            still_reading, frame = video_stream.read()
+            if not still_reading:
+                video_stream.release()
+                break
+            if resize_factor > 1:
+                frame = cv2.resize(frame, (frame.shape[1]//resize_factor, frame.shape[0]//resize_factor))
+
+            if rotate:
+                frame = cv2.rotate(frame, cv2.ROTATE_90_CLOCKWISE)
+
+            y1, y2, x1, x2 = crop
+            if x2 == -1: x2 = frame.shape[1]
+            if y2 == -1: y2 = frame.shape[0]
+
+            frame = frame[y1:y2, x1:x2]
+            full_frames.append(frame)
+
+    print ("Number of frames available for inference: "+str(len(full_frames)))
+    if not full_frames:
+        raise ValueError("No frames could be read from the input face file.")
+
+    temp_audio_path = os.path.join(temp_dir, 'temp_audio.wav')
+    if not audio_path.endswith('.wav'):
+        print('Extracting raw audio...')
+        command = f'ffmpeg -y -i "{audio_path}" -strict -2 "{temp_audio_path}"'
+        try:
+            subprocess.run(command, shell=True, check=True, capture_output=True)
+            audio_path = temp_audio_path
+        except subprocess.CalledProcessError as e:
+            print(f"FFmpeg error: {e.stderr.decode()}")
+            raise RuntimeError(f"Failed to extract audio from {audio_path}. Error: {e.stderr.decode()}")
+    else:
+        # Copy the wav file to temp if it's already wav to maintain consistency in naming
+        shutil.copy(audio_path, temp_audio_path)
+        audio_path = temp_audio_path
+
+
+    wav = audio.load_wav(audio_path, 16000)
+    mel = audio.melspectrogram(wav)
+    print("Mel spectrogram shape:", mel.shape)
+
+    if np.isnan(mel.reshape(-1)).sum() > 0:
+        raise ValueError('Mel contains nan! Using a TTS voice? Add a small epsilon noise to the wav file and try again')
+
+    mel_chunks = []
+    mel_idx_multiplier = 80./fps
+    i = 0
+    while 1:
+        start_idx = int(i * mel_idx_multiplier)
+        if start_idx + mel_step_size > len(mel[0]):
+            mel_chunks.append(mel[:, len(mel[0]) - mel_step_size:])
+            break
+        mel_chunks.append(mel[:, start_idx : start_idx + mel_step_size])
+        i += 1
+
+    print("Length of mel chunks: {}".format(len(mel_chunks)))
+
+    # Ensure full_frames matches mel_chunks length, or loop if static
+    if not is_static_input:
+        full_frames = full_frames[:len(mel_chunks)]
+    else:
+        # If static, replicate the first frame for the duration of the audio
+        full_frames = [full_frames[0]] * len(mel_chunks)
+
+
+    gen = datagen(full_frames.copy(), mel_chunks, box, is_static_input, wav2lip_batch_size, img_size, pads, face_det_batch_size, nosmooth)
+
+    output_avi_path = os.path.join(temp_dir, 'result.avi')
+
+    model_loaded = False
+    model = None
+    frame_h, frame_w = 0, 0
+    out = None
+
+    for i, (img_batch, mel_batch, frames, coords) in enumerate(tqdm(gen, desc="Wav2Lip Inference",
+                                            total=int(np.ceil(float(len(mel_chunks))/wav2lip_batch_size)))):
+        if not model_loaded:
+            model = load_model(checkpoint_path)
+            model_loaded = True
+            print ("Model loaded successfully")
+
+            frame_h, frame_w = full_frames[0].shape[:-1]
+            out = cv2.VideoWriter(output_avi_path,
+                                    cv2.VideoWriter_fourcc(*'DIVX'), fps, (frame_w, frame_h))
+        if out is None: # In case no frames were generated for some reason
+            raise RuntimeError("Video writer could not be initialized.")
+
+
+        img_batch = torch.FloatTensor(np.transpose(img_batch, (0, 3, 1, 2))).to(device)
+        mel_batch = torch.FloatTensor(np.transpose(mel_batch, (0, 3, 1, 2))).to(device)
+
+        with torch.no_grad():
+            pred = model(mel_batch, img_batch)
+
+        pred = pred.cpu().numpy().transpose(0, 2, 3, 1) * 255.
+
+        for p, f, c in zip(pred, frames, coords):
+            y1, y2, x1, x2 = c
+            p = cv2.resize(p.astype(np.uint8), (x2 - x1, y2 - y1))
+
+            f[y1:y2, x1:x2] = p
+            out.write(f)
+
+    if out:
+        out.release()
+    else:
+        print("Warning: Video writer was not initialized or no frames were processed.")
+
+
+    final_output_path = os.path.join(output_dir, output_filename)
+    command = f'ffmpeg -y -i "{audio_path}" -i "{output_avi_path}" -strict -2 -q:v 1 "{final_output_path}"'
+
+    try:
+        subprocess.run(command, shell=True, check=True, capture_output=True)
+        print(f"Output saved to: {final_output_path}")
+    except subprocess.CalledProcessError as e:
+        print(f"FFmpeg final merge error: {e.stderr.decode()}")
+        raise RuntimeError(f"Failed to merge audio and video. Error: {e.stderr.decode()}")
+
+    # Clean up temporary files (optional, but good practice)
+    # shutil.rmtree(temp_dir) # Be careful with this if you want to inspect temp files
+
+    return final_output_path
+
+# No `if __name__ == '__main__':` block here, as it's meant to be imported

