{"id":17,"date":"2025-12-10T17:33:17","date_gmt":"2025-12-10T08:33:17","guid":{"rendered":"https:\/\/www.em99.tech\/?p=17"},"modified":"2025-12-11T16:06:03","modified_gmt":"2025-12-11T07:06:03","slug":"gpusoroban%e3%81%95%e3%82%93%e3%82%92%e4%bd%bf%e3%81%a3%e3%81%a6%e3%81%bf%e3%81%9f2denoiser%e7%92%b0%e5%a2%83%e6%a7%8b%e7%af%89","status":"publish","type":"post","link":"https:\/\/www.em99.tech\/?p=17","title":{"rendered":"GPUSOROBAN\u3055\u3093\u3092\u4f7f\u3063\u3066\u307f\u305f(2)Denoiser\u74b0\u5883\u69cb\u7bc9"},"content":{"rendered":"\n

\u672c\u8a18\u4e8b\u306f\u3001
\u524d\u56de\u69cb\u7bc9\u3057\u305f GPU \u5bfe\u5fdc PyTorch \u74b0\u5883\u3092\u524d\u63d0\u306b\u3001Denoiser \u306e GPU \u52d5\u4f5c\u691c\u8a3c\u3068\u52d5\u753b\u97f3\u58f0\u306e\u30ce\u30a4\u30ba\u9664\u53bb\u30c6\u30b9\u30c8\u3092\u884c\u3063\u305f\u8a18\u9332<\/strong> \u3067\u3059\u3002<\/p>\n\n\n\n

\u52d5\u753b\u3092\u64ae\u308b\u306e\u3082\u30b9\u30de\u30db\u306e\u30ab\u30e1\u30e9\u306e\u6a5f\u80fd\u3092\u5229\u7528\u3059\u308b\u7a0b\u5ea6\u3067\u3001\u64ae\u3063\u305f\u52d5\u753b\u3082\u30a4\u30f3\u30b9\u30bf\u30b0\u30e9\u30e0\u306e\u30b9\u30c8\u30fc\u30ea\u30fc\u30ba\u3078\u305d\u306e\u307e\u307e\u4e0a\u3052\u308b\u304f\u3089\u3044\u3057\u304b\u3057\u3066\u3044\u306a\u304b\u3063\u305f\u7d20\u4eba\u306a\u79c1\u3067\u3059\u304c\u3001
\u300c\u6700\u8fd1\u306fAI\u3067\u30ce\u30a4\u30ba\u9664\u53bb\u3068\u304b\u7c21\u5358\u306b\u3067\u304d\u306a\u308b\u3093\u3058\u3083\u306a\u3044\u3067\u3059\u304b\u306d\uff1f\u300d
\u3068\u76f8\u8ac7\u3092\u3044\u305f\u3060\u304d\u307e\u3057\u305f\u3002
\u4e16\u306e\u4e2d\u7c21\u5358\u306a\u3082\u306e\u3063\u3066\u7121\u3044\u3067\u3059\u3088…\u3068\u3082\u8a00\u3048\u305a\u3001\u4e7e\u3044\u305f\u7b11\u3044\u3067\u7b54\u3048\u3066\u304a\u3044\u305f\u306e\u3067\u3059\u304c\u3001\u6700\u8fd1\u8208\u5473\u306e\u3042\u308b\u5206\u91ce\u3067\u3082\u3042\u308b\u306e\u3067\u8a66\u3057\u3066\u307f\u308b\u3053\u3068\u306b\u3057\u307e\u3057\u305f\u3002<\/p>\n\n\n\n

