wav2lip model added

Wav2Lip/audio.py

@@ -0,0 +1,219 @@
+import librosa
+import librosa.filters
+import numpy as np
+# import tensorflow as tf
+from scipy import signal
+from scipy.io import wavfile
+# from hparams import hparams as hp
+
+class HParams:
+	def __init__(self, **kwargs):
+		self.data = {}
+
+		for key, value in kwargs.items():
+			self.data[key] = value
+
+	def __getattr__(self, key):
+		if key not in self.data:
+			raise AttributeError("'HParams' object has no attribute %s" % key)
+		return self.data[key]
+
+	def set_hparam(self, key, value):
+		self.data[key] = value
+
+
+# Default hyperparameters
+hp = HParams(
+	num_mels=80,  # Number of mel-spectrogram channels and local conditioning dimensionality
+	#  network
+	rescale=True,  # Whether to rescale audio prior to preprocessing
+	rescaling_max=0.9,  # Rescaling value
+	
+	# Use LWS (https://github.com/Jonathan-LeRoux/lws) for STFT and phase reconstruction
+	# It"s preferred to set True to use with https://github.com/r9y9/wavenet_vocoder
+	# Does not work if n_ffit is not multiple of hop_size!!
+	use_lws=False,
+	
+	n_fft=800,  # Extra window size is filled with 0 paddings to match this parameter
+	hop_size=200,  # For 16000Hz, 200 = 12.5 ms (0.0125 * sample_rate)
+	win_size=800,  # For 16000Hz, 800 = 50 ms (If None, win_size = n_fft) (0.05 * sample_rate)
+	sample_rate=16000,  # 16000Hz (corresponding to librispeech) (sox --i <filename>)
+	
+	frame_shift_ms=None,  # Can replace hop_size parameter. (Recommended: 12.5)
+	
+	# Mel and Linear spectrograms normalization/scaling and clipping
+	signal_normalization=True,
+	# Whether to normalize mel spectrograms to some predefined range (following below parameters)
+	allow_clipping_in_normalization=True,  # Only relevant if mel_normalization = True
+	symmetric_mels=True,
+	# Whether to scale the data to be symmetric around 0. (Also multiplies the output range by 2, 
+	# faster and cleaner convergence)
+	max_abs_value=4.,
+	# max absolute value of data. If symmetric, data will be [-max, max] else [0, max] (Must not 
+	# be too big to avoid gradient explosion, 
+	# not too small for fast convergence)
+	# Contribution by @begeekmyfriend
+	# Spectrogram Pre-Emphasis (Lfilter: Reduce spectrogram noise and helps model certitude 
+	# levels. Also allows for better G&L phase reconstruction)
+	preemphasize=True,  # whether to apply filter
+	preemphasis=0.97,  # filter coefficient.
+	
+	# Limits
+	min_level_db=-100,
+	ref_level_db=20,
+	fmin=55,
+	# Set this to 55 if your speaker is male! if female, 95 should help taking off noise. (To 
+	# test depending on dataset. Pitch info: male~[65, 260], female~[100, 525])
+	fmax=7600,  # To be increased/reduced depending on data.
+
+	###################### Our training parameters #################################
+	img_size=96,
+	fps=25,
+	
+	batch_size=16,
+	initial_learning_rate=1e-4,
+	nepochs=200000000000000000,  ### ctrl + c, stop whenever eval loss is consistently greater than train loss for ~10 epochs
+	num_workers=16,
+	checkpoint_interval=3000,
+	eval_interval=3000,
+    save_optimizer_state=True,
+
+    syncnet_wt=0.0, # is initially zero, will be set automatically to 0.03 later. Leads to faster convergence. 
+	syncnet_batch_size=64,
+	syncnet_lr=1e-4,
+	syncnet_eval_interval=10000,
+	syncnet_checkpoint_interval=10000,
+
+	disc_wt=0.07,
+	disc_initial_learning_rate=1e-4,
+)
+
+def load_wav(path, sr):
+    return librosa.core.load(path, sr=sr)[0]
+
+def save_wav(wav, path, sr):
+    wav *= 32767 / max(0.01, np.max(np.abs(wav)))
+    #proposed by @dsmiller
+    wavfile.write(path, sr, wav.astype(np.int16))
+
+def save_wavenet_wav(wav, path, sr):
+    librosa.output.write_wav(path, wav, sr=sr)
+
+def preemphasis(wav, k, preemphasize=True):
+    if preemphasize:
+        return signal.lfilter([1, -k], [1], wav)
+    return wav
+
+def inv_preemphasis(wav, k, inv_preemphasize=True):
+    if inv_preemphasize:
+        return signal.lfilter([1], [1, -k], wav)
