Steven Atkinson

[output.wav](https://drive.google.com/file/d/1NrpQLBbCDHyu0RPsne4YcjIpi5-rEP6w/view?usp=sharing) to go with it for the tutorials

Neat. I'm potentially interested, but would want more data points. It would be nice to assemble a set of test cases. I assume this is a high-gain model? Orange is...

Ah, I remember: Models have the ability to `pad_start` when making predictions. This prepends enough silence that the first sample of the output can be predicted for. This is used...

Closing for now. I'll open something new (or hopefully find this w/ the search function) if it pops up again.

It looks like the fix is to get this implemented in PyTorch. The exception raised is: ``` The operator 'aten::angle' is not currently implemented for the MPS device. If you...

For now, a warning is printed: https://github.com/sdatkinson/neural-amp-modeler/blob/3899ccd703c73d159c9442e24d743dd3b4219558/nam/train/lightning_module.py#L401 and we use CPU to compute the MRSTFT loss when MPS can't. I tried only doing the angle op on CPU, but the...

Insight: [`torch.angle`](https://github.com/csteinmetz1/auraloss/blob/1576b0cd6e927abc002b23cf3bfc455b660f663c/auraloss/freq.py#L206) causes the problem, but this is only used for returning the phase. Since the default (which is used here) is to [ignore the phase component of the loss...

I think I'm going to fork auraloss, implement like this, then PR in case upstream likes it.

Ah, [this is already in 0.4](https://github.com/csteinmetz1/auraloss/blob/0853e9e732cd32a7b23f11ef4f0d2975d259e653/auraloss/freq.py#L151). Let's just use that then 😄

Correction: it's only on [the current tip of `main`](https://github.com/csteinmetz1/auraloss/commit/0853e9e732cd32a7b23f11ef4f0d2975d259e653). 0.4.0 is older. Bummer. Well, I can at least use that commit.