CoolFace
Apppublic

vlanding/pizda

sourceHugging Faceupdated 3y agoView on Hugging Face
0likes
app.py262 linesDownload Raw Back to root
1import cv22import onnxruntime as rt3import sys4from insightface.app import FaceAnalysis5sys.path.insert(1, './recognition')6from scrfd import SCRFD7from arcface_onnx import ArcFaceONNX8import os.path as osp9import os10from pathlib import Path11from tqdm import tqdm12import ffmpeg13import random14import multiprocessing as mp15from concurrent.futures import ThreadPoolExecutor16from insightface.model_zoo.inswapper import INSwapper17import psutil18from enum import Enum19from insightface.app.common import Face20from insightface.utils.storage import ensure_available21import re22import subprocess23 24class RefacerMode(Enum):25     CPU, CUDA, COREML, TENSORRT = range(1, 5)26 27class Refacer:28    def __init__(self,force_cpu=False,colab_performance=False):29        self.first_face = False30        self.force_cpu = force_cpu31        self.colab_performance = colab_performance32        self.__check_encoders()33        self.__check_providers()34        self.total_mem = psutil.virtual_memory().total35        self.__init_apps()36 37    def __check_providers(self):38        if self.force_cpu :39            self.providers = ['CPUExecutionProvider']40        else:41            self.providers = rt.get_available_providers()42        rt.set_default_logger_severity(4)43        self.sess_options = rt.SessionOptions()44        self.sess_options.execution_mode = rt.ExecutionMode.ORT_SEQUENTIAL45        self.sess_options.graph_optimization_level = rt.GraphOptimizationLevel.ORT_ENABLE_ALL46 47        if len(self.providers) == 1 and 'CPUExecutionProvider' in self.providers:48            self.mode = RefacerMode.CPU49            self.use_num_cpus = mp.cpu_count()-150            self.sess_options.intra_op_num_threads = int(self.use_num_cpus/3)51            print(f"CPU mode with providers {self.providers}")52        elif self.colab_performance:53            self.mode = RefacerMode.TENSORRT54            self.use_num_cpus = mp.cpu_count()-155            self.sess_options.intra_op_num_threads = int(self.use_num_cpus/3)56            print(f"TENSORRT mode with providers {self.providers}")57        elif 'CoreMLExecutionProvider' in self.providers:58            self.mode = RefacerMode.COREML59            self.use_num_cpus = mp.cpu_count()-160            self.sess_options.intra_op_num_threads = int(self.use_num_cpus/3)61            print(f"CoreML mode with providers {self.providers}")62        elif 'CUDAExecutionProvider' in self.providers:63            self.mode = RefacerMode.CUDA64            self.use_num_cpus = 265            self.sess_options.intra_op_num_threads = 166            if 'TensorrtExecutionProvider' in self.providers:67                self.providers.remove('TensorrtExecutionProvider')68            print(f"CUDA mode with providers {self.providers}")69        """70        elif 'TensorrtExecutionProvider' in self.providers:71            self.mode = RefacerMode.TENSORRT72            #self.use_num_cpus = 173            #self.sess_options.intra_op_num_threads = 174            self.use_num_cpus = mp.cpu_count()-175            self.sess_options.intra_op_num_threads = int(self.use_num_cpus/3)76            print(f"TENSORRT mode with providers {self.providers}")77        """78        79 80    def __init_apps(self):81        assets_dir = ensure_available('models', 'buffalo_l', root='~/.insightface')82 83        model_path = os.path.join(assets_dir, 'det_10g.onnx')84        sess_face = rt.InferenceSession(model_path, self.sess_options, providers=self.providers)85        self.face_detector = SCRFD(model_path,sess_face)86        self.face_detector.prepare(0,input_size=(640, 640))87 88        model_path = os.path.join(assets_dir , 'w600k_r50.onnx')89        sess_rec = rt.InferenceSession(model_path, self.sess_options, providers=self.providers)90        self.rec_app = ArcFaceONNX(model_path,sess_rec)91        self.rec_app.prepare(0)92 93        model_path = 'inswapper_128.onnx'94        sess_swap = rt.InferenceSession(model_path, self.sess_options, providers=self.providers)95        self.face_swapper = INSwapper(model_path,sess_swap)96 97    def prepare_faces(self, faces):98        self.replacement_faces=[]99        for face in faces:100            #image1 = cv2.imread(face.origin)101            if "origin" in face:102                face_threshold = face['threshold']103                bboxes1, kpss1 = self.face_detector.autodetect(face['origin'], max_num=1)  104                if len(kpss1)<1:105                    raise Exception('No face detected on "Face to replace" image')106                feat_original = self.rec_app.get(face['origin'], kpss1[0])107            else:108                face_threshold = 0109                self.first_face = True110                feat_original = None111                print('No origin image: First face change')112            #image2 = cv2.imread(face.destination)113            _faces = self.__get_faces(face['destination'],max_num=1)114            if len(_faces)<1:115                raise Exception('No face detected on "Destination face" image')116            self.replacement_faces.append((feat_original,_faces[0],face_threshold))117 118    def __convert_video(self,video_path,output_video_path):119        if self.video_has_audio:120            print("Merging audio with the refaced video...")121            new_path = output_video_path + str(random.randint(0,999)) + "_c.mp4"122            #stream = ffmpeg.input(output_video_path)123            in1 = ffmpeg.input(output_video_path)124            in2 = ffmpeg.input(video_path)125            out = ffmpeg.output(in1.video, in2.audio, new_path,video_bitrate=self.ffmpeg_video_bitrate,vcodec=self.ffmpeg_video_encoder)126            out.run(overwrite_output=True,quiet=True)127        else:128            new_path = output_video_path129            print("The video doesn't have audio, so post-processing is not necessary")130        131        print(f"The process has finished.