Add pattern for MultivariateNormal(affine) #245
Conversation
fritzo
changed the title
fritzo
changed the title
|
OK, I've fixed most of the shape errors and we're down to numerical errors only 😄 |
|
Yay all tests pass 😄 Thanks for all the help @fehiepsi ! |
| @@ -13,8 +14,49 @@ | |||
| from funsor.torch import Tensor | |||
|
|
|||
|
|
|||
| # This version constructs factors using funsor.distributions. | |||
| @pytest.mark.parametrize('state_dim,obs_dim', [(3, 2), (2, 3)]) | |||
| def test_distributions(state_dim, obs_dim): | |||
There was a problem hiding this comment.
@jpchen you can follow the idioms of this test in your experiment.
|
|
||
| # Compute log_prob using funsors. | ||
| scale_diag = Tensor(scale_tril.data.diagonal(dim1=-1, dim2=-2), scale_tril.inputs) | ||
| log_prob = (-0.5 * scale_diag.shape[0] * math.log(2 * math.pi) |
There was a problem hiding this comment.
Just curious: why we use shape[0] instead of shape[-1], scale_diag.log().sum() instead of scale_diag.log().sum(-1), and 0.5 * (const ** 2).sum() instead of 0.5 * (const ** 2).sum(-1) here?
There was a problem hiding this comment.
They are equivalent: scale_diag.shape[0] == scale_diag.shape[-1]. Here scale_diag is a funsor, and it separates "batch" .inputs from "event" .shape. In fact scale_diag.shape == (dim,) regardless of batching. This also allows us to call .sum() below rather than .sum(-1), since there is only one tensor dimension.
There was a problem hiding this comment.
That's nice! Thanks for explaining.
Addresses #72
pair coded with @eb8680
Tested
MultivariateNormal(x + y, ...)etc.dist.MultivariateNormaldirectly