I wanted to quickly build a training pipeline for this competition, but mp3 files are way too slow to load.
\n\nI used this : \n```\n!pip install pydub\nimport pydub\nimport numpy as np
\n\ndef load_mp3(f):\n a = pydub.AudioSegment.from_mp3(f)\n y = np.array(a.get_array_of_samples()).astype(float)\n return y / 2**15\n```\nWhich takes >1h to load the whole data
\n\nThis is a bit faster but still too slow :\n```\n!pip install audio2numpy\nfrom audio2numpy import open_audio
\n\ndef load_mp3(f):\n return open_audio(f)[0]\n```
\n\nAnybody familiar with the mp3 format, aware of a quicker way to load samples ?\nThe only thing I can think of right now is converting everything to .wav files, which are faster to load.
The idea being that we'll probably need to load them on the fly during training/inference, since I don't think they'll fit in the ram
", "messages": [ { "id": "887763", "postDate": "06/15/2020 21:05:17", "content": "I wanted to quickly build a training pipeline for this competition, but mp3 files are way too slow to load.
\n\nI used this : \n```\n!pip install pydub\nimport pydub\nimport numpy as np
\n\ndef load_mp3(f):\n a = pydub.AudioSegment.from_mp3(f)\n y = np.array(a.get_array_of_samples()).astype(float)\n return y / 2**15\n```\nWhich takes >1h to load the whole data
\n\nThis is a bit faster but still too slow :\n```\n!pip install audio2numpy\nfrom audio2numpy import open_audio
\n\ndef load_mp3(f):\n return open_audio(f)[0]\n```
\n\nAnybody familiar with the mp3 format, aware of a quicker way to load samples ?\nThe only thing I can think of right now is converting everything to .wav files, which are faster to load.
The idea being that we'll probably need to load them on the fly during training/inference, since I don't think they'll fit in the ram
", "rawMarkdown": "I wanted to quickly build a training pipeline for this competition, but mp3 files are way too slow to load.\n\nI used this : \n```\n!pip install pydub\nimport pydub\nimport numpy as np\n\ndef load_mp3(f):\n a = pydub.AudioSegment.from_mp3(f)\n y = np.array(a.get_array_of_samples()).astype(float)\n return y / 2**15\n```\nWhich takes >1h to load the whole data \n\n\nThis is a bit faster but still too slow :\n```\n!pip install audio2numpy\nfrom audio2numpy import open_audio\n\ndef load_mp3(f):\n return open_audio(f)[0]\n```\n\nAnybody familiar with the mp3 format, aware of a quicker way to load samples ?\nThe only thing I can think of right now is converting everything to `.wav` files, which are faster to load.\n\nThe idea being that we'll probably need to load them on the fly during training/inference, since I don't think they'll fit in the ram", "votes": null }, { "id": "887811", "postDate": "06/15/2020 22:41:16", "content": "@theoviel \nI think we are not allowed to have an internet access under the competition rule.\nhttps://www.kaggle.com/c/birdsong-recognition/overview/code-requirements\nSo magic command, !pip install XXX will fail in inference time(not in training).
\n\nIn this case, I guess the best way to save our inference time is to use audioread directly not to make librosa fail in loading, as in the article below.\nhttps://stackoverflow.com/questions/59854527/librosa-load-takes-too-long-to-loadsample-mp3-files
update
\n\nI found the code to decode mp3 file to wave file and have just fixed a little not to return an error (file access error) in kernel environment.\nhttps://github.com/beetbox/audioread/blob/master/decode.py
\n\n```\nfrom future import print_function\nimport audioread\nimport sys\nimport os\nimport wave\nimport contextlib
\n\ndef decode(filename):\n if not os.path.exists(filename):\n print(\"File not found.\", file=sys.stderr)\n sys.exit(1)
\n\ntry:\n with audioread.audio_open(filename) as f:\n print('Input file: %i channels at %i Hz; %.1f seconds.' %\n (f.channels, f.samplerate, f.duration),\n file=sys.stderr)\n print('Backend:', str(type(f).__module__).split('.')[1],\n file=sys.stderr)\n\n with contextlib.closing(wave.open(fname.split('/')[-1] + '.wav', 'w')) as of:\n of.setnchannels(f.channels)\n of.setframerate(f.samplerate)\n of.setsampwidth(2)\n\n for buf in f:\n of.writeframes(buf)\n\nexcept audioread.DecodeError:\n print(\"File could not be decoded.\", file=sys.stderr)\n sys.exit(1)\nfname = '../input/birdsong-recognition/train_audio/swaspa/XC302566.mp3'\ndecode(fname)\nx, sr = librosa.load('/kaggle/working/' + fname.split('/')[-1] + '.wav')\nfig, ax = plt.subplots(1, 1, figsize=(10, 4))\nax.plot(range(len(x)), x, color='deeppink');\n```
\n\nOutput:
\n
You can install packages even without internet by adding the source code as an external data source, this is not an issue :)
\n\nThanks for sharing the snippet, although it is not faster than what I used, it can be convenient to save files as .wav
", "rawMarkdown": "You can install packages even without internet by adding the source code as an external data source, this is not an issue :)\n\nThanks for sharing the snippet, although it is not faster than what I used, it can be convenient to save files as .wav", "votes": null }, { "id": "888443", "postDate": "06/16/2020 11:06:25", "content": "@theoviel \nYes. As you mentioned, adding package source as an external data is one alternative seen in the some past competitions. Thank you for your advice :-)
\n\nPlease let me write additional comments, sir.
