search

Rikorose / DeepFilterNet

Noise supression using deep filtering
https://huggingface.co/spaces/hshr/DeepFilterNet2
Other
2.18k stars 203 forks source link

Crashes when p_reverb=1.0 (both with DNS5 and SoundSpaces rirs) #584

Open enricguso opened 1 week ago

enricguso commented 1 week ago

When changing the default DFN3 recipe p_reverb to 1.0, it crashes during training, both for DNS and SoundSpaces RIR datasets.

2024-07-03 12:04:30 | TRACE | libdfdata.torchdataloader:df:dataset:1275 | Sampled RIR ....guso│ in24_rirs_train_SSmp3d_left_SSmp3d_left_Vt2qJdWjCF2_84_160.wav with shape [1, 8072] │ thread '' panicked at 'assertion failed: k <= self.len()', /rustc/5680fa18feaa87f3ff04063800│ aec256c3d4b4be/library/core/src/slice/mod.rs:3420:9 │ note: run with RUST_BACKTRACE=1 environment variable to display a backtrace │ Aborted (core dumped) │

Removing that particular RIR from the dataset seems to fix it, but the error appears very often making it hard to train.

Anyone has found a fix for this? Thanks

Rikorose commented 3 days ago

Removing that particular RIR from the dataset seems to fix it, but the error appears very often making it hard to train.

Do you mean that removing ..._SSmp3d_left_SSmp3d_left_Vt2qJdWjCF2_84_160.wav RIR seems to fix it? It then might be the problem with the shape. Maybe try to reshape it to a 1D Tensor [8072].

If this does not help, insert some dbg() messages within RandReverbSim::transform() in augmentation.rs