{"id":36,"date":"2025-12-11T14:50:44","date_gmt":"2025-12-11T05:50:44","guid":{"rendered":"https:\/\/www.em99.tech\/?p=36"},"modified":"2025-12-11T14:50:44","modified_gmt":"2025-12-11T05:50:44","slug":"gpusoraban%e3%81%95%e3%82%93%e3%82%92%e4%bd%bf%e3%81%a3%e3%81%a6%e3%81%bf%e3%81%9f3deepfilternet3%e3%82%82%e8%a9%a6%e3%81%97%e3%81%be%e3%81%97%e3%81%9f%e3%80%82","status":"publish","type":"post","link":"https:\/\/www.em99.tech\/?p=36","title":{"rendered":"GPUSORABAN\u3055\u3093\u3092\u4f7f\u3063\u3066\u307f\u305f(3)DeepFilterNet3\u3082\u8a66\u3057\u307e\u3057\u305f\u3002"},"content":{"rendered":"\n

\u672c\u8a18\u4e8b\u306f\u3001
\u3053\u308c\u307e\u3067\u306b\u69cb\u7bc9\u3057\u305f GPU \u5bfe\u5fdc PyTorch \u74b0\u5883\u3092\u524d\u63d0\u306b\u3001DeepFilterNet3 \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>\u00a0\u3067\u3059\u3002<\/p>\n\n\n\n

\u524d\u56de\u306e\u78ba\u8a8d\u3067\u3001Denoiser\u304cGPU\u3092\u4f7f\u7528\u3057\u3066\u52d5\u4f5c\u3059\u308b\u74b0\u5883\u69cb\u7bc9\u3068\u5b9f\u969b\u306b\u30ce\u30a4\u30ba\u9664\u53bb\u3092\u8a66\u3059\u3068\u3053\u308d\u307e\u3067\u884c\u3044\u307e\u3057\u305f\u3002ffmpeg\u3067\u30e2\u30ce\u30e9\u30eb\u306b\u5909\u63db\u3057\u305f\u52b9\u679c\u3082\u3042\u3063\u3066\u304b\u3001Denoiser\u3067\u306e\u30ce\u30a4\u30ba\u9664\u53bb\u3068\u97f3\u91cf\u30c0\u30a4\u30ca\u30df\u30af\u30b9\u306e\u5b89\u5b9a\u5316\u306b\u306f\u305d\u308c\u306a\u308a\u306b\u52b9\u679c\u304c\u3042\u308b\u3053\u3068\u304c\u5206\u304b\u308a\u307e\u3057\u305f\u3002(\u672c\u5f53\u306f\u89e3\u6790\u3057\u3066\u6570\u5024\u3067\u8a18\u8f09\u3059\u308b\u307b\u3046\u304c\u5206\u304b\u308a\u6613\u3044\u306e\u3067\u3059\u304c\u3001\u73fe\u72b6\u306f\u79c1\u306e\u500b\u4eba\u7684\u306a\u611f\u60f3\u3067\u3059\u3002)

\u3068\u3053\u308d\u3067\u88dc\u6b63\u5f8c\u306e\u97f3\u58f0\u306b\u3082\u307e\u3060\u30de\u30a4\u30af\u306b\u4f55\u304b\u304c\u3076\u3064\u304b\u308b\u3088\u3046\u306a\u97f3\u304c\u3042\u308a\u307e\u3057\u305f\u3002
\u81ea\u52d5\u3067\u3069\u3053\u307e\u3067\u3067\u304d\u308b\u304b\u3001\u3082\u78ba\u8a8d\u3057\u305f\u3044\u306e\u3067\u3001\u885d\u6483\u97f3\u306b\u5f37\u3044\u3089\u3057\u3044DeepFilterNet\u3092\u8a66\u3059\u3053\u3068\u306b\u3057\u307e\u3057\u305f\u3002

\u8a18\u4e8b\u5185\u306e\u30c7\u30a3\u30ec\u30af\u30c8\u30ea\u306e\u4f4d\u7f6e\u3084\u540d\u524d\u3001\u30d5\u30a1\u30a4\u30eb\u540d\u306a\u3069\u306f\u79c1\u304c\u81ea\u5206\u3067\u5229\u7528\u3057\u3084\u3059\u3044\u3088\u3046\u306b\u5909\u66f4\u3057\u305f\u308a\u3057\u3066\u3044\u307e\u3059\u306e\u3067\u3001\u8a66\u3055\u308c\u308b\u5834\u5408\u306f\u3054\u81ea\u8eab\u306e\u74b0\u5883\u306b\u5408\u308f\u305b\u3066\u8aad\u307f\u66ff\u3048\u3066\u304f\u3060\u3055\u3044\u3002

