diff --git a/animation/SadTalker/src/utils/__pycache__/audio.cpython-38.pyc b/animation/SadTalker/src/utils/__pycache__/audio.cpython-38.pyc new file mode 100644 index 0000000..fb84654 Binary files /dev/null and b/animation/SadTalker/src/utils/__pycache__/audio.cpython-38.pyc differ diff --git a/animation/SadTalker/src/utils/__pycache__/croper.cpython-38.pyc b/animation/SadTalker/src/utils/__pycache__/croper.cpython-38.pyc new file mode 100644 index 0000000..71efb08 Binary files /dev/null and b/animation/SadTalker/src/utils/__pycache__/croper.cpython-38.pyc differ diff --git a/animation/SadTalker/src/utils/__pycache__/face_enhancer.cpython-38.pyc b/animation/SadTalker/src/utils/__pycache__/face_enhancer.cpython-38.pyc new file mode 100644 index 0000000..067fa24 Binary files /dev/null and b/animation/SadTalker/src/utils/__pycache__/face_enhancer.cpython-38.pyc differ diff --git a/animation/SadTalker/src/utils/__pycache__/hparams.cpython-38.pyc b/animation/SadTalker/src/utils/__pycache__/hparams.cpython-38.pyc new file mode 100644 index 0000000..7f056d2 Binary files /dev/null and b/animation/SadTalker/src/utils/__pycache__/hparams.cpython-38.pyc differ diff --git a/animation/SadTalker/src/utils/__pycache__/init_path.cpython-38.pyc b/animation/SadTalker/src/utils/__pycache__/init_path.cpython-38.pyc new file mode 100644 index 0000000..36640ce Binary files /dev/null and b/animation/SadTalker/src/utils/__pycache__/init_path.cpython-38.pyc differ diff --git a/animation/SadTalker/src/utils/__pycache__/paste_pic.cpython-38.pyc b/animation/SadTalker/src/utils/__pycache__/paste_pic.cpython-38.pyc new file mode 100644 index 0000000..9cb21a3 Binary files /dev/null and b/animation/SadTalker/src/utils/__pycache__/paste_pic.cpython-38.pyc differ diff --git a/animation/SadTalker/src/utils/__pycache__/preprocess.cpython-38.pyc b/animation/SadTalker/src/utils/__pycache__/preprocess.cpython-38.pyc new file mode 100644 index 0000000..a6975af Binary files /dev/null and b/animation/SadTalker/src/utils/__pycache__/preprocess.cpython-38.pyc differ diff --git a/animation/SadTalker/src/utils/__pycache__/safetensor_helper.cpython-38.pyc b/animation/SadTalker/src/utils/__pycache__/safetensor_helper.cpython-38.pyc new file mode 100644 index 0000000..9061c81 Binary files /dev/null and b/animation/SadTalker/src/utils/__pycache__/safetensor_helper.cpython-38.pyc differ diff --git a/animation/SadTalker/src/utils/__pycache__/videoio.cpython-38.pyc b/animation/SadTalker/src/utils/__pycache__/videoio.cpython-38.pyc new file mode 100644 index 0000000..b124668 Binary files /dev/null and b/animation/SadTalker/src/utils/__pycache__/videoio.cpython-38.pyc differ diff --git a/animation/SadTalker/src/utils/audio.py b/animation/SadTalker/src/utils/audio.py new file mode 100644 index 0000000..89433eb --- /dev/null +++ b/animation/SadTalker/src/utils/audio.py @@ -0,0 +1,136 @@ +import librosa +import librosa.filters +import numpy as np +# import tensorflow as tf +from scipy import signal +from scipy.io import wavfile +from src.utils.hparams import hparams as hp + +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(sr=hp.sample_rate, n_fft=hp.n_fft, n_mels=hp.num_mels, + fmin=hp.fmin, fmax=hp.fmax) + +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) diff --git a/animation/SadTalker/src/utils/croper.py b/animation/SadTalker/src/utils/croper.py new file mode 100644 index 0000000..3d9a0ac --- /dev/null +++ b/animation/SadTalker/src/utils/croper.py @@ -0,0 +1,144 @@ +import os +import cv2 +import time +import glob +import argparse +import scipy +import numpy as np +from PIL import Image +import torch +from tqdm import tqdm +from itertools import cycle + +from src.face3d.extract_kp_videos_safe import KeypointExtractor +from facexlib.alignment import landmark_98_to_68 + +import numpy as np +from PIL import Image + +class Preprocesser: + def __init__(self, device='cuda'): + self.predictor = KeypointExtractor(device) + + def get_landmark(self, img_np): + """get landmark with dlib + :return: np.array shape=(68, 2) + """ + with torch.no_grad(): + dets = self.predictor.det_net.detect_faces(img_np, 0.97) + + if