0.3.2

gptorch 0.3

Change log

0.3.0

Changes breaking backward compatibility:

• GPR, VFE, SVGP: training inputs order is changed from (y, x) to (x, y) on
  model __init__()s.
• .predict() functions return the same type as the inputs provided
  (numpy.ndarray->numpy.ndarray, torch.Tensor->torch.Tensor)
• Remove util.as_variable()
• Remove util.tensor_type()
• Remove util.KL_Gaussian()
• Remove util.gammaln()
• GPModel method .loss() generally replaces .compute_loss().
• .compute_loss() methods in models generally renamed to .log_likelihood()
  and signs flipped to reflect the fact that the loss is generally the negative
  LL.

Changes not breaking backward compatibility:

• GPR, VFE: Allow specifying training set on .compute_loss() with x, y kwargs
• GPR, VFE: Allow specifying training inputs on ._predict() with x kwarg
• GPU supported with .cuda()
• Remove GPModel.evaluate()
• Don't print inducing inputs on sparse GP initialization
• Suport for priors in gptorch.model.Models

0.3.1

• Fix some places where .compute_loss() wasn't replaced, causing GPModel.optimize() not to work.

0.3.2

• Issue 20 related to installing gptorch on top of pip-installed versions of PyTorch with non-standard device configurations.
• Issue 22 where importing gptorch changes the default dtype in PyTorch from single- to double-precision.
• Added gptorch.__version__

Authors

• Steven Atkinson