Real-ESRGAN

1
import argparse
2
import cv2
3
import glob
4
import mimetypes
5
import numpy as np
6
import os
7
import shutil
8
import subprocess
9
import torch
10
from basicsr.archs.rrdbnet_arch import RRDBNet
11
from basicsr.utils.download_util import load_file_from_url
12
from os import path as osp
13
from tqdm import tqdm
14

15
from realesrgan import RealESRGANer
16
from realesrgan.archs.srvgg_arch import SRVGGNetCompact
17

18
try:
19
    import ffmpeg
20
except ImportError:
21
    import pip
22
    pip.main(['install', '--user', 'ffmpeg-python'])
23
    import ffmpeg
24

25

26
def get_video_meta_info(video_path):
27
    ret = {}
28
    probe = ffmpeg.probe(video_path)
29
    video_streams = [stream for stream in probe['streams'] if stream['codec_type'] == 'video']
30
    has_audio = any(stream['codec_type'] == 'audio' for stream in probe['streams'])
31
    ret['width'] = video_streams[0]['width']
32
    ret['height'] = video_streams[0]['height']
33
    ret['fps'] = eval(video_streams[0]['avg_frame_rate'])
34
    ret['audio'] = ffmpeg.input(video_path).audio if has_audio else None
35
    ret['nb_frames'] = int(video_streams[0]['nb_frames'])
36
    return ret
37

38

39
def get_sub_video(args, num_process, process_idx):
40
    if num_process == 1:
41
        return args.input
42
    meta = get_video_meta_info(args.input)
43
    duration = int(meta['nb_frames'] / meta['fps'])
44
    part_time = duration // num_process
45
    print(f'duration: {duration}, part_time: {part_time}')
46
    os.makedirs(osp.join(args.output, f'{args.video_name}_inp_tmp_videos'), exist_ok=True)
47
    out_path = osp.join(args.output, f'{args.video_name}_inp_tmp_videos', f'{process_idx:03d}.mp4')
48
    cmd = [
49
        args.ffmpeg_bin, f'-i {args.input}', '-ss', f'{part_time * process_idx}',
50
        f'-to {part_time * (process_idx + 1)}' if process_idx != num_process - 1 else '', '-async 1', out_path, '-y'
51
    ]
52
    print(' '.join(cmd))
53
    subprocess.call(' '.join(cmd), shell=True)
54
    return out_path
55

56

57
class Reader:
58

59
    def __init__(self, args, total_workers=1, worker_idx=0):
60
        self.args = args
61
        input_type = mimetypes.guess_type(args.input)[0]
62
        self.input_type = 'folder' if input_type is None else input_type
63
        self.paths = []  # for image&folder type
64
        self.audio = None
65
        self.input_fps = None
66
        if self.input_type.startswith('video'):
67
            video_path = get_sub_video(args, total_workers, worker_idx)
68
            self.stream_reader = (
69
                ffmpeg.input(video_path).output('pipe:', format='rawvideo', pix_fmt='bgr24',
70
                                                loglevel='error').run_async(
71
                                                    pipe_stdin=True, pipe_stdout=True, cmd=args.ffmpeg_bin))
72
            meta = get_video_meta_info(video_path)
73
            self.width = meta['width']
74
            self.height = meta['height']
75
            self.input_fps = meta['fps']
76
            self.audio = meta['audio']
77
            self.nb_frames = meta['nb_frames']
78

79
        else:
80
            if self.input_type.startswith('image'):
81
                self.paths = [args.input]
82
            else:
83
                paths = sorted(glob.glob(os.path.join(args.input, '*')))
84
                tot_frames = len(paths)
85
                num_frame_per_worker = tot_frames // total_workers + (1 if tot_frames % total_workers else 0)
86
                self.paths = paths[num_frame_per_worker * worker_idx:num_frame_per_worker * (worker_idx + 1)]
87

88
            self.nb_frames = len(self.paths)
89
            assert self.nb_frames > 0, 'empty folder'
90
            from PIL import Image
91
            tmp_img = Image.open(self.paths[0])
92
            self.width, self.height = tmp_img.size
93
        self.idx = 0
94

95
    def get_resolution(self):
96
        return self.height, self.width
97

98
    def get_fps(self):
99
        if self.args.fps is not None:
100
            return self.args.fps
101
        elif self.input_fps is not None:
102
            return self.input_fps
103
        return 24
104