\nI simulated loading time with 2 packages you introduced.
Simulation conditions follow data description:
\n\n\n\n\ntest_audio \n The hidden test set audio consists of approximately 150 recordings in mp3 format, each roughly 10 minutes long.
\n
Cond 1. mp3 file with 600 sec audio length\nCond 2. iterate for 150 times
\n\ntest code:
\n```\nAUDIO_PATH = '../input/birdsong-recognition/train_audio/'
tmp = train.query('duration == 600').sample(1)\nbird_code = tmp[\"ebird_code\"].values[0]\nfname = tmp[\"filename\"].values[0]\ntmp_name = f'{AUDIO_PATH}{bird_code}/{fname}'
\n\n%%time\nfor _ in range(150):\n open_audio(tmp_name)
\n\n%%time\nfor _ in range(150):\n a = pydub.AudioSegment.from_mp3(tmp_name)\n y = np.array(a.get_array_of_samples()).astype(float)
\n\n```
\n\nResult:
\n1. audio2numpy
\nCPU times: user 4min 29s, sys: 3min 10s, total: 7min 40s\nWall time: 5min 14s
\n
As you said audio2numpy seems to be faster than pydub (although only one time check).
\nIf I'm not wrong, loading time in inference will not be a big problem with this result.
@theoviel \nI think we are not allowed to have an internet access under the competition rule.\nhttps://www.kaggle.com/c/birdsong-recognition/overview/code-requirements\nSo magic command, !pip install XXX will fail in inference time(not in training).
\n\nIn this case, I guess the best way to save our inference time is to use audioread directly not to make librosa fail in loading, as in the article below.\nhttps://stackoverflow.com/questions/59854527/librosa-load-takes-too-long-to-loadsample-mp3-files
update
\n\nI found the code to decode mp3 file to wave file and have just fixed a little not to return an error (file access error) in kernel environment.\nhttps://github.com/beetbox/audioread/blob/master/decode.py
\n\n```\nfrom future import print_function\nimport audioread\nimport sys\nimport os\nimport wave\nimport contextlib
\n\ndef decode(filename):\n if not os.path.exists(filename):\n print(\"File not found.\", file=sys.stderr)\n sys.exit(1)
\n\n![https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F479538%2F42e70d5cb56ded0611b2733753ef176c%2Fpng.PNG?generation=1592263876710001&alt=media](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F479538%2F42e70d5cb56ded0611b2733753ef176c%2Fpng.PNG?generation=1592263876710001&alt=media\"){kind=link}
try:\n with audioread.audio_open(filename) as f:\n print('Input file: %i channels at %i Hz; %.1f seconds.' %\n (f.channels, f.samplerate, f.duration),\n file=sys.stderr)\n print('Backend:', str(type(f).__module__).split('.')[1],\n file=sys.stderr)\n\n with contextlib.closing(wave.open(fname.split('/')[-1] + '.wav', 'w')) as of:\n of.setnchannels(f.channels)\n of.setframerate(f.samplerate)\n of.setsampwidth(2)\n\n for buf in f:\n of.writeframes(buf)\n\nexcept audioread.DecodeError:\n print(\"File could not be decoded.\", file=sys.stderr)\n sys.exit(1)\nfname = '../input/birdsong-recognition/train_audio/swaspa/XC302566.mp3'\ndecode(fname)\nx, sr = librosa.load('/kaggle/working/' + fname.split('/')[-1] + '.wav')\nfig, ax = plt.subplots(1, 1, figsize=(10, 4))\nax.plot(range(len(x)), x, color='deeppink');\n```
\n\nOutput:
\n
You can install packages even without internet by adding the source code as an external data source, this is not an issue :)
\n\nThanks for sharing the snippet, although it is not faster than what I used, it can be convenient to save files as .wav
", "votes": null, "replies": [] }, { "id": 888443, "author_name": "maxwell110", "author_url": "", "post_date": "06/16/2020 11:06:25", "content": "@theoviel \nYes. As you mentioned, adding package source as an external data is one alternative seen in the some past competitions. Thank you for your advice :-)
\n\nPlease let me write additional comments, sir.
\nI simulated loading time with 2 packages you introduced.
Simulation conditions follow data description:
\n\n\n\n\ntest_audio \n The hidden test set audio consists of approximately 150 recordings in mp3 format, each roughly 10 minutes long.
\n
Cond 1. mp3 file with 600 sec audio length\nCond 2. iterate for 150 times
\n\ntest code:
\n```\nAUDIO_PATH = '../input/birdsong-recognition/train_audio/'
tmp = train.query('duration == 600').sample(1)\nbird_code = tmp[\"ebird_code\"].values[0]\nfname = tmp[\"filename\"].values[0]\ntmp_name = f'{AUDIO_PATH}{bird_code}/{fname}'
\n\n%%time\nfor _ in range(150):\n open_audio(tmp_name)
\n\n%%time\nfor _ in range(150):\n a = pydub.AudioSegment.from_mp3(tmp_name)\n y = np.array(a.get_array_of_samples()).astype(float)
\n\n```
\n\nResult:
\n1. audio2numpy
\nCPU times: user 4min 29s, sys: 3min 10s, total: 7min 40s\nWall time: 5min 14s
\n
As you said audio2numpy seems to be faster than pydub (although only one time check).
\nIf I'm not wrong, loading time in inference will not be a big problem with this result.