+    return wav
+
+def get_hop_size():
+    hop_size = hp.hop_size
+    if hop_size is None:
+        assert hp.frame_shift_ms is not None
+        hop_size = int(hp.frame_shift_ms / 1000 * hp.sample_rate)
+    return hop_size
+
+def linearspectrogram(wav):
+    D = _stft(preemphasis(wav, hp.preemphasis, hp.preemphasize))
+    S = _amp_to_db(np.abs(D)) - hp.ref_level_db
+    
+    if hp.signal_normalization:
+        return _normalize(S)
+    return S
+
+def melspectrogram(wav):
+    D = _stft(preemphasis(wav, hp.preemphasis, hp.preemphasize))
+    S = _amp_to_db(_linear_to_mel(np.abs(D))) - hp.ref_level_db
+    
+    if hp.signal_normalization:
+        return _normalize(S)
+    return S
+
+def _lws_processor():
+    import lws
+    return lws.lws(hp.n_fft, get_hop_size(), fftsize=hp.win_size, mode="speech")
+
+def _stft(y):
+    if hp.use_lws:
+        return _lws_processor(hp).stft(y).T
+    else:
+        return librosa.stft(y=y, n_fft=hp.n_fft, hop_length=get_hop_size(), win_length=hp.win_size)
+
+##########################################################
+#Those are only correct when using lws!!! (This was messing with Wavenet quality for a long time!)
+def num_frames(length, fsize, fshift):
+    """Compute number of time frames of spectrogram
+    """
+    pad = (fsize - fshift)
+    if length % fshift == 0:
+        M = (length + pad * 2 - fsize) // fshift + 1
+    else:
+        M = (length + pad * 2 - fsize) // fshift + 2
+    return M
+
+
+def pad_lr(x, fsize, fshift):
+    """Compute left and right padding
+    """
+    M = num_frames(len(x), fsize, fshift)
+    pad = (fsize - fshift)
+    T = len(x) + 2 * pad
+    r = (M - 1) * fshift + fsize - T
+    return pad, pad + r
+##########################################################
+#Librosa correct padding
+def librosa_pad_lr(x, fsize, fshift):
+    return 0, (x.shape[0] // fshift + 1) * fshift - x.shape[0]
+
+# Conversions
+_mel_basis = None
+
+def _linear_to_mel(spectogram):
+    global _mel_basis
+    if _mel_basis is None:
+        _mel_basis = _build_mel_basis()
+    return np.dot(_mel_basis, spectogram)
+
+def _build_mel_basis():
+    assert hp.fmax <= hp.sample_rate // 2
+    # return librosa.filters.mel(hp.sample_rate, hp.n_fft, n_mels=hp.num_mels,
+    #                            fmin=hp.fmin, fmax=hp.fmax)
+    return librosa.filters.mel(sr=hp.sample_rate, n_fft=hp.n_fft)
+
+def _amp_to_db(x):
+    min_level = np.exp(hp.min_level_db / 20 * np.log(10))
+    return 20 * np.log10(np.maximum(min_level, x))
+
+def _db_to_amp(x):
+    return np.power(10.0, (x) * 0.05)
+
+def _normalize(S):
+    if hp.allow_clipping_in_normalization:
+        if hp.symmetric_mels:
+            return np.clip((2 * hp.max_abs_value) * ((S - hp.min_level_db) / (-hp.min_level_db)) - hp.max_abs_value,
+                           -hp.max_abs_value, hp.max_abs_value)
+        else:
+            return np.clip(hp.max_abs_value * ((S - hp.min_level_db) / (-hp.min_level_db)), 0, hp.max_abs_value)
+    
+    assert S.max() <= 0 and S.min() - hp.min_level_db >= 0
+    if hp.symmetric_mels:
+        return (2 * hp.max_abs_value) * ((S - hp.min_level_db) / (-hp.min_level_db)) - hp.max_abs_value
+    else:
+        return hp.max_abs_value * ((S - hp.min_level_db) / (-hp.min_level_db))
+
+def _denormalize(D):
+    if hp.allow_clipping_in_normalization:
+        if hp.symmetric_mels:
+            return (((np.clip(D, -hp.max_abs_value,
+                              hp.max_abs_value) + hp.max_abs_value) * -hp.min_level_db / (2 * hp.max_abs_value))
+                    + hp.min_level_db)
+        else:
+            return ((np.clip(D, 0, hp.max_abs_value) * -hp.min_level_db / hp.max_abs_value) + hp.min_level_db)
+    
+    if hp.symmetric_mels:
+        return (((D + hp.max_abs_value) * -hp.min_level_db / (2 * hp.max_abs_value)) + hp.min_level_db)
+    else:
+        return ((D * -hp.min_level_db / hp.max_abs_value) + hp.min_level_db)