105
    def get_audio(self):
106
        return self.audio
107

108
    def __len__(self):
109
        return self.nb_frames
110

111
    def get_frame_from_stream(self):
112
        img_bytes = self.stream_reader.stdout.read(self.width * self.height * 3)  # 3 bytes for one pixel
113
        if not img_bytes:
114
            return None
115
        img = np.frombuffer(img_bytes, np.uint8).reshape([self.height, self.width, 3])
116
        return img
117

118
    def get_frame_from_list(self):
119
        if self.idx >= self.nb_frames:
120
            return None
121
        img = cv2.imread(self.paths[self.idx])
122
        self.idx += 1
123
        return img
124

125
    def get_frame(self):
126
        if self.input_type.startswith('video'):
127
            return self.get_frame_from_stream()
128
        else:
129
            return self.get_frame_from_list()
130

131
    def close(self):
132
        if self.input_type.startswith('video'):
133
            self.stream_reader.stdin.close()
134
            self.stream_reader.wait()
135

136

137
class Writer:
138

139
    def __init__(self, args, audio, height, width, video_save_path, fps):
140
        out_width, out_height = int(width * args.outscale), int(height * args.outscale)
141
        if out_height > 2160:
142
            print('You are generating video that is larger than 4K, which will be very slow due to IO speed.',
143
                  'We highly recommend to decrease the outscale(aka, -s).')
144

145
        if audio is not None:
146
            self.stream_writer = (
147
                ffmpeg.input('pipe:', format='rawvideo', pix_fmt='bgr24', s=f'{out_width}x{out_height}',
148
                             framerate=fps).output(
149
                                 audio,
150
                                 video_save_path,
151
                                 pix_fmt='yuv420p',
152
                                 vcodec='libx264',
153
                                 loglevel='error',
154
                                 acodec='copy').overwrite_output().run_async(
155
                                     pipe_stdin=True, pipe_stdout=True, cmd=args.ffmpeg_bin))
156
        else:
157
            self.stream_writer = (
158
                ffmpeg.input('pipe:', format='rawvideo', pix_fmt='bgr24', s=f'{out_width}x{out_height}',
159
                             framerate=fps).output(
160
                                 video_save_path, pix_fmt='yuv420p', vcodec='libx264',
161
                                 loglevel='error').overwrite_output().run_async(
162
                                     pipe_stdin=True, pipe_stdout=True, cmd=args.ffmpeg_bin))
163

164
    def write_frame(self, frame):
165
        frame = frame.astype(np.uint8).tobytes()
166
        self.stream_writer.stdin.write(frame)
167

168
    def close(self):
169
        self.stream_writer.stdin.close()
170
        self.stream_writer.wait()
171

172

173
def inference_video(args, video_save_path, device=None, total_workers=1, worker_idx=0):
174
    # ---------------------- determine models according to model names ---------------------- #
175
    args.model_name = args.model_name.split('.pth')[0]
176
    if args.model_name == 'RealESRGAN_x4plus':  # x4 RRDBNet model
177
        model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4)
178
        netscale = 4
179
        file_url = ['https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth']
180
    elif args.model_name == 'RealESRNet_x4plus':  # x4 RRDBNet model
181
        model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4)
182
        netscale = 4
183
        file_url = ['https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.1/RealESRNet_x4plus.pth']
184
    elif args.model_name == 'RealESRGAN_x4plus_anime_6B':  # x4 RRDBNet model with 6 blocks
185
        model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=6, num_grow_ch=32, scale=4)
186
        netscale = 4
187
        file_url = ['https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth']
188
    elif args.model_name == 'RealESRGAN_x2plus':  # x2 RRDBNet model
189
        model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=2)
190
        netscale = 2
191
        file_url = ['https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth']
192
    elif args.model_name == 'realesr-animevideov3':  # x4 VGG-style model (XS size)
193
        model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=4, act_type='prelu')
194
        netscale = 4
195
        file_url = ['https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-animevideov3.pth']
196
    elif args.model_name == 'realesr-general-x4v3':  # x4 VGG-style model (S size)
197
        model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=32, upscale=4, act_type='prelu')
198
        netscale = 4
199
        file_url = [
200
            'https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-general-wdn-x4v3.pth',
201
            'https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-general-x4v3.pth'
202
        ]
203

204
    # ---------------------- determine model paths ---------------------- #
205
    model_path = os.path.join('weights', args.model_name + '.pth')
206
    if not os.path.isfile(model_path):
207
        ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
208
        for url in file_url:
209
            # model_path will be updated
210
            model_path = load_file_from_url(
211
                url=url, model_dir=os.path.join(ROOT_DIR, 'weights'), progress=True, file_name=None)
212