Facebook\uff08Meta \/ FAIR\uff09\u304cGitHub\u306b\u516c\u958b\u3057\u3066\u3044\u308bDenoiser\u3067\u3059\u304c\u3001\u5c11\u3057\u6642\u9593\u306e\u7d4c\u3063\u305f\u6280\u8853\u306e\u3088\u3046\u3067\u3059\u3002(2020\u5e74\u9803\uff1f)
\u751f\u6210AI\u306b\u76f8\u8ac7\u3057\u305f\u3089\u5019\u88dc\u306b\u6319\u304c\u3063\u305f\u306e\u3067\u3001\u3068\u304f\u306b\u6df1\u304f\u8003\u3048\u305a\u306b\u9078\u629e\u3057\u3066\u307f\u307e\u3057\u305f\u304c\u3001\u524d\u56de<\/a>\u66f8\u3044\u305f\u3088\u3046\u306a\u74b0\u5883\u3067\u3059\u306e\u3067\u3001\u7d50\u679c\u826f\u304b\u3063\u305f\u3088\u3046\u3067\u3059\u3002
Hybrid Demucs\u3084Audiocraft\u306b\u3064\u3044\u3066\u3082\u3001\u4eca\u56de\u306e\u74b0\u5883\u3067GPU\u3092\u4f7f\u7528\u3057\u3066\u52d5\u4f5c\u3059\u308b\u304b\u306e\u78ba\u8a8d\u3082\u3044\u305a\u308c\u3057\u3066\u307f\u3088\u3046\u3068\u601d\u3044\u307e\u3059\u304c\u3001\u307e\u305a\u306f\u201dAI\u3092\u4f7f\u3063\u3066\u30ce\u30a4\u30ba\u9664\u53bb\u201d\u304c\u51fa\u6765\u308b\u304b\u3001\u305d\u306e\u969b\u306bGPU\u3092\u4f7f\u7528\u3067\u304d\u308b\u304b\u3001Denoiser(AI)\u3092\u4f7f\u3063\u3066\u78ba\u8a8d\u3059\u308b\u3053\u3068\u306b\u3057\u307e\u3059\u3002

1.\u4eca\u56de\u306e\u30b5\u30f3\u30d7\u30eb\u52d5\u753b\u306e\u8ab2\u984c

\u3000\u30fb\u30d6\u30ec\u30b9\u30ce\u30a4\u30ba\u304c\u6fc0\u3057\u3044
\u3000\u30fb\u97f3\u58f0\u306e\u5de6\u53f3\u306e\u30d0\u30e9\u30f3\u30b9\u304c\u60aa\u3044

\u4eca\u56de\u5165\u624b\u3057\u305f\u30b5\u30f3\u30d7\u30eb\u52d5\u753b\u306e\u97f3\u58f0\u306f\u3053\u306e2\u70b9\u3092\u4f55\u3068\u304b\u3057\u305f\u3044\u3001\u3068\u3044\u3046\u3082\u306e\u3067\u3057\u305f\u3002
\u30c6\u30b9\u30c8\u7684\u306b\u5b9f\u65bd\u3059\u308b\u3082\u306e\u306a\u306e\u3067\u3001\u3053\u306e\u3046\u3061\u300c\u97f3\u58f0\u306e\u5de6\u53f3\u306e\u30d0\u30e9\u30f3\u30b9\u304c\u60aa\u3044\u300d\u306f\u5358\u7d14\u306b\u30b9\u30c6\u30ec\u30aa\u3092\u30e2\u30ce\u30e9\u30eb\u306b\u3057\u3066\u89e3\u6c7a\u3059\u308b\u3053\u3068\u306b\u3002
\u3042\u304f\u307e\u3067\u547c\u5438\u97f3\u3092\u542b\u3081\u305f\u30ce\u30a4\u30ba\u4f4e\u6e1b\u3068\u8a71\u8005\u306e\u58f0\u306e\u660e\u77ad\u5ea6\u30a2\u30c3\u30d7\u3001\u5b89\u5b9a\u5316\u3092\u76ee\u6a19\u306b\u3059\u308b\u3053\u3068\u3057\u307e\u3057\u305f\u3002<\/p>\n\n\n\n