Wav2Lip/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. 

Wav2Lip/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

Wav2Lip/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__('Wav2Lip.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

Wav2Lip/face_detection/detection/__init__.py

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

Wav2Lip/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

Wav2Lip/face_detection/detection/sfd/__init__.py

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

Wav2Lip/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

Wav2Lip/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])

Wav2Lip/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]

Wav2Lip/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

Wav2Lip/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

Wav2Lip/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

Wav2Lip/hparams.py

@@ -0,0 +1,101 @@
+from glob import glob
+import os
+
+def get_image_list(data_root, split):
+	filelist = []
+
+	with open('filelists/{}.txt'.format(split)) as f:
+		for line in f:
+			line = line.strip()
+			if ' ' in line: line = line.split()[0]
+			filelist.append(os.path.join(data_root, line))
+
+	return filelist
+
+class HParams:
+	def __init__(self, **kwargs):
+		self.data = {}
+
+		for key, value in kwargs.items():
+			self.data[key] = value
+
+	def __getattr__(self, key):
+		if key not in self.data:
+			raise AttributeError("'HParams' object has no attribute %s" % key)
+		return self.data[key]
+
+	def set_hparam(self, key, value):
+		self.data[key] = value
+
+
+# Default hyperparameters
+hparams = HParams(
+	num_mels=80,  # Number of mel-spectrogram channels and local conditioning dimensionality
+	#  network
+	rescale=True,  # Whether to rescale audio prior to preprocessing
+	rescaling_max=0.9,  # Rescaling value
+	
+	# Use LWS (https://github.com/Jonathan-LeRoux/lws) for STFT and phase reconstruction
+	# It"s preferred to set True to use with https://github.com/r9y9/wavenet_vocoder
+	# Does not work if n_ffit is not multiple of hop_size!!
+	use_lws=False,
+	
+	n_fft=800,  # Extra window size is filled with 0 paddings to match this parameter
+	hop_size=200,  # For 16000Hz, 200 = 12.5 ms (0.0125 * sample_rate)
+	win_size=800,  # For 16000Hz, 800 = 50 ms (If None, win_size = n_fft) (0.05 * sample_rate)
+	sample_rate=16000,  # 16000Hz (corresponding to librispeech) (sox --i <filename>)
+	
+	frame_shift_ms=None,  # Can replace hop_size parameter. (Recommended: 12.5)
+	
+	# Mel and Linear spectrograms normalization/scaling and clipping
+	signal_normalization=True,
+	# Whether to normalize mel spectrograms to some predefined range (following below parameters)
+	allow_clipping_in_normalization=True,  # Only relevant if mel_normalization = True
+	symmetric_mels=True,
+	# Whether to scale the data to be symmetric around 0. (Also multiplies the output range by 2, 
+	# faster and cleaner convergence)
+	max_abs_value=4.,
+	# max absolute value of data. If symmetric, data will be [-max, max] else [0, max] (Must not 
+	# be too big to avoid gradient explosion, 
+	# not too small for fast convergence)