213
    # use dni to control the denoise strength
214
    dni_weight = None
215
    if args.model_name == 'realesr-general-x4v3' and args.denoise_strength != 1:
216
        wdn_model_path = model_path.replace('realesr-general-x4v3', 'realesr-general-wdn-x4v3')
217
        model_path = [model_path, wdn_model_path]
218
        dni_weight = [args.denoise_strength, 1 - args.denoise_strength]
219

220
    # restorer
221
    upsampler = RealESRGANer(
222
        scale=netscale,
223
        model_path=model_path,
224
        dni_weight=dni_weight,
225
        model=model,
226
        tile=args.tile,
227
        tile_pad=args.tile_pad,
228
        pre_pad=args.pre_pad,
229
        half=not args.fp32,
230
        device=device,
231
    )
232

233
    if 'anime' in args.model_name and args.face_enhance:
234
        print('face_enhance is not supported in anime models, we turned this option off for you. '
235
              'if you insist on turning it on, please manually comment the relevant lines of code.')
236
        args.face_enhance = False
237

238
    if args.face_enhance:  # Use GFPGAN for face enhancement
239
        from gfpgan import GFPGANer
240
        face_enhancer = GFPGANer(
241
            model_path='https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth',
242
            upscale=args.outscale,
243
            arch='clean',
244
            channel_multiplier=2,
245
            bg_upsampler=upsampler)  # TODO support custom device
246
    else:
247
        face_enhancer = None
248

249
    reader = Reader(args, total_workers, worker_idx)
250
    audio = reader.get_audio()
251
    height, width = reader.get_resolution()
252
    fps = reader.get_fps()
253
    writer = Writer(args, audio, height, width, video_save_path, fps)
254

255
    pbar = tqdm(total=len(reader), unit='frame', desc='inference')
256
    while True:
257
        img = reader.get_frame()
258
        if img is None:
259
            break
260

261
        try:
262
            if args.face_enhance:
263
                _, _, output = face_enhancer.enhance(img, has_aligned=False, only_center_face=False, paste_back=True)
264
            else:
265
                output, _ = upsampler.enhance(img, outscale=args.outscale)
266
        except RuntimeError as error:
267
            print('Error', error)
268
            print('If you encounter CUDA out of memory, try to set --tile with a smaller number.')
269
        else:
270
            writer.write_frame(output)
271

272
        torch.cuda.synchronize(device)
273
        pbar.update(1)
274

275
    reader.close()
276
    writer.close()
277

278

279
def run(args):
280
    args.video_name = osp.splitext(os.path.basename(args.input))[0]
281
    video_save_path = osp.join(args.output, f'{args.video_name}_{args.suffix}.mp4')
282

283
    if args.extract_frame_first:
284
        tmp_frames_folder = osp.join(args.output, f'{args.video_name}_inp_tmp_frames')
285
        os.makedirs(tmp_frames_folder, exist_ok=True)
286
        os.system(f'ffmpeg -i {args.input} -qscale:v 1 -qmin 1 -qmax 1 -vsync 0  {tmp_frames_folder}/frame%08d.png')
287
        args.input = tmp_frames_folder
288

289
    num_gpus = torch.cuda.device_count()
290
    num_process = num_gpus * args.num_process_per_gpu
291
    if num_process == 1:
292
        inference_video(args, video_save_path)
293
        return
294

295
    ctx = torch.multiprocessing.get_context('spawn')
296
    pool = ctx.Pool(num_process)
297
    os.makedirs(osp.join(args.output, f'{args.video_name}_out_tmp_videos'), exist_ok=True)
298
    pbar = tqdm(total=num_process, unit='sub_video', desc='inference')
299
    for i in range(num_process):
300
        sub_video_save_path = osp.join(args.output, f'{args.video_name}_out_tmp_videos', f'{i:03d}.mp4')
301
        pool.apply_async(
302
            inference_video,
303
            args=(args, sub_video_save_path, torch.device(i % num_gpus), num_process, i),
304
            callback=lambda arg: pbar.update(1))
305
    pool.close()
306
    pool.join()
307

308
    # combine sub videos
309
    # prepare vidlist.txt
310
    with open(f'{args.output}/{args.video_name}_vidlist.txt', 'w') as f:
311
        for i in range(num_process):
312
            f.write(f'file \'{args.video_name}_out_tmp_videos/{i:03d}.mp4\'\n')
313