1.DeepFilterNet3\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb<\/p>\n\n\n\n

\u4eca\u56de\u3082\u524d\u56de\u306e\u4eee\u60f3\u74b0\u5883\u3092\u5229\u7528\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

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

\u30fbdeepfilternet \u672c\u4f53\u3092\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb<\/p>\n\n\n\n

$ pip install deepfilternet<\/p>\n\n\n\n

\u30fbpy\u30d5\u30a1\u30a4\u30eb\u683c\u7d0d\u7528\u306b\u30c7\u30a3\u30ec\u30af\u30c8\u30ea\u3092\u4f5c\u6210\u3057\u307e\u3057\u305f\u3002<\/p>\n\n\n\n

$ mkdir ~\/scripts

$ cd ~\/scripts<\/p>\n\n\n\n

\u30fb\u30b3\u30fc\u30c9\u3092\u66f8\u3044\u3066.py\u30d5\u30a1\u30a4\u30eb\u306b\u4fdd\u5b58(\u3053\u3053\u3067\u306f\u30d5\u30a1\u30a4\u30eb\u540d\u3092test_dfn3.py\u306b\u3057\u307e\u3057\u305f\u3002)<\/p>\n\n\n\n

import torch\nfrom df.enhance import enhance, init_df, load_audio, save_audio\nimport os\n\n# \u5165\u529b\u30d5\u30a1\u30a4\u30eb\uff08\u4eca\u56de\u4f7f\u3044\u305f\u3044\u30d5\u30a1\u30a4\u30eb\uff09\nINPUT_WAV = os.path.expanduser(\n    \"~\/denoiser_test\/denoised_wavs\/lecture_raw_enhanced.wav\"\n)\n\n# \u51fa\u529b\u30d5\u30a1\u30a4\u30eb\uff08DFN3\u901a\u3057\u305f\u7d50\u679c\uff09\nOUTPUT_WAV = os.path.expanduser(\n    \"~\/denoiser_test\/denoised_wavs\/lecture_raw_enhanced_dfn3.wav\"\n)\n\n# 1\u30c1\u30e3\u30f3\u30af\u306e\u9577\u3055\uff08\u79d2\uff09\nCHUNK_SECONDS = 600  # 10\u5206\u3002\u5fc3\u914d\u306a\u3089 300(5\u5206) \u306b\u3057\u3066\u3082OK\n\n\ndef main():\n    print(\"PyTorch version:\", torch.__version__)\n    print(\"CUDA available:\", torch.cuda.is_available())\n\n    # DeepFilterNet \u521d\u671f\u5316\uff08\u6a19\u6e96\u3067 DeepFilterNet3\uff09\n    model, df_state, _ = init_df()\n    print(\"Model sample rate:\", df_state.sr())\n    print(\"Model device before .to():\", next(model.parameters()).device)\n\n    # GPU\u3078\n    if torch.cuda.is_available():\n        model = model.to(\"cuda\")\n        print(\"Model device after .to():\", next(model.parameters()).device)\n    else:\n        print(\"CUDA not available \u2192 CPU\u52d5\u4f5c\u306b\u306a\u308a\u307e\u3059\")\n\n    # \u97f3\u58f0\u8aad\u307f\u8fbc\u307f\uff08\u30e2\u30c7\u30eb\u306e\u30b5\u30f3\u30d7\u30ea\u30f3\u30b0\u30ec\u30fc\u30c8\u306b\u5408\u308f\u305b\u3066\u30ed\u30fc\u30c9\uff09\n    print(\"Loading:\", INPUT_WAV)\n    audio, sr = load_audio(INPUT_WAV, sr=df_state.sr())\n    # audio: [channels, samples]\n    print(\"Input sr:\", sr, \"Shape:\", audio.shape)\n\n    # \u30e2\u30ce\u30e9\u30eb\u524d\u63d0\u306a\u306e\u3067 1ch \u306b\u3057\u3066\u6271\u3046\n    audio = audio.squeeze(0)  # [samples]\n    total_samples = audio.shape[0]\n    chunk_samples = CHUNK_SECONDS * df_state.sr()\n\n    print(f\"Total samples: {total_samples}\")\n    print(f\"Chunk samples: {chunk_samples} ({CHUNK_SECONDS} sec)\")\n\n    enhanced_chunks = []\n\n    start = 0\n    chunk_idx = 0\n\n    with torch.no_grad():\n        while start < total_samples:\n            end = min(start + chunk_samples, total_samples)\n            chunk_idx += 1\n            print(f\"Processing chunk {chunk_idx}: samples {start} - {end}\")\n\n            # [1, chunk_len] \u306e\u5f62\u306b\u623b\u3059\n            chunk = audio[start:end].unsqueeze(0)\n\n            # DFN3 \u3067\u30ce\u30a4\u30ba\u9664\u53bb\n            