info_install.txt

@@ -0,0 +1,54 @@
+install version python 3.7
+conda create the environment
+install all with the code forge 
+conda install -c conda-forge numpy=1.17.1 scipy=1.3.1 numba=0.48 tqdm=4.45.0 -y
+conda install -c conda-forge librosa=0.7.0 -y
+conda install -c conda-forge opencv=4.1.0 -y
+pip install --no-deps opencv-contrib-python==4.1.0.25
+pip install https://mirrors.aliyun.com/pytorch-wheels/cpu/torchvision-0.3.0-cp37-cp37m-win_amd64.whl
+pip install https://mirror.sjtu.edu.cn/pytorch-wheels/cpu/torch-1.1.0-cp37-cp37m-win_amd64.whl    
+
+
+all installs
+(wav2lip_env) D:\DEV PATEL\2025\Wav2Lip-master>pip freeze
+audioread @ file:///D:/bld/audioread_1660497578082/work
+certifi @ file:///home/conda/feedstock_root/build_artifacts/certifi_1725278078093/work/certifi
+cffi @ file:///D:/bld/cffi_1666183927951/work
+cycler @ file:///home/conda/feedstock_root/build_artifacts/cycler_1635519461629/work
+decorator @ file:///home/conda/feedstock_root/build_artifacts/decorator_1641555617451/work
+fonttools @ file:///D:/bld/fonttools_1666390069478/work
+joblib @ file:///home/conda/feedstock_root/build_artifacts/joblib_1691577114857/work
+kiwisolver @ file:///D:/bld/kiwisolver_1657953189205/work
+librosa==0.7.0
+llvmlite==0.31.0
+matplotlib @ file:///C:/ci/matplotlib-suite_1634667159685/work
+mkl-service==2.3.0
+munkres==1.1.4
+numba==0.48.0
+numpy==1.17.1
+opencv-contrib-python==4.1.0.25
+packaging @ file:///home/conda/feedstock_root/build_artifacts/packaging_1696202382185/work
+Pillow==9.3.0
+ply @ file:///home/conda/feedstock_root/build_artifacts/ply_1712242996588/work
+pycparser @ file:///home/conda/feedstock_root/build_artifacts/pycparser_1636257122734/work
+pyparsing @ file:///home/conda/feedstock_root/build_artifacts/pyparsing_1724616129934/work
+PyQt5-sip @ file:///D:/bld/pyqt-split_1665676787902/work/pyqt_sip
+python-dateutil @ file:///home/conda/feedstock_root/build_artifacts/python-dateutil_1709299778482/work
+resampy @ file:///home/conda/feedstock_root/build_artifacts/resampy_1657206395424/work
+scikit-learn @ file:///D:/bld/scikit-learn_1611079929791/work
+scipy==1.3.1
+six @ file:///home/conda/feedstock_root/build_artifacts/six_1620240208055/work
+soundfile @ file:///home/conda/feedstock_root/build_artifacts/pysoundfile_1676571469739/work
+threadpoolctl @ file:///home/conda/feedstock_root/build_artifacts/threadpoolctl_1643647933166/work
+toml @ file:///home/conda/feedstock_root/build_artifacts/toml_1604308577558/work
+torch @ https://mirror.sjtu.edu.cn/pytorch-wheels/cpu/torch-1.1.0-cp37-cp37m-win_amd64.whl
+torchvision @ https://mirrors.aliyun.com/pytorch-wheels/cpu/torchvision-0.3.0-cp37-cp37m-win_amd64.whl
+tornado @ file:///D:/bld/tornado_1656937938087/work
+tqdm==4.45.0
+typing_extensions @ file:///home/conda/feedstock_root/build_artifacts/typing_extensions_1688315532570/work
+unicodedata2 @ file:///D:/bld/unicodedata2_1649112131705/work
+wincertstore==0.2
+
+drive 
+model download for chrckpoint - https://drive.google.com/drive/folders/153HLrqlBNxzZcHi17PEvP09kkAfzRshM
+face recognition model - https://www.adrianbulat.com/downloads/python-fan/s3fd-619a316812.pth  (alternate link -  https://iiitaphyd-my.sharepoint.com/:u:/g/personal/prajwal_k_research_iiit_ac_in/EZsy6qWuivtDnANIG73iHjIBjMSoojcIV0NULXV-yiuiIg?e=qTasa8)

input/audio/audio_hindi_tony_stark.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be12df3fb3d817bae39f6b02e532e67f33af7d04cafdf4ac59d971d9aa239b21
+size 1260284

input/audio/harvard.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:971b4163670445c415c6b0fb6813c38093409ecac2f6b4d429ae3574d24ad470
+size 3249924

input/audio/processed_tony_stark.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:16b1d2a79ff219503cb8f8df1e4d0bc6ea4aac791528b112372d71a1a2766ca9
+size 580318