314
    cmd = [
315
        args.ffmpeg_bin, '-f', 'concat', '-safe', '0', '-i', f'{args.output}/{args.video_name}_vidlist.txt', '-c',
316
        'copy', f'{video_save_path}'
317
    ]
318
    print(' '.join(cmd))
319
    subprocess.call(cmd)
320
    shutil.rmtree(osp.join(args.output, f'{args.video_name}_out_tmp_videos'))
321
    if osp.exists(osp.join(args.output, f'{args.video_name}_inp_tmp_videos')):
322
        shutil.rmtree(osp.join(args.output, f'{args.video_name}_inp_tmp_videos'))
323
    os.remove(f'{args.output}/{args.video_name}_vidlist.txt')
324

325

326
def main():
327
    """Inference demo for Real-ESRGAN.
328
    It mainly for restoring anime videos.
329

330
    """
331
    parser = argparse.ArgumentParser()
332
    parser.add_argument('-i', '--input', type=str, default='inputs', help='Input video, image or folder')
333
    parser.add_argument(
334
        '-n',
335
        '--model_name',
336
        type=str,
337
        default='realesr-animevideov3',
338
        help=('Model names: realesr-animevideov3 | RealESRGAN_x4plus_anime_6B | RealESRGAN_x4plus | RealESRNet_x4plus |'
339
              ' RealESRGAN_x2plus | realesr-general-x4v3'
340
              'Default:realesr-animevideov3'))
341
    parser.add_argument('-o', '--output', type=str, default='results', help='Output folder')
342
    parser.add_argument(
343
        '-dn',
344
        '--denoise_strength',
345
        type=float,
346
        default=0.5,
347
        help=('Denoise strength. 0 for weak denoise (keep noise), 1 for strong denoise ability. '
348
              'Only used for the realesr-general-x4v3 model'))
349
    parser.add_argument('-s', '--outscale', type=float, default=4, help='The final upsampling scale of the image')
350
    parser.add_argument('--suffix', type=str, default='out', help='Suffix of the restored video')
351
    parser.add_argument('-t', '--tile', type=int, default=0, help='Tile size, 0 for no tile during testing')
352
    parser.add_argument('--tile_pad', type=int, default=10, help='Tile padding')
353
    parser.add_argument('--pre_pad', type=int, default=0, help='Pre padding size at each border')
354
    parser.add_argument('--face_enhance', action='store_true', help='Use GFPGAN to enhance face')
355
    parser.add_argument(
356
        '--fp32', action='store_true', help='Use fp32 precision during inference. Default: fp16 (half precision).')
357
    parser.add_argument('--fps', type=float, default=None, help='FPS of the output video')
358
    parser.add_argument('--ffmpeg_bin', type=str, default='ffmpeg', help='The path to ffmpeg')
359
    parser.add_argument('--extract_frame_first', action='store_true')
360
    parser.add_argument('--num_process_per_gpu', type=int, default=1)
361

362
    parser.add_argument(
363
        '--alpha_upsampler',
364
        type=str,
365
        default='realesrgan',
366
        help='The upsampler for the alpha channels. Options: realesrgan | bicubic')
367
    parser.add_argument(
368
        '--ext',
369
        type=str,
370
        default='auto',
371
        help='Image extension. Options: auto | jpg | png, auto means using the same extension as inputs')
372
    args = parser.parse_args()
373

374
    args.input = args.input.rstrip('/').rstrip('\\')
375
    os.makedirs(args.output, exist_ok=True)
376

377
    if mimetypes.guess_type(args.input)[0] is not None and mimetypes.guess_type(args.input)[0].startswith('video'):
378
        is_video = True
379
    else:
380
        is_video = False
381

382
    if is_video and args.input.endswith('.flv'):
383
        mp4_path = args.input.replace('.flv', '.mp4')
384
        os.system(f'ffmpeg -i {args.input} -codec copy {mp4_path}')
385
        args.input = mp4_path
386

387
    if args.extract_frame_first and not is_video:
388
        args.extract_frame_first = False
389

390
    run(args)
391

392
    if args.extract_frame_first:
393
        tmp_frames_folder = osp.join(args.output, f'{args.video_name}_inp_tmp_frames')
394
        shutil.rmtree(tmp_frames_folder)
395

396

397
if __name__ == '__main__':
398
    main()
399