Crate brainharmony 

brainharmony-rs — Brain-Harmony multimodal brain foundation model inference in Rust

Pure-Rust inference for the Brain-Harmony multimodal brain foundation model, built on Burn 0.20.

Brain-Harmony unifies morphology (T1 MRI) and function (fMRI) into 1D tokens using a Vision Transformer with:

• Brain gradient + geometric harmonics positioning for spatial embeddings
• Flexible patch embedding via Conv2d with dynamic patch size
• JEPA architecture (encoder + predictor with momentum target)

Three entry points

Type	Loads	Use case
BrainHarmonyEncoder	encoder only	produce latent embeddings
ClassificationHead	classification layer	downstream classification
MLPHead	3-layer MLP head	stage 2 finetuning

Quick start — encode brain signal

use brainharmony::{BrainHarmonyEncoder, ModelConfig, DataConfig};

let (enc, ms) = BrainHarmonyEncoder::<B>::from_weights(
    "model.safetensors",
    "gradient_mapping_400.csv",
    "schaefer400_roi_eigenmodes.csv",
    &ModelConfig::default(),
    &DataConfig::default(),
    &device,
)?;
let result = enc.encode_safetensors("data/signal.safetensors")?;
result.save_safetensors("embeddings.safetensors")?;

Backends

Feature	Backend	Notes
ndarray (default)	CPU (NdArray + Rayon)	Add blas-accelerate on macOS
wgpu	GPU (Metal / Vulkan)	--no-default-features --features wgpu
wgpu-f16	GPU (half precision)	--no-default-features --features wgpu-f16

Re-exports

pub use config::DataConfig;

pub use config::ModelConfig;

pub use config::YamlConfig;

pub use inference::BrainHarmonyEncoder;

pub use inference::EmbeddingResult;

pub use predictor_api::BrainHarmonyPredictor;

pub use classification::ClassificationHead;

pub use classification::MLPHead;

pub use classification::predict_classes;

pub use data::GradientData;

pub use data::GeohData;

pub use data::SignalInput;

pub use masks::MaskConfig;

pub use masks::full_context_mask;

pub use masks::jepa_masks;

pub use error::BrainHarmonyError;

pub use error::Result;

pub use model::encoder::apply_masks;

pub use weights::WeightFilter;

pub use weights::WeightMap;

pub use csv_export::save_embeddings_csv;

pub use hf_download::resolve as resolve_weights;

pub use hf_download::ResolvedWeights;

pub use hf_download::DEFAULT_REPO;

Modules

classification

config

csv_export

CSV export utilities for Brain-Harmony embeddings.

data

error

hf_download

inference

masks

model

predictor_api

prelude

weights

Functions

init_threads

Configure the global Rayon thread pool.