def test_lower_cholesky_affine(self):
for input_dim in [2, 3]:
loc = torch.randn(input_dim)
scale_tril = torch.randn(input_dim).exp().diag() + 0.03 * torch.randn(input_dim, input_dim)
scale_tril = scale_tril.tril(0)
transform = T.LowerCholeskyAffine(loc, scale_tril)
self._test_inverse(input_dim, transform)
self._test_jacobian(input_dim, transform)
for shape in [(3,), (3, 4)]:
self._test_shape(shape + (input_dim,), transform)
def transform_factory(input_dim):
loc = torch.randn(input_dim)
scale_tril = torch.randn(input_dim).exp().diag(
) + 0.03 * torch.randn(input_dim, input_dim)
scale_tril = scale_tril.tril(0)
return T.LowerCholeskyAffine(loc, scale_tril)