Crate nir_rs 

Pure-Rust implementation of the Neuromorphic Intermediate Representation (NIR).

NIR is a framework-agnostic graph format for spiking neural networks (analogous to ONNX for conventional nets). This crate provides a typed in-memory graph model plus HDF5 .nir read/write that interoperates with the Python reference implementation.

Status

v0.4 — developer experience: a public load-inspect-save example makes the HDF5 workflow executable end to end, on top of the v0.3 I/O and v0.2 graph model — the closed NirNode enum (wire-accurate type strings), NirGraph with ordered nodes/edges, Tensor / metadata types, and structured NirError.

I/O lives behind the opt-in hdf5 feature because it links the native libhdf5 library; the graph model itself has no system dependencies. See the io module for the feature gate, the file layout, and the version-string policy.

The independent opt-in serde feature implements Serde traits for the graph model. Formats such as JSON are useful for debugging and tests only: they are not a stable NIR schema or an interchange format. Use HDF5 .nir through io::read / io::write for interoperability. JSON also cannot represent non-finite floats faithfully, so NaN and infinities are not guaranteed to round-trip.

nir-rs = { version = "0.4", features = ["hdf5"] }

Run the complete fixture workflow from a checkout with:

cargo run --example load_inspect_lif --features hdf5

Example

use nir_rs::nodes::{Affine, Input, Lif, Output};
use nir_rs::types::Tensor;
use nir_rs::{NirGraph, NirNode};

let mut g = NirGraph::new();
g.insert_node(
    "input",
    NirNode::Input(Input {
        shape: vec![4],
        metadata: Default::default(),
    }),
)?;
g.insert_node(
    "fc",
    NirNode::Affine(Affine {
        weight: Tensor::from_f32(vec![2, 4], vec![0.1; 8])?,
        bias: Tensor::from_f32(vec![2], vec![0.0, 0.0])?,
        metadata: Default::default(),
    }),
)?;
g.insert_node(
    "lif",
    NirNode::Lif(Lif {
        tau: Tensor::from_f64(vec![2], vec![10.0, 10.0])?,
        r: Tensor::from_f64(vec![2], vec![1.0, 1.0])?,
        v_leak: Tensor::from_f64(vec![2], vec![0.0, 0.0])?,
        v_threshold: Tensor::from_f64(vec![2], vec![1.0, 1.0])?,
        v_reset: None,
        metadata: Default::default(),
    }),
)?;
g.insert_node(
    "output",
    NirNode::Output(Output {
        shape: vec![2],
        metadata: Default::default(),
    }),
)?;
g.add_edge("input", "fc");
g.add_edge("fc", "lif");
g.add_edge("lif", "output");
g.validate_structure()?;

// With `features = ["hdf5"]`, the graph exchanges as a `.nir` file:
// nir_rs::io::write("model.nir", &g)?;
// let reloaded = nir_rs::io::read("model.nir")?;

Non-goals

• SNN training or simulation
• Mapping graphs onto specific neuromorphic hardware
• Framework-specific converters (belong in producer/consumer tools)

Upstream

• neuromorphs/NIR
• neuroir.org
• Paper: Pedersen et al., Nat. Commun. 15, 8122 (2024) — please cite if you use NIR (see the repository README / CITATION.cff)

Wire type names must match the Python IR (CubaLIF, Conv2d, …), not informal marketing aliases.

Re-exports

pub use error::NirError;

pub use error::Result;

pub use graph::NirGraph;

pub use nodes::NirNode;

pub use types::DType;

pub use types::MetadataMap;

pub use types::MetadataValue;

pub use types::Tensor;

pub use types::TensorData;

Modules

error

Structured errors for nir-rs.

graph

In-memory NIR graph model.

io

HDF5 .nir read/write.

nodes

Wire-accurate NIR computational node types.

types

Shared value types for NIR tensors and metadata.