The competition almost come to the end and I realize that I did a huge mistake! I compute the MelSpectrogram on CPU. I can accelerate a lot by batch processing on GPU.
\nI experiment with a batch of 8 wave signals, 121 ms on CPU and 1.14 ms on GPU
\naudio, sr = torchaudio.load(filename)\ntransform = torchaudio.transforms.MelSpectrogram(sample_rate=sr, \n n_fft=N_FFT, \n win_length=N_FFT, \n hop_length=HOP_LEN,\n center=True,\n pad_mode=\"reflect\",\n power=2.0,\n norm='slaney',\n onesided=True,\n n_mels=256,\n mel_scale=\"htk\"\n )\ntransform_cuda = deepcopy(transform)\ntransform_cuda.to(cuda)\naudio_batch = audio.repeat(8,1)\naudio_batch_cuda = audio_batch.to(cuda)\n\n%%time\ntransform(audio_batch)\n# CPU times: user 132 ms, sys: 109 ms, total: 240 ms\n# Wall time: 121 ms\n\n%%time\ntransform(audio_batch)\n# CPU times: user 0 ns, sys: 1.99 ms, total: 1.99 ms\n# Wall time: 1.14 ms\nHope it is useful :D
\np/s: I think this is a huge advantage of torchaudio compared to librosa, right ?
", "messages": [ { "id": "1749687", "postDate": "04/08/2022 21:37:14", "content": "The competition almost come to the end and I realize that I did a huge mistake! I compute the MelSpectrogram on CPU. I can accelerate a lot by batch processing on GPU.
\nI experiment with a batch of 8 wave signals, 121 ms on CPU and 1.14 ms on GPU
\naudio, sr = torchaudio.load(filename)\ntransform = torchaudio.transforms.MelSpectrogram(sample_rate=sr, \n n_fft=N_FFT, \n win_length=N_FFT, \n hop_length=HOP_LEN,\n center=True,\n pad_mode=\"reflect\",\n power=2.0,\n norm='slaney',\n onesided=True,\n n_mels=256,\n mel_scale=\"htk\"\n )\ntransform_cuda = deepcopy(transform)\ntransform_cuda.to(cuda)\naudio_batch = audio.repeat(8,1)\naudio_batch_cuda = audio_batch.to(cuda)\n\n%%time\ntransform(audio_batch)\n# CPU times: user 132 ms, sys: 109 ms, total: 240 ms\n# Wall time: 121 ms\n\n%%time\ntransform(audio_batch)\n# CPU times: user 0 ns, sys: 1.99 ms, total: 1.99 ms\n# Wall time: 1.14 ms\nHope it is useful :D
\np/s: I think this is a huge advantage of torchaudio compared to librosa, right ?
", "rawMarkdown": "The competition almost come to the end and I realize that I did a huge mistake! I compute the MelSpectrogram on CPU. I can accelerate a lot by batch processing on GPU.\n\nI experiment with a batch of 8 wave signals, 121 ms on CPU and 1.14 ms on GPU \n\n```python\naudio, sr = torchaudio.load(filename)\ntransform = torchaudio.transforms.MelSpectrogram(sample_rate=sr, \n n_fft=N_FFT, \n win_length=N_FFT, \n hop_length=HOP_LEN,\n center=True,\n pad_mode=\"reflect\",\n power=2.0,\n norm='slaney',\n onesided=True,\n n_mels=256,\n mel_scale=\"htk\"\n )\ntransform_cuda = deepcopy(transform)\ntransform_cuda.to(cuda)\naudio_batch = audio.repeat(8,1)\naudio_batch_cuda = audio_batch.to(cuda)\n\n%%time\ntransform(audio_batch)\n# CPU times: user 132 ms, sys: 109 ms, total: 240 ms\n# Wall time: 121 ms\n\n%%time\ntransform(audio_batch)\n# CPU times: user 0 ns, sys: 1.99 ms, total: 1.99 ms\n# Wall time: 1.14 ms\n```\n\nHope it is useful :D\n\np/s: I think this is a huge advantage of torchaudio compared to librosa, right ?", "votes": null }, { "id": "1749839", "postDate": "04/09/2022 03:53:31", "content": "Yes, torchaudio >>> librosa.
", "rawMarkdown": "Yes, torchaudio >>> librosa.", "votes": null }, { "id": "1750153", "postDate": "04/09/2022 11:31:32", "content": "Wow, this is indeed helpful. I entered a competition related to audio files classification and went with librosa, however this seems much faster. Thanks!
