Resample Batch of Waveforms

[ahmed-fau]

Hello,
I would like to use the Resample class for up/down sampling the learned features during the training. Exactly the same way as F.interpolate. However, the batch dimension is not accepted by the Resample class so that I need to reshape the tensor to have only the channel and temporal dimensions.

Any help on how to resample a batch of waveforms/multi-channel features without any need to reshape?

Thanks in advance.

[vincentqb]

Thanks for pointing this out: I've opened a pull request to address this.

[ahmed-fau]

Thanks, @vincentqb for taking care of this. I should then let you know that the same behavior occurs when using other transform modules like Spectrogram and MelSpectrogram. The modules cannot accept three-dimensional tensors (batch, channel, time). Only two-dimensional tensors (channel, time) are supported so that the batch dimension is dealt as the cannel one, while the remaining dimensions should be flattened in order to work.

[vincentqb]

Can you provide an example for these? Which version are you using? These should support batching since #391.

[ahmed-fau]

Can you provide an example for these? Which version are you using? These should support batching since #391.

Sorry for the confusion. I tried this on a previous version. Batching for spectrograms works fine with version 0.4