The AffineTransform method seems to create two estimates for each sample. Why are both being created and are they required to do accurate estimates using the predict function?
Are you trying to transform multivariate distributions? I just encountered a similar confusion and the solution is to use LowerCholeskyAffine.
from pyro.distributions.transforms.lower_cholesky_affine import LowerCholeskyAffine
transform = LowerCholeskyAffine(loc=target_mean, scale_tril=transformation_matrix)
transformed_sample = transform(sample)