len(dets) == 0: + return None + det = dets[0] + + img = img_np[int(det[1]):int(det[3]), int(det[0]):int(det[2]), :] + lm = landmark_98_to_68(self.predictor.detector.get_landmarks(img)) # [0] + + #### keypoints to the original location + lm[:,0] += int(det[0]) + lm[:,1] += int(det[1]) + + return lm + + def align_face(self, img, lm, output_size=1024): + """ + :param filepath: str + :return: PIL Image + """ + lm_chin = lm[0: 17] # left-right + lm_eyebrow_left = lm[17: 22] # left-right + lm_eyebrow_right = lm[22: 27] # left-right + lm_nose = lm[27: 31] # top-down + lm_nostrils = lm[31: 36] # top-down + lm_eye_left = lm[36: 42] # left-clockwise + lm_eye_right = lm[42: 48] # left-clockwise + lm_mouth_outer = lm[48: 60] # left-clockwise + lm_mouth_inner = lm[60: 68] # left-clockwise + + # Calculate auxiliary vectors. + eye_left = np.mean(lm_eye_left, axis=0) + eye_right = np.mean(lm_eye_right, axis=0) + eye_avg = (eye_left + eye_right) * 0.5 + eye_to_eye = eye_right - eye_left + mouth_left = lm_mouth_outer[0] + mouth_right = lm_mouth_outer[6] + mouth_avg = (mouth_left + mouth_right) * 0.5 + eye_to_mouth = mouth_avg - eye_avg + + # Choose oriented crop rectangle. + x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1] # Addition of binocular difference and double mouth difference + x /= np.hypot(*x) # hypot函数计算直角三角形的斜边长，用斜边长对三角形两条直边做归一化 + x *= max(np.hypot(*eye_to_eye) * 2.0, np.hypot(*eye_to_mouth) * 1.8) # 双眼差和眼嘴差，选较大的作为基准尺度 + y = np.flipud(x) * [-1, 1] + c = eye_avg + eye_to_mouth * 0.1 + quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y]) # 定义四边形，以面部基准位置为中心上下左右平移得到四个顶点 + qsize = np.hypot(*x) * 2 # 定义四边形的大小（边长），为基准尺度的2倍 + + # Shrink. + # 如果计算出的四边形太大了，就按比例缩小它 + shrink = int(np.floor(qsize / output_size * 0.5)) + if shrink > 1: + rsize = (int(np.rint(float(img.size[0]) / shrink)), int(np.rint(float(img.size[1]) / shrink))) + img = img.resize(rsize, Image.ANTIALIAS) + quad /= shrink + qsize /= shrink + else: + rsize = (int(np.rint(float(img.size[0]))), int(np.rint(float(img.size[1])))) + + # Crop. + border = max(int(np.rint(qsize * 0.1)), 3) + crop = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))), + int(np.ceil(max(quad[:, 1])))) + crop = (max(crop[0] - border, 0), max(crop[1] - border, 0), min(crop[2] + border, img.size[0]), + min(crop[3] + border, img.size[1])) + if crop[2] - crop[0] < img.size[0] or crop[3] - crop[1] < img.size[1]: + # img = img.crop(crop) + quad -= crop[0:2] + + # Pad. + pad = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))), + int(np.ceil(max(quad[:, 1])))) + pad = (max(-pad[0] + border, 0), max(-pad[1] + border, 0), max(pad[2] - img.size[0] + border, 0), + max(pad[3] - img.size[1] + border, 0)) + # if enable_padding and max(pad) > border - 4: + # pad = np.maximum(pad, int(np.rint(qsize * 0.3))) + # img = np.pad(np.float32(img), ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'reflect') + # h, w, _ = img.shape + # y, x, _ = np.ogrid[:h, :w, :1] + # mask = np.maximum(1.0 - np.minimum(np.float32(x) / pad[0], np.float32(w - 1 - x) / pad[2]), + # 1.0 - np.minimum(np.float32(y) / pad[1], np.float32(h - 1 - y) / pad[3])) + # blur = qsize * 0.02 + # img += (scipy.ndimage.gaussian_filter(img, [blur, blur, 0]) - img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0) + # img += (np.median(img, axis=(0, 1)) - img) * np.clip(mask, 0.0, 1.0) + # img = Image.fromarray(np.uint8(np.clip(np.rint(img), 0, 255)), 'RGB') + # quad += pad[:2] + + # Transform. + quad = (quad + 0.5).flatten() + lx = max(min(quad[0], quad[2]), 0) + ly = max(min(quad[1], quad[7]), 0) + rx = min(max(quad[4], quad[6]), img.size[0]) + ry = min(max(quad[3], quad[5]), img.size[0]) + + # Save aligned image. + return rsize, crop, [lx, ly, rx, ry] + + def crop(self, img_np_list, still=False, xsize=512): # first frame for all video + img_np = img_np_list[0] + lm = self.get_landmark(img_np) + + if lm is None: + raise 'can not detect the landmark from source image' + rsize, crop, quad = self.align_face(img=Image.fromarray(img_np), lm=lm, output_size=xsize) + clx, cly, crx, cry = crop + lx, ly, rx, ry = quad + lx, ly, rx, ry = int(lx), int(ly), int(rx), int(ry) + for _i in range(len(img_np_list)): + _inp = img_np_list[_i] + _inp = cv2.resize(_inp, (rsize[0], rsize[1])) + _inp = _inp[cly:cry, clx:crx] + if not still: + _inp = _inp[ly:ry, lx:rx] + img_np_list[_i] = _inp + return img_np_list, crop, quad + diff --git a/animation/SadTalker/src/utils/face_enhancer.py b/animation/SadTalker/src/utils/face_enhancer.py new file mode 100644 index 0000000..15851a1 --- /dev/null +++ b/animation/SadTalker/src/utils/face_enhancer.py @@ -0,0 +1,123 @@ +import os +import torch + +from gfpgan import GFPGANer + +from tqdm import tqdm + +from src.utils.videoio import load_video_to_cv2 + +import cv2 + + +class GeneratorWithLen(object): + """ From https://stackoverflow.com/a/7460929 """ + + def __init__(self, gen, length): + self.gen = gen + self.length = length + + def __len__(self): + return self.length + + def __iter__(self): + return self.gen + +def enhancer_list(images, method='gfpgan', bg_upsampler='realesrgan'): + gen = enhancer_generator_no_len(images, method=method, bg_upsampler=bg_upsampler) + return list(gen) + +def enhancer_generator_with_len(images, method='gfpgan', bg_upsampler='realesrgan'): + """ Provide a generator with a __len__ method so that it can passed to functions that + call len()""" + + if os.path.isfile(images): # handle video to images + # TODO: Create a generator version of load_video_to_cv2 + images = load_video_to_cv2(images) + + gen = enhancer_generator_no_len(images, method=method, bg_upsampler=bg_upsampler) + gen_with_len = GeneratorWithLen(gen, len(images)) + return gen_with_len + +def enhancer_generator_no_len(images, method='gfpgan', bg_upsampler='realesrgan'): + """ Provide a generator function so that all of the enhanced images don't need + to be stored in memory at the same time. This can save tons of RAM compared to + the enhancer function. """ + + print('face enhancer....') + if not isinstance(images, list) and os.path.isfile(images): # handle video to images + images = load_video_to_cv2(images) + + # ------------------------ set up GFPGAN restorer ------------------------ + if method == 'gfpgan': + arch = 'clean' + channel_multiplier = 2 + model_name = 'GFPGANv1.4' + url = 'https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.4.pth' + elif method == 'RestoreFormer': + arch = 'RestoreFormer' + channel_multiplier = 2 + model_name = 'RestoreFormer' + url = 'https://github.com/TencentARC/GFPGAN/releases/download/v1.3.4/RestoreFormer.pth' + elif method == 'codeformer': # TODO: + arch = 'CodeFormer' + channel_multiplier = 2 + model_name = 'CodeFormer' + url = 'https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/codeformer.pth' + else: + raise ValueError(f'Wrong model version {method}.') + + + # ------------------------ set up background upsampler ------------------------ + if bg_upsampler == 'realesrgan': + if not torch.cuda.is_available(): # CPU + import warnings + warnings.warn('The unoptimized RealESRGAN is slow on CPU. We do not use it. ' + 'If you really want to use it, please modify the corresponding codes.') + bg_upsampler = None + else: + from basicsr.archs.rrdbnet_arch import RRDBNet + from realesrgan import RealESRGANer + model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=2) + bg_upsampler = RealESRGANer( + scale=2, + model_path='https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth', + model=model, + tile=400, + tile_pad=10, + pre_pad=0, + half=True) # need to set False in CPU mode + else: + bg_upsampler = None + + # determine model paths + model_path = os.path.join('gfpgan/weights', model_name + '.pth') + + if not os.path.isfile(model_path): + model_path = os.path.join('checkpoints', model_name + '.pth') + + if not os.path.isfile(model_path): + # download pre-trained models from url + model_path = url + + restorer = GFPGANer( + model_path=model_path, + upscale=2, + arch=arch, + channel_multiplier=channel_multiplier, + bg_upsampler=bg_upsampler) + + # ------------------------ restore ------------------------ + for idx in tqdm(range(len(images)), 'Face Enhancer:'): + + img = cv2.cvtColor(images[idx], cv2.COLOR_RGB2BGR) + + # restore faces and background if necessary + cropped_faces, restored_faces, r_img = restorer.enhance( + img, + has_aligned=False, + only_center_face=False, + paste_back=True) + + r_img = cv2.cvtColor(r_img, cv2.COLOR_BGR2RGB) + yield r_img diff --git a/animation/SadTalker/src/utils/hparams.py b/animation/SadTalker/src/utils/hparams.py new file mode 100644 index 0000000..743c5c7 --- /dev/null +++ b/animation/SadTalker/src/utils/hparams.py @@ -0,0 +1,160 @@ +from glob import glob +import os + +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 ) + + 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=300000, ### ctrl + c, stop whenever eval loss is consistently greater than train loss for ~10 epochs + num_workers=20, + checkpoint_interval=3000, + eval_interval=3000, + writer_interval=300, + 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=1000, + syncnet_checkpoint_interval=10000, + + disc_wt=0.07, + disc_initial_learning_rate=1e-4, +) + + + +# Default hyperparameters +hparamsdebug = 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 ) + + 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=2, + initial_learning_rate=1e-3, + nepochs=100000, ### ctrl + c, stop whenever eval loss is consistently greater than train loss for ~10 epochs + num_workers=0, + checkpoint_interval=10000, + eval_interval=10, + writer_interval=5, + 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) diff --git a/animation/SadTalker/src/utils/init_path.py b/animation/SadTalker/src/utils/init_path.py new file mode 100644 index 0000000..5f38d11 --- /dev/null +++ b/animation/SadTalker/src/utils/init_path.py @@ -0,0 +1,47 @@ +import os +import glob + +def init_path(checkpoint_dir, config_dir, size=512, old_version=False, preprocess='crop'): + + if old_version: + #### load all the checkpoint of `pth` + sadtalker_paths = { + 'wav2lip_checkpoint' : os.path.join(checkpoint_dir, 'wav2lip.pth'), + 'audio2pose_checkpoint' : os.path.join(checkpoint_dir, 'auido2pose_00140-model.pth'), + 'audio2exp_checkpoint' : os.path.join(checkpoint_dir, 'auido2exp_00300-model.pth'), + 'free_view_checkpoint' : os.path.join(checkpoint_dir, 'facevid2vid_00189-model.pth.tar'), + 'path_of_net_recon_model' : os.path.join(checkpoint_dir, 'epoch_20.pth') + } + + use_safetensor = False + elif len(glob.glob(os.path.join(checkpoint_dir, '*.safetensors'))): + print('using safetensor as default') + sadtalker_paths = { + "checkpoint":os.path.join(checkpoint_dir, 'SadTalker_V0.0.2_'+str(size)+'.safetensors'), + } + use_safetensor = True + else: + print("WARNING: The new version of the model will be updated by safetensor, you may need to download it mannully. We run the old version of the checkpoint this time!") + use_safetensor = False + + sadtalker_paths = { + 'wav2lip_checkpoint' : os.path.join(checkpoint_dir, 'wav2lip.pth'), + 'audio2pose_checkpoint' : os.path.join(checkpoint_dir, 'auido2pose_00140-model.pth'), + 'audio2exp_checkpoint' : os.path.join(checkpoint_dir, 'auido2exp_00300-model.pth'), + 'free_view_checkpoint' : os.path.join(checkpoint_dir, 'facevid2vid_00189-model.pth.tar'), + 'path_of_net_recon_model' : os.path.join(checkpoint_dir, 'epoch_20.pth') + } + + sadtalker_paths['dir_of_BFM_fitting'] = os.path.join(config_dir) # , 'BFM_Fitting' + sadtalker_paths['audio2pose_yaml_path'] = os.path.join(config_dir, 'auido2pose.yaml') + sadtalker_paths['audio2exp_yaml_path'] = os.path.join(config_dir, 'auido2exp.yaml') + sadtalker_paths['use_safetensor'] = use_safetensor # os.path.join(config_dir, 'auido2exp.yaml') + + if 'full' in preprocess: + sadtalker_paths['mappingnet_checkpoint'] = os.path.join(checkpoint_dir, 'mapping_00109-model.pth.tar') + sadtalker_paths['facerender_yaml'] = os.path.join(config_dir, 'facerender_still.yaml') + else: + sadtalker_paths['mappingnet_checkpoint'] = os.path.join(checkpoint_dir, 'mapping_00229-model.pth.tar') + sadtalker_paths['facerender_yaml'] = os.path.join(config_dir, 'facerender.yaml') + + return sadtalker_paths \ No newline at end of file diff --git a/animation/SadTalker/src/utils/model2safetensor.py b/animation/SadTalker/src/utils/model2safetensor.py new file mode 100644 index 0000000..50c4850 --- /dev/null +++ b/animation/SadTalker/src/utils/model2safetensor.py @@ -0,0 +1,141 @@ +import torch +import yaml +import os + +import safetensors +from safetensors.torch import save_file +from yacs.config import CfgNode as CN +import sys + +sys.path.append('/apdcephfs/private_shadowcun/SadTalker') + +from src.face3d.models import networks + +from src.facerender.modules.keypoint_detector import HEEstimator, KPDetector +from src.facerender.modules.mapping import MappingNet +from src.facerender.modules.generator import OcclusionAwareGenerator, OcclusionAwareSPADEGenerator + +from src.audio2pose_models.audio2pose import Audio2Pose +from src.audio2exp_models.networks import SimpleWrapperV2 +from src.test_audio2coeff import load_cpk + +size = 256 +############ face vid2vid +config_path = os.path.join('src', 'config', 'facerender.yaml') +current_root_path = '.' + +path_of_net_recon_model = os.path.join(current_root_path, 'checkpoints', 'epoch_20.pth') +net_recon = networks.define_net_recon(net_recon='resnet50', use_last_fc=False, init_path='') +checkpoint = torch.load(path_of_net_recon_model, map_location='cpu') +net_recon.load_state_dict(checkpoint['net_recon']) + +with open(config_path) as f: + config = yaml.safe_load(f) + +generator = OcclusionAwareSPADEGenerator(**config['model_params']['generator_params'], + **config['model_params']['common_params']) +kp_extractor = KPDetector(**config['model_params']['kp_detector_params'], + **config['model_params']['common_params']) +he_estimator = HEEstimator(**config['model_params']['he_estimator_params'], + **config['model_params']['common_params']) +mapping = MappingNet(**config['model_params']['mapping_params']) + +def load_cpk_facevid2vid(checkpoint_path, generator=None, discriminator=None, + kp_detector=None, he_estimator=None, optimizer_generator=None, + optimizer_discriminator=None, optimizer_kp_detector=None, + optimizer_he_estimator=None, device="cpu"): + + checkpoint = torch.load(checkpoint_path, map_location=torch.device(device)) + if generator is not None: + generator.load_state_dict(checkpoint['generator']) + if kp_detector is not None: + kp_detector.load_state_dict(checkpoint['kp_detector']) + if he_estimator is not None: + he_estimator.load_state_dict(checkpoint['he_estimator']) + if discriminator is not None: + try: + discriminator.load_state_dict(checkpoint['discriminator']) + except: + print ('No discriminator in the state-dict. Dicriminator will be randomly initialized') + if optimizer_generator is not None: + optimizer_generator.load_state_dict(checkpoint['optimizer_generator']) + if optimizer_discriminator is not None: + try: + optimizer_discriminator.load_state_dict(checkpoint['optimizer_discriminator']) + except RuntimeError as e: + print ('No discriminator optimizer in the state-dict. Optimizer will be not initialized') + if optimizer_kp_detector is not None: + optimizer_kp_detector.load_state_dict(checkpoint['optimizer_kp_detector']) + if optimizer_he_estimator is not None: + optimizer_he_estimator.load_state_dict(checkpoint['optimizer_he_estimator']) + + return checkpoint['epoch'] + + +def load_cpk_facevid2vid_safetensor(checkpoint_path, generator=None, + kp_detector=None, he_estimator=None, + device="cpu"): + + checkpoint = safetensors.torch.load_file(checkpoint_path) + + if generator is not None: + x_generator = {} + for k,v in checkpoint.items(): + if 'generator' in k: + x_generator[k.replace('generator.', '')] = v + generator.load_state_dict(x_generator) + if kp_detector is not None: + x_generator = {} + for k,v in checkpoint.items(): + if 'kp_extractor' in k: + x_generator[k.replace('kp_extractor.', '')] = v + kp_detector.load_state_dict(x_generator) + if he_estimator is not None: + x_generator = {} + for k,v in checkpoint.items(): + if 'he_estimator' in k: + x_generator[k.replace('he_estimator.', '')] = v + he_estimator.load_state_dict(x_generator) + + return None + +free_view_checkpoint = '/apdcephfs/private_shadowcun/SadTalker/checkpoints/facevid2vid_'+str(size)+'-model.pth.tar' +load_cpk_facevid2vid(free_view_checkpoint, kp_detector=kp_extractor, generator=generator, he_estimator=he_estimator) + +wav2lip_checkpoint = os.path.join(current_root_path, 'checkpoints', 'wav2lip.pth') + +audio2pose_checkpoint = os.path.join(current_root_path, 'checkpoints', 'auido2pose_00140-model.pth') +audio2pose_yaml_path = os.path.join(current_root_path, 'src', 'config', 'auido2pose.yaml') + +audio2exp_checkpoint = os.path.join(current_root_path, 'checkpoints', 'auido2exp_00300-model.pth') +audio2exp_yaml_path = os.path.join(current_root_path, 'src', 'config', 'auido2exp.yaml') + +fcfg_pose = open(audio2pose_yaml_path) +cfg_pose = CN.load_cfg(fcfg_pose) +cfg_pose.freeze() +audio2pose_model = Audio2Pose(cfg_pose, wav2lip_checkpoint) +audio2pose_model.eval() +load_cpk(audio2pose_checkpoint, model=audio2pose_model, device='cpu') + +# load audio2exp_model +netG = SimpleWrapperV2() +netG.eval() +load_cpk(audio2exp_checkpoint, model=netG, device='cpu') + +class SadTalker(torch.nn.Module): + def __init__(self, kp_extractor, generator, netG, audio2pose, face_3drecon): + super(SadTalker, self).__init__() + self.kp_extractor = kp_extractor + self.generator = generator + self.audio2exp = netG + self.audio2pose = audio2pose + self.face_3drecon = face_3drecon + + +model = SadTalker(kp_extractor, generator, netG, audio2pose_model, net_recon) + +# here, we want to convert it to safetensor +save_file(model.state_dict(), "checkpoints/SadTalker_V0.0.2_"+str(size)+".safetensors") + +### test +load_cpk_facevid2vid_safetensor('checkpoints/SadTalker_V0.0.2_'+str(size)+'.safetensors', kp_detector=kp_extractor, generator=generator, he_estimator=None) \ No newline at end of file diff --git a/animation/SadTalker/src/utils/paste_pic.py b/animation/SadTalker/src/utils/paste_pic.py new file mode 100644 index 0000000..f9989e2 --- /dev/null +++ b/animation/SadTalker/src/utils/paste_pic.py @@ -0,0 +1,69 @@ +import cv2, os +import numpy as np +from tqdm import tqdm +import uuid + +from src.utils.videoio import save_video_with_watermark + +def paste_pic(video_path, pic_path, crop_info, new_audio_path, full_video_path, extended_crop=False): + + if not os.path.isfile(pic_path): + raise ValueError('pic_path must be a valid path to video/image file') + elif pic_path.split('.')[-1] in ['jpg', 'png', 'jpeg']: + # loader for first frame + full_img = cv2.imread(pic_path) + else: + # loader for videos + video_stream = cv2.VideoCapture(pic_path) + fps = video_stream.get(cv2.CAP_PROP_FPS) + full_frames = [] + while 1: + still_reading, frame = video_stream.read() + if not still_reading: + video_stream.release() + break + break + full_img = frame + frame_h = full_img.shape[0] + frame_w = full_img.shape[1] + + video_stream = cv2.VideoCapture(video_path) + fps = video_stream.get(cv2.CAP_PROP_FPS) + crop_frames = [] + while 1: + still_reading, frame = video_stream.read() + if not still_reading: + video_stream.release() + break + crop_frames.append(frame) + + if len(crop_info) != 3: + print("you didn't crop the image") + return + else: + r_w, r_h = crop_info[0] + clx, cly, crx, cry = crop_info[1] + lx, ly, rx, ry = crop_info[2] + lx, ly, rx, ry = int(lx), int(ly), int(rx), int(ry) + # oy1, oy2, ox1, ox2 = cly+ly, cly+ry, clx+lx, clx+rx + # oy1, oy2, ox1, ox2 = cly+ly, cly+ry, clx+lx, clx+rx + + if extended_crop: + oy1, oy2, ox1, ox2 = cly, cry, clx, crx + else: + oy1, oy2, ox1, ox2 = cly+ly, cly+ry, clx+lx, clx+rx + + tmp_path = str(uuid.uuid4())+'.mp4' + out_tmp = cv2.VideoWriter(tmp_path, cv2.VideoWriter_fourcc(*'MP4V'), fps, (frame_w, frame_h)) + for crop_frame in tqdm(crop_frames, 'seamlessClone:'): + p = cv2.resize(crop_frame.astype(np.uint8), (ox2-ox1, oy2 - oy1)) + + mask = 255*np.ones(p.shape, p.dtype) + location = ((ox1+ox2) // 2, (oy1+oy2) // 2) + gen_img = cv2.seamlessClone(p, full_img, mask, location, cv2.NORMAL_CLONE) + out_tmp.write(gen_img) + + out_tmp.release() + + save_video_with_watermark(tmp_path, new_audio_path, full_video_path, watermark=False) + os.remove(tmp_path) diff --git a/animation/SadTalker/src/utils/preprocess.py b/animation/SadTalker/src/utils/preprocess.py new file mode 100644 index 0000000..0f784e6 --- /dev/null +++ b/animation/SadTalker/src/utils/preprocess.py @@ -0,0 +1,170 @@ +import numpy as np +import cv2, os, sys, torch +from tqdm import tqdm +from PIL import Image + +# 3dmm extraction +import safetensors +import safetensors.torch +from src.face3d.util.preprocess import align_img +from src.face3d.util.load_mats import load_lm3d +from src.face3d.models import networks + +from scipy.io import loadmat, savemat +from src.utils.croper import Preprocesser + + +import warnings + +from src.utils.safetensor_helper import load_x_from_safetensor +warnings.filterwarnings("ignore") + +def split_coeff(coeffs): + """ + Return: + coeffs_dict -- a dict of torch.tensors + + Parameters: + coeffs -- torch.tensor, size (B, 256) + """ + id_coeffs = coeffs[:, :80] + exp_coeffs = coeffs[:, 80: 144] + tex_coeffs = coeffs[:, 144: 224] + angles = coeffs[:, 224: 227] + gammas = coeffs[:, 227: 254] + translations = coeffs[:, 254:] + return { + 'id': id_coeffs, + 'exp': exp_coeffs, + 'tex': tex_coeffs, + 'angle': angles, + 'gamma': gammas, + 'trans': translations + } + + +class CropAndExtract(): + def __init__(self, sadtalker_path, device): + + self.propress = Preprocesser(device) + self.net_recon = networks.define_net_recon(net_recon='resnet50', use_last_fc=False, init_path='').to(device) + + if sadtalker_path['use_safetensor']: + checkpoint = safetensors.torch.load_file(sadtalker_path['checkpoint']) + self.net_recon.load_state_dict(load_x_from_safetensor(checkpoint, 'face_3drecon')) + else: + checkpoint = torch.load(sadtalker_path['path_of_net_recon_model'], map_location=torch.device(device)) + self.net_recon.load_state_dict(checkpoint['net_recon']) + + self.net_recon.eval() + self.lm3d_std = load_lm3d(sadtalker_path['dir_of_BFM_fitting']) + self.device = device + + def generate(self, input_path, save_dir, crop_or_resize='crop', source_image_flag=False, pic_size=256): + + pic_name = os.path.splitext(os.path.split(input_path)[-1])[0] + + landmarks_path = os.path.join(save_dir, pic_name+'_landmarks.txt') + coeff_path = os.path.join(save_dir, pic_name+'.mat') + png_path = os.path.join(save_dir, pic_name+'.png') + + #load input + if not os.path.isfile(input_path): + raise ValueError('input_path must be a valid path to video/image file') + elif input_path.split('.')[-1] in ['jpg', 'png', 'jpeg']: + # loader for first frame + full_frames = [cv2.imread(input_path)] + fps = 25 + else: + # loader for videos + video_stream = cv2.VideoCapture(input_path) + fps = video_stream.get(cv2.CAP_PROP_FPS) + full_frames = [] + while 1: + still_reading, frame = video_stream.read() + if not still_reading: + video_stream.release() + break + full_frames.append(frame) + if source_image_flag: + break + + x_full_frames= [cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) for frame in full_frames] + + #### crop images as the + if 'crop' in crop_or_resize.lower(): # default crop + x_full_frames, crop, quad = self.propress.crop(x_full_frames, still=True if 'ext' in crop_or_resize.lower() else False, xsize=512) + clx, cly, crx, cry = crop + lx, ly, rx, ry = quad + lx, ly, rx, ry = int(lx), int(ly), int(rx), int(ry) + oy1, oy2, ox1, ox2 = cly+ly, cly+ry, clx+lx, clx+rx + crop_info = ((ox2 - ox1, oy2 - oy1), crop, quad) + elif 'full' in crop_or_resize.lower(): + x_full_frames, crop, quad = self.propress.crop(x_full_frames, still=True if 'ext' in crop_or_resize.lower() else False, xsize=512) + clx, cly, crx, cry = crop + lx, ly, rx, ry = quad + lx, ly, rx, ry = int(lx), int(ly), int(rx), int(ry) + oy1, oy2, ox1, ox2 = cly+ly, cly+ry, clx+lx, clx+rx + crop_info = ((ox2 - ox1, oy2 - oy1), crop, quad) + else: # resize mode + oy1, oy2, ox1, ox2 = 0, x_full_frames[0].shape[0], 0, x_full_frames[0].shape[1] + crop_info = ((ox2 - ox1, oy2 - oy1), None, None) + + frames_pil = [Image.fromarray(cv2.resize(frame,(pic_size, pic_size))) for frame in x_full_frames] + if len(frames_pil) == 0: + print('No face is detected in the input file') + return None, None + + # save crop info + for frame in frames_pil: + cv2.imwrite(png_path, cv2.cvtColor(np.array(frame), cv2.COLOR_RGB2BGR)) + + # 2. get the landmark according to the detected face. + if not os.path.isfile(landmarks_path): + lm = self.propress.predictor.extract_keypoint(frames_pil, landmarks_path) + else: + print(' Using saved landmarks.') + lm = np.loadtxt(landmarks_path).astype(np.float32) + lm = lm.reshape([len(x_full_frames), -1, 2]) + + if not os.path.isfile(coeff_path): + # load 3dmm paramter generator from Deep3DFaceRecon_pytorch + video_coeffs, full_coeffs = [], [] + for idx in tqdm(range(len(frames_pil)), desc='3DMM Extraction In Video:'): + frame = frames_pil[idx] + W,H = frame.size + lm1 = lm[idx].reshape([-1, 2]) + + if np.mean(lm1) == -1: + lm1 = (self.lm3d_std[:, :2]+1)/2. + lm1 = np.concatenate( + [lm1[:, :1]*W, lm1[:, 1:2]*H], 1 + ) + else: + lm1[:, -1] = H - 1 - lm1[:, -1] + + trans_params, im1, lm1, _ = align_img(frame, lm1, self.lm3d_std) + + trans_params = np.array([float(item) for item in np.hsplit(trans_params, 5)]).astype(np.float32) + im_t = torch.tensor(np.array(im1)/255., dtype=torch.float32).permute(2, 0, 1).to(self.device).unsqueeze(0) + + with torch.no_grad(): + full_coeff = self.net_recon(im_t) + coeffs = split_coeff(full_coeff) + + pred_coeff = {key:coeffs[key].cpu().numpy() for key in coeffs} + + pred_coeff = np.concatenate([ + pred_coeff['exp'], + pred_coeff['angle'], + pred_coeff['trans'], + trans_params[2:][None], + ], 1) + video_coeffs.append(pred_coeff) + full_coeffs.append(full_coeff.cpu().numpy()) + + semantic_npy = np.array(video_coeffs)[:,0] + + savemat(coeff_path, {'coeff_3dmm': semantic_npy, 'full_3dmm': np.array(full_coeffs)[0]}) + + return coeff_path, png_path, crop_info diff --git a/animation/SadTalker/src/utils/safetensor_helper.py b/animation/SadTalker/src/utils/safetensor_helper.py new file mode 100644 index 0000000..3cdbdd2 --- /dev/null +++ b/animation/SadTalker/src/utils/safetensor_helper.py @@ -0,0 +1,8 @@ + + +def load_x_from_safetensor(checkpoint, key): + x_generator = {} + for k,v in checkpoint.items(): + if key in k: + x_generator[k.replace(key+'.', '')] = v + return x_generator \ No newline at end of file diff --git a/animation/SadTalker/src/utils/text2speech.py b/animation/SadTalker/src/utils/text2speech.py new file mode 100644 index 0000000..a9cd9f5 --- /dev/null +++ b/animation/SadTalker/src/utils/text2speech.py @@ -0,0 +1,20 @@ +import os +import tempfile +from TTS.api import TTS + + +class TTSTalker(): + def __init__(self) -> None: + model_name = TTS().list_models()[0] + self.tts = TTS(model_name) + + def test(self, text, language='en'): + + tempf = tempfile.NamedTemporaryFile( + delete = False, + suffix = ('.'+'wav'), + ) + + self.tts.tts_to_file(text, speaker=self.tts.speakers[0], language=language, file_path=tempf.name) + + return tempf.name diff --git a/animation/SadTalker/src/utils/videoio.py b/animation/SadTalker/src/utils/videoio.py new file mode 100644 index 0000000..08bfbdd --- /dev/null +++ b/animation/SadTalker/src/utils/videoio.py @@ -0,0 +1,41 @@ +import shutil +import uuid + +import os + +import cv2 + +def load_video_to_cv2(input_path): + video_stream = cv2.VideoCapture(input_path) + fps = video_stream.get(cv2.CAP_PROP_FPS) + full_frames = [] + while 1: + still_reading, frame = video_stream.read() + if not still_reading: + video_stream.release() + break + full_frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)) + return full_frames + +def save_video_with_watermark(video, audio, save_path, watermark=False): + temp_file = str(uuid.uuid4())+'.mp4' + cmd = r'ffmpeg -y -hide_banner -loglevel error -i "%s" -i "%s" -vcodec copy "%s"' % (video, audio, temp_file) + os.system(cmd) + + if watermark is False: + shutil.move(temp_file, save_path) + else: + # watermark + try: + ##### check if stable-diffusion-webui + import webui + from modules import paths + watarmark_path = paths.script_path+"/extensions/SadTalker/docs/sadtalker_logo.png" + except: + # get the root path of sadtalker. + dir_path = os.path.dirname(os.path.realpath(__file__)) + watarmark_path = dir_path+"/../../docs/sadtalker_logo.png" + + cmd = r'ffmpeg -y -hide_banner -loglevel error -i "%s" -i "%s" -filter_complex "[1]scale=100:-1[wm];[0][wm]overlay=(main_w-overlay_w)-10:10" "%s"' % (temp_file, watarmark_path, save_path) + os.system(cmd) + os.remove(temp_file) \ No newline at end of file