{ "id": 312747, "title": "torchaudio.load is much faster than librosa.load", "url": "/competitions/kaggle-pog-series-s01e02/discussion/312747", "author_name": "", "post_date": "2022-03-13T22:32:13.390962600Z", "votes": 12, "comment_count": 10, "views": 0, "content": "

I've just discovered that torchaudio.load is much faster than librosa.load (even with res_type='kaiser_fast').

\n

%%time
\ntorchaudio.load(filename)
\nCPU times: user 59.6 ms, sys: 5.72 ms, total: 65.4 ms
\nWall time: 81.3 ms

\n

%%time
\nlibrosa.load(filename, res_type='kaiser_fast')
\nCPU times: user 366 ms, sys: 2.43 ms, total: 368 ms
\nWall time: 376 ms

\n

However, the default MelSpectrogram of torchaudio is very dark. You can improve it by using the params as below:

\n
    specgram = torchaudio.transforms.MelSpectrogram(sample_rate=sr, \n                                                    n_fft=N_FFT, \n                                                    win_length=N_FFT, \n                                                    hop_length=HOP_LEN*4,\n                                                    center=True,\n                                                    pad_mode=\"reflect\",\n                                                    power=2.0,\n                                                    norm='slaney',\n                                                    onesided=True,\n                                                    n_mels=128,\n                                                    mel_scale=\"htk\"\n                                                   )(audio)[0]\n
\n

Check out my notebook for more details: https://www.kaggle.com/dienhoa/fastai-pogchamp-music-genre-classification

\n

Happy learning !

", "messages": [ { "id": "1721646", "postDate": "03/13/2022 22:32:13", "content": "

I've just discovered that torchaudio.load is much faster than librosa.load (even with res_type='kaiser_fast').

\n

%%time
\ntorchaudio.load(filename)
\nCPU times: user 59.6 ms, sys: 5.72 ms, total: 65.4 ms
\nWall time: 81.3 ms

\n

%%time
\nlibrosa.load(filename, res_type='kaiser_fast')
\nCPU times: user 366 ms, sys: 2.43 ms, total: 368 ms
\nWall time: 376 ms

\n

However, the default MelSpectrogram of torchaudio is very dark. You can improve it by using the params as below:

\n
    specgram = torchaudio.transforms.MelSpectrogram(sample_rate=sr, \n                                                    n_fft=N_FFT, \n                                                    win_length=N_FFT, \n                                                    hop_length=HOP_LEN*4,\n                                                    center=True,\n                                                    pad_mode=\"reflect\",\n                                                    power=2.0,\n                                                    norm='slaney',\n                                                    onesided=True,\n                                                    n_mels=128,\n                                                    mel_scale=\"htk\"\n                                                   )(audio)[0]\n
\n

Check out my notebook for more details: https://www.kaggle.com/dienhoa/fastai-pogchamp-music-genre-classification

\n

Happy learning !

", "rawMarkdown": "I've just discovered that torchaudio.load is much faster than librosa.load (even with res_type='kaiser_fast'). \n\n%%time\ntorchaudio.load(filename)\nCPU times: user 59.6 ms, sys: 5.72 ms, total: 65.4 ms\nWall time: 81.3 ms\n\n%%time\nlibrosa.load(filename, res_type='kaiser_fast')\nCPU times: user 366 ms, sys: 2.43 ms, total: 368 ms\nWall time: 376 ms\n\nHowever, the default MelSpectrogram of torchaudio is very dark. You can improve it by using the params as below:\n```\n specgram = torchaudio.transforms.MelSpectrogram(sample_rate=sr, \n n_fft=N_FFT, \n win_length=N_FFT, \n hop_length=HOP_LEN*4,\n center=True,\n pad_mode=\"reflect\",\n power=2.0,\n norm='slaney',\n onesided=True,\n n_mels=128,\n mel_scale=\"htk\"\n )(audio)[0]\n```\n\nCheck out my notebook for more details: https://www.kaggle.com/dienhoa/fastai-pogchamp-music-genre-classification\n\nHappy learning !", "votes": null }, { "id": "1721751", "postDate": "03/14/2022 01:53:32", "content": "

Very cool! Just keep in mind pretrained models are not allowed for this competition. You should be able to train something using the same pipeline easily though.

