Segfault with python 3.9

[billbrod]

We're running into segfaults on python 3.9, which looks like it's a pytorch issue. In our code, It looks like it happens whenever .synthesize() is called, and, looking at the pytorch issue, that's because of .backward().

Minimum example:

import torch; import plenoptic as po
im = torch.randn((1, 1, 256, 256), requires_grad=True)
v1 = po.simul.PooledV1(.5, im.shape[-2:])
v1(im).sum().backward()

It looks like everything works fine, but then when you quit, you should get a message that the process was terminal by signal SIGSEGV or something that similarly tells you there was a segfault.

[billbrod]

So I think we keep an eye on it?

[billbrod]

Looks like it's fixed now, should be able to get this working.