2.\u3084\u308a\u305f\u3044\u3053\u3068\u306e\u6d41\u308c\u3068\u62c5\u5f53\u90e8\u5206

\u52d5\u753b\u30d5\u30a1\u30a4\u30eb\u304b\u3089\u97f3\u58f0\u53d6\u308a\u51fa\u3057(ffmpeg)
\u2193
\u30b9\u30c6\u30ec\u30aa\u304b\u3089\u30e2\u30ce\u30e9\u30eb\u306b\u5909\u63db(ffmpeg)
\u2193
\u30ce\u30a4\u30ba\u4f4e\u6e1b(Denoiser)
\u2193
\u58f0\u306e\u660e\u77ad\u5ea6\u3084\u97f3\u91cf\u30c0\u30a4\u30ca\u30df\u30af\u30b9\u306e\u5b89\u5b9a\u5316(Denoiser)
\u2193
\u30ce\u30a4\u30ba\u9664\u53bb\u5f8c\u306e\u97f3\u58f0\u3092\u52d5\u753b\u3078\u518d\u7d50\u5408(ffmpeg)
\u2193
AAC \u3078\u306e\u30a8\u30f3\u30b3\u30fc\u30c9\uff08\u97f3\u58f0\u3092\u52d5\u753b\u7528\u30d5\u30a9\u30fc\u30de\u30c3\u30c8\u3078\uff09(ffmpeg)

Denoiser\u90e8\u5206\u304cAI\u3068GPU\u5229\u7528\u90e8\u5206\u3001ffmpeg\u90e8\u5206\u306f\u5b9f\u884c\u6642\u306e\u8a2d\u5b9a\u5024\u3067\u8abf\u6574\u3067\u4f7f\u3046\u306e\u306fCPU\u3067\u3059\u3002<\/p>\n\n\n\n

3.\u74b0\u5883\u69cb\u7bc9

Pytorch\u3001CUDA\u3092\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3057\u3066GPU\u3092\u8a8d\u8b58\u3059\u308b\u3068\u3053\u308d\u307e\u3067\u306f\u524d\u56de<\/a>\u306e\u8a18\u4e8b\u3092\u53c2\u7167\u304f\u3060\u3055\u3044\u3002

\u30fbffmpeg\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb<\/p>\n\n\n\n

$ sudo apt update
$ sudo apt install -y ffmpeg<\/p>\n\n\n\n

\u5fc5\u8981\u306a\u97f3\u58f0\u30d5\u30a3\u30eb\u30bf\u304c\u5165\u3063\u3066\u3044\u308b\u304b\u78ba\u8a8d\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

#\u30ce\u30a4\u30ba\u95a2\u9023
$ ffmpeg -filters | grep -E ‘afftdn|anlmdn’

#\u30a4\u30b3\u30e9\u30a4\u30b6\u30fb\u5e2f\u57df\u7cfb
$ ffmpeg -filters | grep -E ‘equalizer|anequalizer|highpass|lowpass’

#\u30e9\u30a6\u30c9\u30cd\u30b9\u6b63\u898f\u5316
ffmpeg -filters | grep loudnorm<\/p>\n\n\n\n

\u3053\u3053\u3067

afftdn
anlmdn
equalizer
highpass
lowpass
loudnorm

\u3042\u305f\u308a\u304c\u8868\u793a\u3055\u308c\u3066\u3044\u308c\u3070\u5927\u4e08\u592b\u306a\u3088\u3046\u3067\u3059\u3002

\u30fbDenoiser\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb<\/p>\n\n\n\n

conda\u306e\u4eee\u60f3\u74b0\u5883\u306b\u5165\u3063\u3066\u3044\u306a\u3044\u5834\u5408\u306f\u5165\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

$ conda activate denoiserenv<\/p>\n\n\n\n

$ cd ~
$ mkdir -p denoiser_test
$ cd denoiser_test<\/p>\n\n\n\n

$ git clone https:\/\/github.com\/facebookresearch\/denoiser.git
$ cd denoiser<\/p>\n\n\n\n