+	# Contribution by @begeekmyfriend
+	# Spectrogram Pre-Emphasis (Lfilter: Reduce spectrogram noise and helps model certitude 
+	# levels. Also allows for better G&L phase reconstruction)
+	preemphasize=True,  # whether to apply filter
+	preemphasis=0.97,  # filter coefficient.
+	
+	# Limits
+	min_level_db=-100,
+	ref_level_db=20,
+	fmin=55,
+	# Set this to 55 if your speaker is male! if female, 95 should help taking off noise. (To 
+	# test depending on dataset. Pitch info: male~[65, 260], female~[100, 525])
+	fmax=7600,  # To be increased/reduced depending on data.
+
+	###################### Our training parameters #################################
+	img_size=96,
+	fps=25,
+	
+	batch_size=16,
+	initial_learning_rate=1e-4,
+	nepochs=200000000000000000,  ### ctrl + c, stop whenever eval loss is consistently greater than train loss for ~10 epochs
+	num_workers=16,
+	checkpoint_interval=3000,
+	eval_interval=3000,
+    save_optimizer_state=True,
+
+    syncnet_wt=0.0, # is initially zero, will be set automatically to 0.03 later. Leads to faster convergence. 
+	syncnet_batch_size=64,
+	syncnet_lr=1e-4,
+	syncnet_eval_interval=10000,
+	syncnet_checkpoint_interval=10000,
+
+	disc_wt=0.07,
+	disc_initial_learning_rate=1e-4,
+)
+
+
+def hparams_debug_string():
+	values = hparams.values()
+	hp = ["  %s: %s" % (name, values[name]) for name in sorted(values) if name != "sentences"]
+	return "Hyperparameters:\n" + "\n".join(hp)

Wav2Lip/models/__init__.py

@@ -0,0 +1,2 @@
+from .wav2lip import Wav2Lip, Wav2Lip_disc_qual
+from .syncnet import SyncNet_color

Wav2Lip/models/conv.py

@@ -0,0 +1,44 @@
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+class Conv2d(nn.Module):
+    def __init__(self, cin, cout, kernel_size, stride, padding, residual=False, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.conv_block = nn.Sequential(
+                            nn.Conv2d(cin, cout, kernel_size, stride, padding),
+                            nn.BatchNorm2d(cout)
+                            )
+        self.act = nn.ReLU()
+        self.residual = residual
+
+    def forward(self, x):
+        out = self.conv_block(x)
+        if self.residual:
+            out += x
+        return self.act(out)
+
+class nonorm_Conv2d(nn.Module):
+    def __init__(self, cin, cout, kernel_size, stride, padding, residual=False, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.conv_block = nn.Sequential(
+                            nn.Conv2d(cin, cout, kernel_size, stride, padding),
+                            )
+        self.act = nn.LeakyReLU(0.01, inplace=True)
+
+    def forward(self, x):
+        out = self.conv_block(x)
+        return self.act(out)
+
+class Conv2dTranspose(nn.Module):
+    def __init__(self, cin, cout, kernel_size, stride, padding, output_padding=0, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.conv_block = nn.Sequential(
+                            nn.ConvTranspose2d(cin, cout, kernel_size, stride, padding, output_padding),
+                            nn.BatchNorm2d(cout)
+                            )
+        self.act = nn.ReLU()
+
+    def forward(self, x):
+        out = self.conv_block(x)
+        return self.act(out)