", "rawMarkdown": "Wow, this is indeed helpful. I entered a competition related to audio files classification and went with librosa, however this seems much faster. Thanks!", "votes": null }, { "id": "1750642", "postDate": "04/09/2022 23:15:27", "content": "That sounds great. Thank you for sharing the tips.
", "rawMarkdown": "That sounds great. Thank you for sharing the tips.", "votes": null }, { "id": "1753497", "postDate": "04/12/2022 22:12:14", "content": "Thanks for sharing :) didn't know mel_spectrograms could be computed in a batch
", "rawMarkdown": "Thanks for sharing :) didn't know mel_spectrograms could be computed in a batch", "votes": null }, { "id": "1760522", "postDate": "04/19/2022 11:10:58", "content": "Hi, thank you for sharing! How do you create a batch from audiofiles with different durations?
", "rawMarkdown": "Hi, thank you for sharing! How do you create a batch from audiofiles with different durations?", "votes": null }, { "id": "1760814", "postDate": "04/19/2022 14:24:39", "content": "I think we can't. You have to crop your files to have the same duration. Pls tell me if you discover sth else.
", "rawMarkdown": "I think we can't. You have to crop your files to have the same duration. Pls tell me if you discover sth else.", "votes": null } ], "comments": [ { "id": 1749839, "author_name": "pheadrus", "author_url": "", "post_date": "04/09/2022 03:53:31", "content": "Yes, torchaudio >>> librosa.
", "votes": null, "replies": [] }, { "id": 1750153, "author_name": "spyrosrigas", "author_url": "", "post_date": "04/09/2022 11:31:32", "content": "Wow, this is indeed helpful. I entered a competition related to audio files classification and went with librosa, however this seems much faster. Thanks!
", "votes": null, "replies": [] }, { "id": 1750642, "author_name": "satoshiss", "author_url": "", "post_date": "04/09/2022 23:15:27", "content": "That sounds great. Thank you for sharing the tips.
", "votes": null, "replies": [] }, { "id": 1753497, "author_name": "pcmiller", "author_url": "", "post_date": "04/12/2022 22:12:14", "content": "Thanks for sharing :) didn't know mel_spectrograms could be computed in a batch
", "votes": null, "replies": [] }, { "id": 1760522, "author_name": "shevan", "author_url": "", "post_date": "04/19/2022 11:10:58", "content": "Hi, thank you for sharing! How do you create a batch from audiofiles with different durations?
", "votes": null, "replies": [ { "id": 1760814, "author_name": "dienhoa", "author_url": "", "post_date": "04/19/2022 14:24:39", "content": "I think we can't. You have to crop your files to have the same duration. Pls tell me if you discover sth else.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1749687": "The competition almost come to the end and I realize that I did a huge mistake! I compute the MelSpectrogram on CPU. I can accelerate a lot by batch processing on GPU.\n\nI experiment with a batch of 8 wave signals, 121 ms on CPU and 1.14 ms on GPU \n\n```python\naudio, sr = torchaudio.load(filename)\ntransform = torchaudio.transforms.MelSpectrogram(sample_rate=sr, \n n_fft=N_FFT, \n win_length=N_FFT, \n hop_length=HOP_LEN,\n center=True,\n pad_mode=\"reflect\",\n power=2.0,\n norm='slaney',\n onesided=True,\n n_mels=256,\n mel_scale=\"htk\"\n )\ntransform_cuda = deepcopy(transform)\ntransform_cuda.to(cuda)\naudio_batch = audio.repeat(8,1)\naudio_batch_cuda = audio_batch.to(cuda)\n\n%%time\ntransform(audio_batch)\n# CPU times: user 132 ms, sys: 109 ms, total: 240 ms\n# Wall time: 121 ms\n\n%%time\ntransform(audio_batch)\n# CPU times: user 0 ns, sys: 1.99 ms, total: 1.99 ms\n# Wall time: 1.14 ms\n```\n\nHope it is useful :D\n\np/s: I think this is a huge advantage of torchaudio compared to librosa, right ?", "1749839": "Yes, torchaudio >>> librosa.", "1750153": "Wow, this is indeed helpful. I entered a competition related to audio files classification and went with librosa, however this seems much faster. Thanks!", "1750642": "That sounds great. Thank you for sharing the tips.", "1753497": "Thanks for sharing :) didn't know mel_spectrograms could be computed in a batch", "1760522": "Hi, thank you for sharing! How do you create a batch from audiofiles with different durations?", "1760814": "I think we can't. You have to crop your files to have the same duration. Pls tell me if you discover sth else." }, "source": "meta" }