VariationalPosterior.to_torch

VariationalPosterior.to_torch(orig_flag=True, *, dtype=None, device=None)

Export a snapshot as a torch probability distribution.

Parameters:

orig_flagbool, optional

Return a distribution in original parameter coordinates (default True), or in the internal unconstrained coordinates if False.

dtypetorch.dtype, optional

Export precision, either torch.float64 (default) or torch.float32. This does not change the NumPy posterior.

devicestr or torch.device, optional

Device for the distribution’s tensors, default CPU.

Returns:

distributiontorch.distributions.Distribution

A Gaussian mixture, transformed to original coordinates when requested, with event shape (D,). Its log_prob supports differentiation with respect to its input. Mixture sampling is not reparameterized.

Raises:

ImportError

If torch is absent; install pyvbmc[torch].

ValueError

If the posterior or transformer cannot define the distribution.

Notes

Parameters are copied; later changes to either object do not affect the other. Conversion draws no random numbers. Sampling the exported distribution uses torch’s RNG, controlled by torch.manual_seed, independently of self.rng. No torch objects are stored on this VP. In original space, density inputs must be finite and strictly inside parameter bounds. Samples rounded to a bound are moved to the nearest representable interior point; the analytic density transform is not clamped. Explicit float32 exports have less resolution near bounds.

See Torch distribution for a short example and installation details in Optional integrations. The worked PyVBMC Example 9: Torch and JAX models and posterior exports uses the export for posterior probability predictions and density gradients.