This bug is observed in #646.
Note we specifically avoided implementing IndependentDistribution.enumerate_support() in torch.distributions because the size of the cartesian product space grows exponentially in event_shape.numel(), so most uses would be impractical or lead to OOM errors.
I see. Thanks for explaining! I am going to ask this question in slack but you are so fast. :))
Most helpful comment
Note we specifically avoided implementing
IndependentDistribution.enumerate_support()in torch.distributions because the size of the cartesian product space grows exponentially inevent_shape.numel(), so most uses would be impractical or lead to OOM errors.