Struct DiagonalGaussianDistribution 

pub struct DiagonalGaussianDistribution<T: Float> {
    pub mean: Tensor<T>,
    pub logvar: Tensor<T>,
}

Diagonal Gaussian over latent space — the same parameterization diffusers.models.autoencoders.vae.DiagonalGaussianDistribution uses. Holds mean and logvar tensors (both [B, L, h, w]) split from the encoder’s concatenated parameters output.

logvar is clamped to [-30, 20] on construction, matching the diffusers reference exactly.

Fields

mean: Tensor<T>

Mean of the diagonal Gaussian, [B, L, h, w].

logvar: Tensor<T>

Log-variance of the diagonal Gaussian (clamped), [B, L, h, w].

Implementations

impl<T: Float> DiagonalGaussianDistribution<T>

pub fn from_parameters( params: &Tensor<T>, latent_channels: usize, ) -> FerrotorchResult<Self>

Split the encoder’s [B, 2*L, h, w] parameters tensor into mean / logvar halves along the channel axis.

Errors

Returns FerrotorchError::ShapeMismatch when the channel dimension is not 2 * latent_channels. Propagates downstream errors from the chunk / clamp ops.

pub fn mode(&self) -> &Tensor<T>

Return the distribution mode — the deterministic latent if the caller does not want to sample. Matches DiagonalGaussianDistribution.mode() in diffusers.

pub fn sample_with_seed(&self, seed: u64) -> FerrotorchResult<Tensor<T>>

Sample from the distribution with a deterministic seed.

Computes mean + std * eps where std = exp(0.5 * logvar) and eps is a fresh N(0, 1) tensor produced by Box-Muller over a seeded xorshift64 PRNG (mirroring the CPU branch of ferrotorch_core::randn but using seed as the state).

This is fully deterministic for a given seed on a given host — but the bitwise output will NOT match a CUDA PyTorch reference run, because PyTorch’s CUDA RNG uses Philox, not xorshift + Box-Muller. Tests should compare statistical properties (mean, std) or use .mode() for bit-exact reference matches.

Errors

Returns the underlying FerrotorchError for shape or arithmetic failures.

Trait Implementations

impl<T: Debug + Float> Debug for DiagonalGaussianDistribution<T>

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.