musikalkemist / pytorchforaudio Goto Github PK

I get the issue on the script "07 Preprocessing data on GPU" even though I run the exact same script. Works on CPU though.
The error hits on the line " signal = self._resample_if_necessary(signal, sr)".

NB: Running on PyTorch 1.10.0+cu113 with RTX 2060(6 GB)

I'm getting the following error when running the urbansounddataset.py file:
resampled = torch.nn.functional.conv1d(waveform[:, None], kernel, stride=orig_freq)
RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.FloatTensor) should be the same. I've tried to figure out why by checking my variables, but i'm just struggling with that. Do you have any idea? I'm using windows, and torch 1.13.0+cu117.

I trained this urban sound classifier for 10 epochs, but there one problem
The loss stays extactly same except for 1st epochs. What could be the reason?

In case of image we know the maximum is 255 and we normalize it easily. and here the output signal of UrbanDataset is not normalized, is it a good approach to normalize each spectorgram with respect to itself?

if we normalize each audio spectogram based on its max and minimum, we kind of miss the amplitude features, am i right?

Thank you for you answer