Wav2Lip/models/syncnet.py

@@ -0,0 +1,66 @@
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from .conv import Conv2d
+
+class SyncNet_color(nn.Module):
+    def __init__(self):
+        super(SyncNet_color, self).__init__()
+
+        self.face_encoder = nn.Sequential(
+            Conv2d(15, 32, kernel_size=(7, 7), stride=1, padding=3),
+
+            Conv2d(32, 64, kernel_size=5, stride=(1, 2), padding=1),
+            Conv2d(64, 64, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(64, 64, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(64, 128, kernel_size=3, stride=2, padding=1),
+            Conv2d(128, 128, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(128, 128, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(128, 128, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(128, 256, kernel_size=3, stride=2, padding=1),
+            Conv2d(256, 256, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(256, 256, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(256, 512, kernel_size=3, stride=2, padding=1),
+            Conv2d(512, 512, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(512, 512, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(512, 512, kernel_size=3, stride=2, padding=1),
+            Conv2d(512, 512, kernel_size=3, stride=1, padding=0),
+            Conv2d(512, 512, kernel_size=1, stride=1, padding=0),)
+
+        self.audio_encoder = nn.Sequential(
+            Conv2d(1, 32, kernel_size=3, stride=1, padding=1),
+            Conv2d(32, 32, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(32, 32, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(32, 64, kernel_size=3, stride=(3, 1), padding=1),
+            Conv2d(64, 64, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(64, 64, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(64, 128, kernel_size=3, stride=3, padding=1),
+            Conv2d(128, 128, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(128, 128, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(128, 256, kernel_size=3, stride=(3, 2), padding=1),
+            Conv2d(256, 256, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(256, 256, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(256, 512, kernel_size=3, stride=1, padding=0),
+            Conv2d(512, 512, kernel_size=1, stride=1, padding=0),)
+
+    def forward(self, audio_sequences, face_sequences): # audio_sequences := (B, dim, T)
+        face_embedding = self.face_encoder(face_sequences)
+        audio_embedding = self.audio_encoder(audio_sequences)
+
+        audio_embedding = audio_embedding.view(audio_embedding.size(0), -1)
+        face_embedding = face_embedding.view(face_embedding.size(0), -1)
+
+        audio_embedding = F.normalize(audio_embedding, p=2, dim=1)
+        face_embedding = F.normalize(face_embedding, p=2, dim=1)
+
+
+        return audio_embedding, face_embedding

