Adds failing test case for batched base distribution #18
Conversation
|
|
||
|
|
||
| def test_batched_base_distribution(affine_ar_bijector): | ||
| base_dist = dist.Normal(torch.zeros(1), torch.ones(1)) |
There was a problem hiding this comment.
This causes base_dist.batch_shape == torch.Size([1,]) so you get a [100, 1, 1] shaped sample from base_dist
There was a problem hiding this comment.
@feynmanliang hmm, I get a different result here.. I get a [100, 1] shaped sample from the transformed distribution... Could you share some code to reproduce?
On the other hand, for base_dist = dist.Normal(0, 1) I get an error since base_dist.batch_shape == torch.Size([]) and it tries to input a torch.Size([100]) into the MLP from the base distribution.
|
After thinking about this a while, I think we have a solution. Just to summarize what we discussed, there will be a |
|
There's been some upstream work to PyTorch/Pyro building on our discussion on the Transform/Bijector interface (thanks @fritzo @fehiepsi et al.): pytorch/pytorch#50581 We should integrate these changes before we proceed with shape tests |
|
In pyro-ppl/pyro#2753 I am updating @stefanwebb's original flow implementations in Pyro; hopefully you will be able to adapt some of those changes here. |
Closes #28