#\u4f9d\u5b58\u95a2\u4fc2\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb
$ pip install -r requirements.txt
$ pip install -e .<\/p>\n\n\n\n

Denoiser \u304c GPU \u3092\u4f7f\u3063\u3066\u304f\u308c\u308b\u304b\u30c6\u30b9\u30c8<\/p>\n\n\n\n

$ python\n\n>>>import torch\nfrom denoiser import pretrained\n\nprint(\"torch.cuda.is_available():\", torch.cuda.is_available())\n\nmodel = pretrained.dns64()\nprint(\"loaded model device:\", next(model.parameters()).device)\n\nif torch.cuda.is_available():\n model = model.to(\"cuda\")\n print(\"after .to('cuda'):\", next(model.parameters()).device)\nelse:\n print(\"CUDA not available, staying on CPU.\")<\/code><\/pre>\n\n\n\n
\u3053\u3053\u3067

torch.cuda.is_available(): True

after .to(‘cuda’): cuda:0

\u3068\u51fa\u308c\u3070\u3001\u30e2\u30c7\u30eb\u3092 GPU \u306b\u8f09\u305b\u308b\u6e96\u5099\u304c\u5b8c\u4e86\u3067\u3059\u3002

\u30d5\u30a9\u30eb\u30c0\u69cb\u6210\u3092\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u3057\u307e\u3057\u305f\u3002

~\/denoiser_test
     |-denoised_wavs\/
     |-noisy_wavs\/
         |-work\/
<\/p>\n\n\n\n
4.\u5b9f\u969b\u306b\u52d5\u304b\u3057\u3066\u307f\u308b<\/p>\n\n\n\n
~\/denoiser_test\/noisy_wavs\u3000\u3078mp4\u30d5\u30a1\u30a4\u30eb\u3092\u683c\u7d0d\u3057\u3066\u304f\u3060\u3055\u3044\u3002
\u79c1\u306fWinSCP\u3092\u4f7f\u7528\u3057\u307e\u3057\u305f\u3002WinSCP\u3067\u306e\u63a5\u7d9a\u65b9\u6cd5\u3082GPUSORBAN\u3055\u3093\u306e

WinSCP\u3067Winodws PC\u304b\u3089\u30d5\u30a1\u30a4\u30eb\u8ee2\u9001\u3059\u308b\u65b9\u6cd5<\/a>