Wav2Lip/models/wav2lip.py

@@ -0,0 +1,192 @@
+import torch
+from torch import nn
+from torch.nn import functional as F
+import math
+import streamlit as st
+
+from .conv import Conv2dTranspose, Conv2d, nonorm_Conv2d
+
+class Wav2Lip(nn.Module):
+    def __init__(self):
+        super(Wav2Lip, self).__init__()
+
+        self.face_encoder_blocks = nn.ModuleList([
+            nn.Sequential(Conv2d(6, 16, kernel_size=7, stride=1, padding=3)), # 96,96
+
+            nn.Sequential(Conv2d(16, 32, kernel_size=3, stride=2, padding=1), # 48,48
+            Conv2d(32, 32, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(32, 32, kernel_size=3, stride=1, padding=1, residual=True)),
+
+            nn.Sequential(Conv2d(32, 64, kernel_size=3, stride=2, padding=1),    # 24,24
+            Conv2d(64, 64, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(64, 64, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(64, 64, kernel_size=3, stride=1, padding=1, residual=True)),
+
+            nn.Sequential(Conv2d(64, 128, kernel_size=3, stride=2, padding=1),   # 12,12
+            Conv2d(128, 128, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(128, 128, kernel_size=3, stride=1, padding=1, residual=True)),
+
+            nn.Sequential(Conv2d(128, 256, kernel_size=3, stride=2, padding=1),       # 6,6
+            Conv2d(256, 256, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(256, 256, kernel_size=3, stride=1, padding=1, residual=True)),
+
+            nn.Sequential(Conv2d(256, 512, kernel_size=3, stride=2, padding=1),     # 3,3
+            Conv2d(512, 512, kernel_size=3, stride=1, padding=1, residual=True),),
+            
+            nn.Sequential(Conv2d(512, 512, kernel_size=3, stride=1, padding=0),     # 1, 1
+            Conv2d(512, 512, kernel_size=1, stride=1, padding=0)),])
+
+        self.audio_encoder = nn.Sequential(
+            Conv2d(1, 32, kernel_size=3, stride=1, padding=1),
+            Conv2d(32, 32, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(32, 32, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(32, 64, kernel_size=3, stride=(3, 1), padding=1),
+            Conv2d(64, 64, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(64, 64, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(64, 128, kernel_size=3, stride=3, padding=1),
+            Conv2d(128, 128, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(128, 128, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(128, 256, kernel_size=3, stride=(3, 2), padding=1),
+            Conv2d(256, 256, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(256, 512, kernel_size=3, stride=1, padding=0),
+            Conv2d(512, 512, kernel_size=1, stride=1, padding=0)
+        )
+
+        self.face_decoder_blocks = nn.ModuleList([
+            nn.Sequential(Conv2d(512, 512, kernel_size=1, stride=1, padding=0),),
+
+            nn.Sequential(Conv2dTranspose(1024, 512, kernel_size=3, stride=1, padding=0), # 3,3
+            Conv2d(512, 512, kernel_size=3, stride=1, padding=1, residual=True),),
+
+            nn.Sequential(Conv2dTranspose(1024, 512, kernel_size=3, stride=2, padding=1, output_padding=1),
+            Conv2d(512, 512, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(512, 512, kernel_size=3, stride=1, padding=1, residual=True),), # 6, 6
+
+            nn.Sequential(Conv2dTranspose(768, 384, kernel_size=3, stride=2, padding=1, output_padding=1),
+            Conv2d(384, 384, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(384, 384, kernel_size=3, stride=1, padding=1, residual=True),), # 12, 12
+
+            nn.Sequential(Conv2dTranspose(512, 256, kernel_size=3, stride=2, padding=1, output_padding=1),
+            Conv2d(256, 256, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(256, 256, kernel_size=3, stride=1, padding=1, residual=True),), # 24, 24
+
+            nn.Sequential(Conv2dTranspose(320, 128, kernel_size=3, stride=2, padding=1, output_padding=1), 
+            Conv2d(128, 128, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(128, 128, kernel_size=3, stride=1, padding=1, residual=True),), # 48, 48
+
+            nn.Sequential(Conv2dTranspose(160, 64, kernel_size=3, stride=2, padding=1, output_padding=1),
+            Conv2d(64, 64, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(64, 64, kernel_size=3, stride=1, padding=1, residual=True),),]) # 96,96
+
+        self.output_block = nn.Sequential(Conv2d(80, 32, kernel_size=3, stride=1, padding=1),