", "rawMarkdown": "Very cool! Just keep in mind pretrained models are not allowed for this competition. You should be able to train something using the same pipeline easily though.", "votes": null }, { "id": "1721821", "postDate": "03/14/2022 03:44:31", "content": "

Thanks. I will update my notebook with no pretrained model

", "rawMarkdown": "Thanks. I will update my notebook with no pretrained model", "votes": null }, { "id": "1722817", "postDate": "03/14/2022 22:15:33", "content": "

Great tip, I've been limiting my files to 2 seconds because loading them up was so slow using librosa.

", "rawMarkdown": "Great tip, I've been limiting my files to 2 seconds because loading them up was so slow using librosa.", "votes": null }, { "id": "1723436", "postDate": "03/15/2022 12:17:56", "content": "

Great point @dienhoa An additional point to add that I discovered today, Librosa converts stereo audio content (which is a matrix) to mono (a vector) which makes it easier to work on. Otherwise torch.load is way faster.

", "rawMarkdown": "Great point @dienhoa An additional point to add that I discovered today, Librosa converts stereo audio content (which is a matrix) to mono (a vector) which makes it easier to work on. Otherwise torch.load is way faster.", "votes": null }, { "id": "1723446", "postDate": "03/15/2022 12:27:37", "content": "

Thanks a lot @hadicurtay ! Wow I didn’t know that and found it’s weird that I got a matrix and I just put the 1st row to the model. Do you have a link that I can read about that ? How can we convert stereo to mono by torchaudio ?

", "rawMarkdown": "Thanks a lot @hadicurtay ! Wow I didn’t know that and found it’s weird that I got a matrix and I just put the 1st row to the model. Do you have a link that I can read about that ? How can we convert stereo to mono by torchaudio ?", "votes": null }, { "id": "1723830", "postDate": "03/15/2022 19:06:29", "content": "

You can merge two channels by mean function in torch

\n
audio, sample_rate_orig = torchaudio.load(filename)\naudio = torch.mean(audio, dim=0).unsqueeze(0)  # from stereo to mono\n
", "rawMarkdown": "You can merge two channels by mean function in torch\n```\naudio, sample_rate_orig = torchaudio.load(filename)\naudio = torch.mean(audio, dim=0).unsqueeze(0) # from stereo to mono\n```", "votes": null }, { "id": "1723836", "postDate": "03/15/2022 19:15:13", "content": "

I haven't tried using torchaudio to do that but going to try @shevan 's solution

", "rawMarkdown": "I haven't tried using torchaudio to do that but going to try @shevan 's solution", "votes": null }, { "id": "1741412", "postDate": "03/31/2022 19:34:29", "content": "

I think this is because librosa by default resamples audio to 22.05kHz, to disable this option you also have to call it like this

\n
audio, sr = librosa.load(path, sr=None)\n
", "rawMarkdown": "I think this is because librosa by default resamples audio to 22.05kHz, to disable this option you also have to call it like this \n```\naudio, sr = librosa.load(path, sr=None)\n```", "votes": null }, { "id": "1741414", "postDate": "03/31/2022 19:37:37", "content": "

Wow thanks ! did you compare the execution time between librosa and torch with your approach?

", "rawMarkdown": "Wow thanks ! did you compare the execution time between librosa and torch with your approach?", "votes": null }, { "id": "1742394", "postDate": "04/01/2022 20:28:54", "content": "

To response to my own question, yes you are right @martynoveduard , with sr=None torch and librosa just have the same loading time, thanks so much for the help

", "rawMarkdown": "To response to my own question, yes you are right @martynoveduard , with sr=None torch and librosa just have the same loading time, thanks so much for the help", "votes": null } ], "comments": [ { "id": 1721751, "author_name": "robikscube", "author_url": "", "post_date": "03/14/2022 01:53:32", "content": "

Very cool! Just keep in mind pretrained models are not allowed for this competition. You should be able to train something using the same pipeline easily though.

", "votes": null, "replies": [ { "id": 1721821, "author_name": "dienhoa", "author_url": "", "post_date": "03/14/2022 03:44:31", "content": "

Thanks. I will update my notebook with no pretrained model

", "votes": null, "replies": [] } ] }, { "id": 1722817, "author_name": "yousseftaoudi", "author_url": "", "post_date": "03/14/2022 22:15:33", "content": "

Great tip, I've been limiting my files to 2 seconds because loading them up was so slow using librosa.

", "votes": null, "replies": [] }, { "id": 1723436, "author_name": "hadicurtay", "author_url": "", "post_date": "03/15/2022 12:17:56", "content": "

Great point @dienhoa An additional point to add that I discovered today, Librosa converts stereo audio content (which is a matrix) to mono (a vector) which makes it easier to work on. Otherwise torch.load is way faster.