\u3067\u8a73\u3057\u304f\u8aac\u660e\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n
\u30fbmp4 \u304b\u3089\u97f3\u58f0\u3060\u3051\u629c\u304d\u51fa\u3057\uff08wav\uff09<\/p>\n\n\n\n
\u4eca\u56de\u4f7f\u7528\u3057\u305fmp4\u30d5\u30a1\u30a4\u30eb\u306e\u30d5\u30a1\u30a4\u30eb\u540d\uff1alecture.mp4<\/p>\n\n\n\n
$ cd ~\/denoiser_test\/noisy_wavs\n\n$ ffmpeg -i lecture.mp4 \\\n-vn -acodec pcm_s16le -ar 48000 -ac 2 \\\nwork\/lecture_raw.wav<\/code><\/pre>\n\n\n\n
-vn \u2026 \u30d3\u30c7\u30aa\u7121\u52b9
pcm_s16le \u2026 \u7121\u5727\u7e2e\uff08\u51e6\u7406\u3057\u3084\u3059\u3044\uff09
-ar 48000 \/ -ac 2 \u2026 48kHz \/ \u30b9\u30c6\u30ec\u30aa\u306b\u7d71\u4e00<\/p>\n\n\n\n
\u30fb\u30b9\u30c6\u30ec\u30aa\u3092\u30e2\u30ce\u30e9\u30eb\u306b\u5909\u63db<\/p>\n\n\n\n
$ cd work\n\n$ ffmpeg -i lecture_raw.wav \\\n -af \"pan=mono|c0=0.5c0+0.5c1\" \\\n lecture_raw_mono.wav<\/code><\/pre>\n\n\n\n
\u30fb\u30ce\u30a4\u30ba\u4f4e\u6e1b\u51e6\u7406<\/p>\n\n\n\n
$ cd ~\/denoiser_test\/denoiser\n\n$ python -m denoiser.enhance \\\n --dns48 \\\n --device cuda \\\n --streaming \\\n --batch_size 1 \\\n --num_workers 1 \\\n --noisy_dir ..\/noisy_wavs\/work \\\n --out_dir ..\/denoised_wavs<\/code><\/pre>\n\n\n\n
\u30d0\u30c3\u30af\u30b9\u30e9\u30c3\u30b7\u30e5\u306e\u6b21\u306b\u306f\u534a\u89d2\u30b9\u30da\u30fc\u30b9\u304c\u5fc5\u8981\u3067\u3059\u3002\u30b9\u30da\u30fc\u30b9\u306a\u3051\u308c\u3070\u52d5\u4f5c\u3057\u306a\u3051\u308c\u3070\u3044\u3044\u306e\u306b\u3001CPU\u3092\u4f7f\u3063\u3066\u8d70\u308a\u51fa\u3059\u3068\u3044\u3046…
\u5225\u306e\u30bf\u30fc\u30df\u30ca\u30eb\u3092\u8d77\u52d5\u3057\u3066\u4ee5\u4e0b\u306e\u30b3\u30de\u30f3\u30c9\u3092\u5b9f\u884c\u3059\u308b\u3068GPU\u304c\u52d5\u4f5c\u3057\u3066\u3044\u308b\u304b\u78ba\u8a8d\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
$ watch -n 1 nvidia-smi<\/p>\n\n\n\n
\u4fdd\u5b58\u306fout_dir(\u4eca\u56de\u3060\u3068~\/denoiser_test\/denoised_wavs)\u306b\u3055\u308c\u307e\u3059\u3002
(\u5143\u306e\u30d5\u30a1\u30a4\u30eb\u540d)_enhanced.wav\u30d5\u30a1\u30a4\u30eb\u304c\u305d\u308c\u3067\u3059\u3002
\u540c\u3058\u3068\u3053\u308d\u306b(\u5143\u306e\u30d5\u30a1\u30a4\u30eb\u540d)_noisy.wav\u30d5\u30a1\u30a4\u30eb\u3082\u751f\u6210\u3055\u308c\u307e\u3059\u304c\u3001\u3053\u308c\u306f\u4ed5\u69d8\u3060\u305d\u3046\u3067\u3059\u3002
\u7406\u7531\u306f\u751f\u6210AI\u306b\u805e\u3044\u3066\u307f\u3066\u304f\u3060\u3055\u3044\u306d\u3002<\/p>\n\n\n\n
\u30fb\u97f3\u58f0\u3092\u6620\u50cf\u306b\u623b\u3059<\/p>\n\n\n\n
\u623b\u3059\u3068\u3044\u3046\u304b\u3001\u52d5\u753b\u30d5\u30a1\u30a4\u30eb\u3092\u30b9\u30c8\u30ea\u30fc\u30df\u30f3\u30b0\u3057\u3066\u65b0\u3057\u3044\u97f3\u58f0\u3068\u5408\u6210\u3059\u308b\u3001\u3068\u3044\u3046\u30a4\u30e1\u30fc\u30b8\u3067\u3059\u306d\u3002<\/p>\n\n\n\n
$ ffmpeg -i ~\/denoiser_test\/noisy_wavs\/lecture.mp4 \\\n -i ~\/denoiser_test\/denoised_wavs\/lecture_raw_mono_enhanced.wav \\\n -map 0:v -map 1:a \\\n -c:v copy \\\n -c:a aac -b:a 192k \\\n ~\/denoiser_test\/lecture_denoised.mp4<\/code><\/pre>\n\n\n\n
\u4eca\u56de\u306e\u30c6\u30b9\u30c8\u306f\u3053\u3053\u307e\u3067\u3067\u3059\u3002
\u51fa\u6765\u4e0a\u304c\u3063\u305f\u30d5\u30a1\u30a4\u30eb\u306e\u97f3\u58f0\u306f\u5c11\u3057\u306f\u805e\u304d\u3084\u3059\u304f\u306a\u3063\u305f\u6c17\u304c\u3057\u307e\u3059\u3002
\u30ce\u30a4\u30ba\u304c\u5927\u304d\u304b\u3063\u305f\u3068\u3053\u308d\u3092\u88dc\u6b63\u3057\u3066\u3044\u308b\u90e8\u5206\u3067\u5225\u306e\u30ce\u30a4\u30ba\u304c\u767a\u751f\u3057\u3066\u3044\u308b\u306e\u306f\u30ce\u30a4\u30ba\u88dc\u6b63\u306e\u305b\u3044\u304b\u3082\u3057\u308c\u307e\u305b\u3093\u306d\u3002\u3082\u3046\u4e00\u5ea6\u3053\u306e\u30d5\u30ed\u30fc\u3092\u6d41\u3057\u305f\u3089\u3069\u3046\u306a\u308b\u306e\u304b\u306a\uff1f