+            nn.Conv2d(32, 3, kernel_size=1, stride=1, padding=0),
+            nn.Sigmoid()) 
+
+    def forward(self, audio_sequences, face_sequences):
+        # audio_sequences = (B, T, 1, 80, 16)
+        B = audio_sequences.size(0)
+
+        input_dim_size = len(face_sequences.size())
+        if input_dim_size > 4:
+            audio_sequences = torch.cat([audio_sequences[:, i] for i in range(audio_sequences.size(1))], dim=0)
+            face_sequences = torch.cat([face_sequences[:, :, i] for i in range(face_sequences.size(2))], dim=0)
+        # st.write(f"audio_sequences: {audio_sequences.shape}")
+        audio_embedding = self.audio_encoder(audio_sequences) # B, 512, 1, 1
+        audio_embedding = audio_embedding[:, :, 1, :].unsqueeze(2)
+        # st.write(f"audio_embedding: {audio_embedding.shape}")
+        # st.write(audio_embedding[0, 0, :, :])
+
+        feats = []
+        x = face_sequences
+        for f in self.face_encoder_blocks:
+            x = f(x)
+            feats.append(x)
+
+        x = audio_embedding
+        for f in self.face_decoder_blocks:
+            # st.write(x.shape)
+            # st.write(feats[-1].shape)
+            x = f(x)
+            # st.write(x.shape)
+            try:
+                x = torch.cat((x, feats[-1]), dim=1)
+            except Exception as e:
+                print(x.size())
+                print(feats[-1].size())
+                raise e
+            
+            feats.pop()
+
+        x = self.output_block(x)
+
+        if input_dim_size > 4:
+            x = torch.split(x, B, dim=0) # [(B, C, H, W)]
+            outputs = torch.stack(x, dim=2) # (B, C, T, H, W)
+
+        else:
+            outputs = x
+            
+        return outputs
+
+class Wav2Lip_disc_qual(nn.Module):
+    def __init__(self):
+        super(Wav2Lip_disc_qual, self).__init__()
+
+        self.face_encoder_blocks = nn.ModuleList([
+            nn.Sequential(nonorm_Conv2d(3, 32, kernel_size=7, stride=1, padding=3)), # 48,96
+
+            nn.Sequential(nonorm_Conv2d(32, 64, kernel_size=5, stride=(1, 2), padding=2), # 48,48
+            nonorm_Conv2d(64, 64, kernel_size=5, stride=1, padding=2)),
+
+            nn.Sequential(nonorm_Conv2d(64, 128, kernel_size=5, stride=2, padding=2),    # 24,24
+            nonorm_Conv2d(128, 128, kernel_size=5, stride=1, padding=2)),
+
+            nn.Sequential(nonorm_Conv2d(128, 256, kernel_size=5, stride=2, padding=2),   # 12,12
+            nonorm_Conv2d(256, 256, kernel_size=5, stride=1, padding=2)),
+
+            nn.Sequential(nonorm_Conv2d(256, 512, kernel_size=3, stride=2, padding=1),       # 6,6
+            nonorm_Conv2d(512, 512, kernel_size=3, stride=1, padding=1)),
+
+            nn.Sequential(nonorm_Conv2d(512, 512, kernel_size=3, stride=2, padding=1),     # 3,3
+            nonorm_Conv2d(512, 512, kernel_size=3, stride=1, padding=1),),
+            
+            nn.Sequential(nonorm_Conv2d(512, 512, kernel_size=3, stride=1, padding=0),     # 1, 1
+            nonorm_Conv2d(512, 512, kernel_size=1, stride=1, padding=0)),])
+
+        self.binary_pred = nn.Sequential(nn.Conv2d(512, 1, kernel_size=1, stride=1, padding=0), nn.Sigmoid())
+        self.label_noise = .0
+
+    def get_lower_half(self, face_sequences):
+        return face_sequences[:, :, face_sequences.size(2)//2:]
+
+    def to_2d(self, face_sequences):
+        B = face_sequences.size(0)
+        face_sequences = torch.cat([face_sequences[:, :, i] for i in range(face_sequences.size(2))], dim=0)
+        return face_sequences
+
+    def perceptual_forward(self, false_face_sequences):
+        false_face_sequences = self.to_2d(false_face_sequences)
+        false_face_sequences = self.get_lower_half(false_face_sequences)
+
+        false_feats = false_face_sequences
+        for f in self.face_encoder_blocks:
+            false_feats = f(false_feats)
+
+        false_pred_loss = F.binary_cross_entropy(self.binary_pred(false_feats).view(len(false_feats), -1), 
+                                        torch.ones((len(false_feats), 1)).cuda())
+
+        return false_pred_loss
+
+    def forward(self, face_sequences):
+        face_sequences = self.to_2d(face_sequences)
+        face_sequences = self.get_lower_half(face_sequences)
+
+        x = face_sequences
+        for f in self.face_encoder_blocks:
+            x = f(x)
+
+        return self.binary_pred(x).view(len(x), -1)