", "votes": null, "replies": [ { "id": 1723446, "author_name": "dienhoa", "author_url": "", "post_date": "03/15/2022 12:27:37", "content": "

Thanks a lot @hadicurtay ! Wow I didn’t know that and found it’s weird that I got a matrix and I just put the 1st row to the model. Do you have a link that I can read about that ? How can we convert stereo to mono by torchaudio ?

", "votes": null, "replies": [] }, { "id": 1723830, "author_name": "shevan", "author_url": "", "post_date": "03/15/2022 19:06:29", "content": "

You can merge two channels by mean function in torch

\n
audio, sample_rate_orig = torchaudio.load(filename)\naudio = torch.mean(audio, dim=0).unsqueeze(0)  # from stereo to mono\n
", "votes": null, "replies": [] }, { "id": 1723836, "author_name": "hadicurtay", "author_url": "", "post_date": "03/15/2022 19:15:13", "content": "

I haven't tried using torchaudio to do that but going to try @shevan 's solution

", "votes": null, "replies": [] } ] }, { "id": 1741412, "author_name": "martynoveduard", "author_url": "", "post_date": "03/31/2022 19:34:29", "content": "

I think this is because librosa by default resamples audio to 22.05kHz, to disable this option you also have to call it like this

\n
audio, sr = librosa.load(path, sr=None)\n
", "votes": null, "replies": [ { "id": 1741414, "author_name": "dienhoa", "author_url": "", "post_date": "03/31/2022 19:37:37", "content": "

Wow thanks ! did you compare the execution time between librosa and torch with your approach?

", "votes": null, "replies": [] }, { "id": 1742394, "author_name": "dienhoa", "author_url": "", "post_date": "04/01/2022 20:28:54", "content": "

To response to my own question, yes you are right @martynoveduard , with sr=None torch and librosa just have the same loading time, thanks so much for the help

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1721646": "I've just discovered that torchaudio.load is much faster than librosa.load (even with res_type='kaiser_fast'). \n\n%%time\ntorchaudio.load(filename)\nCPU times: user 59.6 ms, sys: 5.72 ms, total: 65.4 ms\nWall time: 81.3 ms\n\n%%time\nlibrosa.load(filename, res_type='kaiser_fast')\nCPU times: user 366 ms, sys: 2.43 ms, total: 368 ms\nWall time: 376 ms\n\nHowever, the default MelSpectrogram of torchaudio is very dark. You can improve it by using the params as below:\n```\n specgram = torchaudio.transforms.MelSpectrogram(sample_rate=sr, \n n_fft=N_FFT, \n win_length=N_FFT, \n hop_length=HOP_LEN*4,\n center=True,\n pad_mode=\"reflect\",\n power=2.0,\n norm='slaney',\n onesided=True,\n n_mels=128,\n mel_scale=\"htk\"\n )(audio)[0]\n```\n\nCheck out my notebook for more details: https://www.kaggle.com/dienhoa/fastai-pogchamp-music-genre-classification\n\nHappy learning !", "1721751": "Very cool! Just keep in mind pretrained models are not allowed for this competition. You should be able to train something using the same pipeline easily though.", "1721821": "Thanks. I will update my notebook with no pretrained model", "1722817": "Great tip, I've been limiting my files to 2 seconds because loading them up was so slow using librosa.", "1723436": "Great point @dienhoa An additional point to add that I discovered today, Librosa converts stereo audio content (which is a matrix) to mono (a vector) which makes it easier to work on. Otherwise torch.load is way faster.", "1723446": "Thanks a lot @hadicurtay ! Wow I didn’t know that and found it’s weird that I got a matrix and I just put the 1st row to the model. Do you have a link that I can read about that ? How can we convert stereo to mono by torchaudio ?", "1723830": "You can merge two channels by mean function in torch\n```\naudio, sample_rate_orig = torchaudio.load(filename)\naudio = torch.mean(audio, dim=0).unsqueeze(0) # from stereo to mono\n```", "1723836": "I haven't tried using torchaudio to do that but going to try @shevan 's solution", "1741412": "I think this is because librosa by default resamples audio to 22.05kHz, to disable this option you also have to call it like this \n```\naudio, sr = librosa.load(path, sr=None)\n```", "1741414": "Wow thanks ! did you compare the execution time between librosa and torch with your approach?", "1742394": "To response to my own question, yes you are right @martynoveduard , with sr=None torch and librosa just have the same loading time, thanks so much for the help" }, "source": "meta" }