enhanced_chunk = enhance(model, df_state, chunk)\n\n            # CPU\u5074\u306b\u96c6\u3081\u3066\u304a\u304f\n            enhanced_chunk = enhanced_chunk.squeeze(0).cpu()\n            enhanced_chunks.append(enhanced_chunk)\n\n            start = end\n\n    # \u30c1\u30e3\u30f3\u30af\u3092\u7d50\u5408\n    enhanced_full = torch.cat(enhanced_chunks, dim=-1)\n    print(\"Enhanced total samples:\", enhanced_full.shape[-1])\n\n    # \u4fdd\u5b58\n    save_audio(OUTPUT_WAV, enhanced_full, df_state.sr())\n    print(\"Saved:\", OUTPUT_WAV)\n\n\nif __name__ == \"__main__\":\n    main()<\/code><\/pre>\n\n\n\n
\u5b9f\u884c\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
$ python test_dfn3.py<\/p>\n\n\n\n
\u79c1\u306e\u74b0\u5883\u3067\u306f\u3053\u308c\u3067\u7121\u4e8b\u52d5\u4f5c\u3057\u307e\u3057\u305f\u3002
\u52b9\u679c\u304c\u3042\u3063\u305f\u304b\u3001\u3068\u3044\u3046\u3068\u4eca\u56de\u306f\u201d\u5fae\u5999\u201d\u3068\u8a00\u308f\u3056\u308b\u3092\u5f97\u306a\u3044\u7d50\u679c\u3067\u3057\u305f\u3002
\u305f\u3060\u3001\u60aa\u304f\u306a\u3063\u305f\u90e8\u5206\u306f\u898b\u53d7\u3051\u3089\u308c\u306a\u304b\u3063\u305f\u306e\u3067\u3001
\u3044\u3063\u305f\u3093\u4ee5\u4e0b\u306e\u3088\u3046\u306a\u30d5\u30ed\u30fc\u306b\u3057\u3088\u3046\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n\n\n\n
mp4 \u2192 ffmpeg\u3067\u97f3\u58f0\u62bd\u51fa\uff08WAV\uff09
\u2193
\u30b9\u30c6\u30ec\u30aa \u2192 \u30e2\u30ce\u30e9\u30eb\uff08\u6574\u3048\u308b\uff09
\u2193
DFN3\uff08\u885d\u6483\u97f3\u30fb\u63a5\u89e6\u97f3\u30fb\u77ac\u9593\u30ce\u30a4\u30ba\u5bfe\u7b56\uff09
\u2193
Denoiser\uff08\u80cc\u666f\u30ce\u30a4\u30ba\u30fb\u30d6\u30ec\u30b9\u97f3\u3092\u8efd\u6e1b\u3001\u660e\u77ad\u5ea6UP\uff09
\u2193
ffmpeg\u3067\u52d5\u753b\u3068\u5408\u6210<\/p>\n\n\n\n
\u6b21\u56de\u306f\u3053\u306e\u30d5\u30ed\u30fc\u3067\u660e\u77ad\u306a\u54b3\u6255\u3044\u5165\u308a\u306e\u97f3\u58f0\u304c\u3069\u308c\u304f\u3089\u3044\u88dc\u6b63\u3055\u308c\u308b\u304b\u3092\u8a66\u3057\u3066\u307f\u307e\u3059\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u672c\u8a18\u4e8b\u306f\u3001\u3053\u308c\u307e\u3067\u306b\u69cb\u7bc9\u3057\u305f GPU \u5bfe\u5fdc PyTorch \u74b0\u5883\u3092\u524d\u63d0\u306b\u3001DeepFilterNet3 \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\u00a0\u3067\u3059\u3002 \u524d\u56de\u306e\u78ba\u8a8d\u3067\u3001Denoiser\u304cGPU\u3092\u4f7f\u7528\u3057 […]<\/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-36","post","type-post","status-publish","format-standard","hentry","category-tech"],"_links":{"self":[{"href":"https:\/\/www.em99.tech\/index.php?rest_route=\/wp\/v2\/posts\/36","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=36"}],"version-history":[{"count":2,"href":"https:\/\/www.em99.tech\/index.php?rest_route=\/wp\/v2\/posts\/36\/revisions"}],"predecessor-version":[{"id":38,"href":"https:\/\/www.em99.tech\/index.php?rest_route=\/wp\/v2\/posts\/36\/revisions\/38"}],"wp:attachment":[{"href":"https:\/\/www.em99.tech\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=36"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.em99.tech\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=36"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.em99.tech\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=36"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}