\nThe refaced video can be found at {os.path.abspath(new_path)}")132        return new_path133 134    def __get_faces(self,frame,max_num=0):135 136        bboxes, kpss = self.face_detector.detect(frame,max_num=max_num,metric='default')137 138        if bboxes.shape[0] == 0:139            return []140        ret = []141        for i in range(bboxes.shape[0]):142            bbox = bboxes[i, 0:4]143            det_score = bboxes[i, 4]144            kps = None145            if kpss is not None:146                kps = kpss[i]147            face = Face(bbox=bbox, kps=kps, det_score=det_score)148            face.embedding = self.rec_app.get(frame, kps)149            ret.append(face)150        return ret151 152    def process_first_face(self,frame):153        faces = self.__get_faces(frame,max_num=1)154        if len(faces) != 0:155            frame = self.face_swapper.get(frame, faces[0], self.replacement_faces[0][1], paste_back=True)156        return frame157 158    def process_faces(self,frame):159        faces = self.__get_faces(frame,max_num=0)160        for rep_face in self.replacement_faces:161            for i in range(len(faces) - 1, -1, -1):162                sim = self.rec_app.compute_sim(rep_face[0], faces[i].embedding)163                if sim>=rep_face[2]:164                    frame = self.face_swapper.get(frame, faces[i], rep_face[1], paste_back=True)165                    del faces[i]166                    break167        return frame168 169    def __check_video_has_audio(self,video_path):170        self.video_has_audio = False171        probe = ffmpeg.probe(video_path)172        audio_stream = next((stream for stream in probe['streams'] if stream['codec_type'] == 'audio'), None)173        if audio_stream is not None:174            self.video_has_audio = True175        176    def reface_group(self, faces, frames, output):177        with ThreadPoolExecutor(max_workers = self.use_num_cpus) as executor:178            if self.first_face:179                results = list(tqdm(executor.map(self.process_first_face, frames), total=len(frames),desc="Processing frames"))180            else:181                results = list(tqdm(executor.map(self.process_faces, frames), total=len(frames),desc="Processing frames"))182            for result in results:183                output.write(result)184 185    def reface(self, video_path, faces):186        self.__check_video_has_audio(video_path)187        output_video_path = os.path.join('out',Path(video_path).name)188        self.prepare_faces(faces)189 190        cap = cv2.VideoCapture(video_path)191        total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))192        print(f"Total frames: {total_frames}")193        194        fps = cap.get(cv2.CAP_PROP_FPS)195        frame_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))196        frame_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))197 198        fourcc = cv2.VideoWriter_fourcc(*'mp4v')199        output = cv2.VideoWriter(output_video_path, fourcc, fps, (frame_width, frame_height))200        201        frames=[]202        self.k = 1203        with tqdm(total=total_frames,desc="Extracting frames") as pbar:204            while cap.isOpened():205                flag, frame = cap.read()206                if flag and len(frame)>0:207                    frames.append(frame.copy())208                    pbar.update()209                else:210                    break211                if (len(frames) > 1000):212                    self.reface_group(faces,frames,output)213                    frames=[]214 215            cap.release()216            pbar.close()217 218        self.reface_group(faces,frames,output)219        frames=[]220        output.release()221        222        return self.__convert_video(video_path,output_video_path)223    224    def __try_ffmpeg_encoder(self, vcodec):225        print(f"Trying FFMPEG {vcodec} encoder")226        command = ['ffmpeg', '-y', '-f','lavfi','-i','testsrc=duration=1:size=1280x720:rate=30','-vcodec',vcodec,'testsrc.mp4']227        try:228            subprocess.run(command, check=True, capture_output=True).stderr229        except subprocess.CalledProcessError as e:230            print(f"FFMPEG {vcodec} encoder doesn't work -> Disabled.")231            return False232        print(f"FFMPEG {vcodec} encoder works")233        return True234        235    def __check_encoders(self):236        self.ffmpeg_video_encoder='libx264'237        self.ffmpeg_video_bitrate='0'238 239        pattern = r"encoders: ([a-zA-Z0-9_]+(?: [a-zA-Z0-9_]+)*)"240        command = ['ffmpeg', '-codecs', '--list-encoders']241        commandout = subprocess.run(command, check=True, capture_output=True).stdout242        result = commandout.decode('utf-8').split('\n')243        for r in result:244            if "264" in r: 245                encoders = re.search(pattern, r).group(1).split(' ')246                for v_c in Refacer.VIDEO_CODECS:247                    for v_k in encoders:248                        if v_c == v_k:249                            if self.__try_ffmpeg_encoder(v_k):250                                self.ffmpeg_video_encoder=v_k251                                self.ffmpeg_video_bitrate=Refacer.VIDEO_CODECS[v_k]252                                print(f"Video codec for FFMPEG: {self.ffmpeg_video_encoder}")253                                return254 255    VIDEO_CODECS = {256         'h264_videotoolbox':'0', #osx HW acceleration257         'h264_nvenc':'0', #NVIDIA HW acceleration258         #'h264_qsv', #Intel HW acceleration259         #'h264_vaapi', #Intel HW acceleration260         #'h264_omx', #HW acceleration261         'libx264':'0' #No HW acceleration262    }