Wav2Lip/video_generator.py

@@ -0,0 +1,273 @@
+from pathlib import Path
+import os
+import numpy as np
+import cv2
+from Wav2Lip import audio
+import subprocess
+from tqdm import tqdm
+from Wav2Lip import face_detection
+from Wav2Lip.models import Wav2Lip
+import platform
+# import tensorflow as tf
+import torch
+# import face_alignment
+# import streamlit as st
+
+# checkpoint_path = 'wav2lip_model/wav2lip_gan.tflite'
+checkpoint_path = 'wav2lip_model/wav2lip_gan.pth'
+outfile = 'generated_video.mp4'
+static = False
+fps = 25.
+pads = [0, 10, 0, 0]
+face_det_batch_size = 16
+wav2lip_batch_size = 128
+resize_factor = 1
+crop = [0, -1, 0, -1]
+box = [-1, -1, -1, -1]
+rotate = False
+nosmooth = False
+img_size = 96
+device = 'cpu'
+mel_step_size = 16
+
+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()
+	
+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)
+
+    # detector = face_detection.build_detector("DSFDDetector", confidence_threshold=.5, nms_iou_threshold=.3)
+    # detector = face_alignment.FaceAlignment(face_alignment.LandmarksType.TWO_D, 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)):
+                predictions.extend(detector.get_detections_for_batch(np.array(images[i:i + batch_size])))
+                # predictions.extend(detector.batched_detect(np.array(images[i:i + batch_size])))
+                # predictions.extend(detector.get_landmarks_from_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 = 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 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 box[0] == -1:
+		if not 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 = 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 get_full_frames(face, fps=fps):
+		
+    if not os.path.isfile(face):
+        
+        raise ValueError('face argument must be a valid path to video/image file')
+
+    elif face.split('.')[1] in ['jpg', 'png', 'jpeg']:
+        
+        full_frames = [cv2.imread(face)]
+        fps = fps
+
+    else:
+        
+        video_stream = cv2.VideoCapture(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 resize_factor > 1:
+                frame = cv2.resize(frame, (frame.shape[1]//resize_factor, frame.shape[0]//resize_factor))
+
+            if rotate:
+                frame = cv2.rotate(frame, cv2.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)))
+
+    return full_frames
+
+def get_mel_chunks(voice_audio):
+	
+    if not voice_audio.endswith('.wav'):
+        print('Extracting raw audio...')
+        # st.write(voice_audio)
+        command = 'ffmpeg -y -i {} -strict -2 {}'.format(voice_audio, 'temp/temp.wav')
+        subprocess.call(command, shell=True)
+        voice_audio = 'temp/temp.wav'
+        
+    wav = audio.load_wav(voice_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)))
+
+    return mel_chunks
+	
+def create_video(voice_audio, face):
+	
+    global static
+
+    mel_chunks = get_mel_chunks(voice_audio)
+    full_frames = get_full_frames(face)
+    full_frames = full_frames[:len(mel_chunks)]
+
+    batch_size = wav2lip_batch_size
+
+    if face and face.split('.')[1] in ['jpg', 'png', 'jpeg']:
+        static = True
+
+    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(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(voice_audio, 'temp/result.avi', outfile)
+    subprocess.call(command, shell=platform.system() != 'Windows')