\u3055\u3066\u6b21\u306fHybrid Demucs\u304c\u52d5\u4f5c\u3059\u308b\u74b0\u5883\u69cb\u7bc9\u304c\u3067\u304d\u308b\u304b\u3092\u8a66\u3059\u304b\u3001\u6559\u5e2b\u30c7\u30fc\u30bf\u3092\u5927\u91cf\u306b\u7528\u610f\u3057\u3066\u5225\u306e\u30b7\u30b9\u30c6\u30e0\u3092\u4f5c\u308b\u304b…\u3069\u3046\u3057\u307e\u3057\u3087\u3046\u306d\u7b11<\/p>\n","protected":false},"excerpt":{"rendered":"
\u672c\u8a18\u4e8b\u306f\u3001\u524d\u56de\u69cb\u7bc9\u3057\u305f GPU \u5bfe\u5fdc PyTorch \u74b0\u5883\u3092\u524d\u63d0\u306b\u3001Denoiser \u306e GPU \u52d5\u4f5c\u691c\u8a3c\u3068\u52d5\u753b\u97f3\u58f0\u306e\u30ce\u30a4\u30ba\u9664\u53bb\u30c6\u30b9\u30c8\u3092\u884c\u3063\u305f\u8a18\u9332 \u3067\u3059\u3002 \u52d5\u753b\u3092\u64ae\u308b\u306e\u3082\u30b9\u30de\u30db\u306e\u30ab\u30e1\u30e9\u306e\u6a5f\u80fd\u3092\u5229\u7528\u3059\u308b\u7a0b\u5ea6\u3067\u3001\u64ae\u3063\u305f\u52d5\u753b\u3082 […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[],"class_list":["post-17","post","type-post","status-publish","format-standard","hentry","category-tech"],"_links":{"self":[{"href":"https:\/\/www.em99.tech\/index.php?rest_route=\/wp\/v2\/posts\/17","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.em99.tech\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.em99.tech\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.em99.tech\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.em99.tech\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=17"}],"version-history":[{"count":24,"href":"https:\/\/www.em99.tech\/index.php?rest_route=\/wp\/v2\/posts\/17\/revisions"}],"predecessor-version":[{"id":46,"href":"https:\/\/www.em99.tech\/index.php?rest_route=\/wp\/v2\/posts\/17\/revisions\/46"}],"wp:attachment":[{"href":"https:\/\/www.em99.tech\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=17"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.em99.tech\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=17"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.em99.tech\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=17"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}