search

googlecreativelab / open-nsynth-super

Open NSynth Super is an experimental physical interface for the NSynth algorithm
Apache License 2.0
2.42k stars 270 forks source link

05_clean_files.py fails with audioread.NoBackendError #53

Open 9600 opened 6 years ago

9600 commented 6 years ago

Run on a 2x GPU system with 4x instruments per corner. Completes fine with one batch and not the other. Guess the error is misleading...

@jesseengel, wondering if you've seen this? (and doesn't seem like anyone else is watching this repo!)

//

(magenta) andrew@blast:/data/Liles/nsynth/open-nsynth-super/audio/workdir$ (magenta) andrew@blast:/data/Liles/nsynth/open-nsynth-super/audio/workdir$ python 05_clean_files.py 1 Normalising input WAVs 2%|█▊ | 992/57784 [03:26<3:16:57, 4.81it/s]100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████| 57784/57784 [3:24:43<00:00, 4.70it/s] Converting to WAV 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████| 104431/104431 [41:26<00:00, 42.00it/s] (magenta) andrew@blast:/data/Liles/nsynth/open-nsynth-super/audio/workdir$ (magenta) andrew@blast:/data/Liles/nsynth/open-nsynth-super/audio/workdir$ (magenta) andrew@blast:/data/Liles/nsynth/open-nsynth-super/audio/workdir$ (magenta) andrew@blast:/data/Liles/nsynth/open-nsynth-super/audio/workdir$ (magenta) andrew@blast:/data/Liles/nsynth/open-nsynth-super/audio/workdir$ (magenta) andrew@blast:/data/Liles/nsynth/open-nsynth-super/audio/workdir$ python 05_clean_files.py 0 Normalising input WAVs 27%|█████████████████████████████▊ | 15537/57784 [58:29<2:39:03, 4.43it/s]sox WARN gain: gain clipped 63999 samples; decrease volume? sox WARN dither: dither clipped 56045 samples; decrease volume? 28%|███████████████████████████████ | 16454/57784 [1:01:56<2:35:35, 4.43it/s]sox WARN gain: gain clipped 63902 samples; decrease volume? sox WARN dither: dither clipped 56048 samples; decrease volume? 70%|███████████████████████████████████████████████████████████████████████████▉ | 40228/57784 [2:31:36<1:06:09, 4.42it/s]sox WARN gain: gain clipped 64000 samples; decrease volume? sox WARN dither: dither clipped 55997 samples; decrease volume? 81%|█████████████████████████████████████████████████████████████████████████████████████████▌ | 46647/57784 [2:55:49<41:58, 4.42it/s]Traceback (most recent call last): File "05_clean_files.py", line 46, in audio, sr = librosa.core.load(os.path.join(original_path, fpath), sr=16000) File "/data/Liles/nsynth/miniconda2/envs/magenta/lib/python2.7/site-packages/librosa/core/audio.py", line 112, in load with audioread.audio_open(os.path.realpath(path)) as input_file: File "/data/Liles/nsynth/miniconda2/envs/magenta/lib/python2.7/site-packages/audioread/init.py", line 116, in audio_open raise NoBackendError() audioread.NoBackendError

jesseengel commented 6 years ago

Hey, Isn't this something you saw and fixed in #50?

I've seen this before as an upstream audioread/librosa issue (ex: https://github.com/librosa/librosa/issues/390, https://github.com/librosa/librosa/issues/589, https://github.com/librosa/librosa/issues/650).

9600 commented 6 years ago

@jesseengel looks like it's upstream. Was failing on just one file, which seems OK:

$ soxi gen_drone1_0.596_drone2_0.357_piano_0.047_pitch_48.wav

Input File : 'gen_drone1_0.596_drone2_0.357_piano_0.047_pitch_48.wav' Channels : 1 Sample Rate : 16000 Precision : 24-bit Duration : 00:00:04.00 = 64000 samples ~ 300 CDDA sectors File Size : 256k Bit Rate : 512k Sample Encoding: 32-bit Floating Point PCM

By modifying 05_clean_files.py so that it didn't just crash out and instead logged the error and continued, I could then bodge it by copying the neighbouring interpolation point WAV for those instruments and proceed to the next stage.