Jake Gardner

@atharvahans1 We have a series of base kernel test cases that you can extend here: https://github.com/cornellius-gp/gpytorch/blob/master/gpytorch/test/base_kernel_test_case.py If after implementing the two abstract methods your kernel passes those tests, it's an...

@fleskovar yes, but a closely following line prevents it from working by default: https://github.com/cornellius-gp/gpytorch/blob/fc2053b0fc00517880fbc11adc7f5802242eec6a/gpytorch/models/exact_prediction_strategies.py#L232 The reason this is done is that otherwise making predictions with the model repeatedly would either...

I think an implementation that basically does the above (although of course also transpose other) would be a perfectly reasonable first pass at this. The only thing to note is...

@j-wilson Yeah, both the implementation and the basic test looks pretty good to me :+1:

To be honest, I'm not totally sure what's going on here because I'm not familiar enough with the exact internals of Pyro. This might have something to do with the...

This issue is basically superseded by #1578, as the bugs are caused by the same problem, and a fix to #1578 will resolve this issue as well. I've confirmed that...

@Balandat Since batch shapes of an arbitrary kernel are probably hard to modify on the fly for this, maybe we should special case operations that modify the batch shape to...

This is probably related to the total size of train and test crossing the `max_eager_kernel_size` boundary: https://github.com/cornellius-gp/gpytorch/blob/ade5db8df1f8e4acbc64ec1dd8809e58506827f1/gpytorch/settings.py#L395 Basically, we added an optimization a while back that unless the kernel matrix...

So if I understand that paper correctly, they basically derive a kernel (Equation 4) that allows you to define arbitrary functions over lengthscales and then when you compute `k(x_i, x_j)`...

@stanbiryukov Yeah, so basically what I'd do is define an new kernel: ```python class NonstationaryKernel(Kernel): def __init__(self, neural_net): # Register base modules, call super init etc def forward(self, x1, x2,...