Struct ZunaDecoder 

pub struct ZunaDecoder<B: Backend> {
    pub model_cfg: ModelConfig,
    pub data_cfg: DataConfig,
    pub global_sigma: f32,
    /* private fields */
}

Standalone ZUNA decoder.

Reconstructs EEG signals from latent embeddings produced by crate::encoder::ZunaEncoder using a rectified-flow diffusion loop.

Load with ZunaDecoder::load (decoder weights only — saves ~50 % memory vs the full crate::ZunaInference).

Fields

model_cfg: ModelConfig

Architecture hyperparameters (from config.json).

data_cfg: DataConfig

Preprocessing / tokenisation parameters.

global_sigma: f32

Diffusion noise standard deviation (σ).

Implementations

impl<B: Backend> ZunaDecoder<B>

pub fn load( config_path: &Path, weights_path: &Path, device: B::Device, ) -> Result<(Self, f64)>

Load decoder weights from a HuggingFace config.json and model.safetensors. Encoder tensors are read from disk but not kept in memory.

Returns (decoder, weight_load_ms).

pub fn describe(&self) -> String

One-line description of the loaded decoder.

pub fn decode_embeddings( &self, embeddings: &EncodingResult, steps: usize, cfg: f32, data_norm: f32, ) -> Result<InferenceResult>

Reconstruct EEG signals from pre-computed embeddings.

This is the pure decode path — no FIF loading or preprocessing happens here. The EncodingResult must have been produced by [ZunaEncoder::encode_fif] or [ZunaEncoder::encode_batch].

Arguments

• embeddings — pre-computed encoder latents
• steps — diffusion denoising steps (50 = full quality, 10 = fast)
• cfg — classifier-free guidance scale (1.0 = disabled)
• data_norm — divisor used during preprocessing; multiplied back into the output to restore the original signal scale

pub fn decode_tensor( &self, enc_out: Tensor<B, 3>, tok_idx: Tensor<B, 2, Int>, steps: usize, cfg: f32, ) -> Tensor<B, 3>

Decode a single epoch from a raw encoder output tensor [1, S, output_dim].

tok_idx is [S, 4]. Returns the reconstructed token matrix [1, S, input_dim] before inversion of the chop-and-reshape.