README.md

metadata

license: other
library_name: jax
tags:
  - nanopore
  - signal-reconstruction
  - audio-codec
  - flax
  - jax
pipeline_tag: feature-extraction

NanoRecon

NanoRecon is a Flax/JAX SimVQ-based model for nanopore signal reconstruction. This repository contains inference-ready model weights and the matching architecture config.

Files

• flax_model.msgpack: Flax variables for inference only, containing params and mutable vq codebook variables.
• config.json: model architecture and signal-shape metadata needed to instantiate the model.

The uploaded package is limited to inference assets and does not include local machine paths or private dataset references.

Loading Example

import json
from pathlib import Path

import jax
import jax.numpy as jnp
from flax.core import freeze
from flax.serialization import from_bytes

from codec.models import build_audio_model

repo_dir = Path("/path/to/NanoRecon")
cfg = json.loads((repo_dir / "config.json").read_text())
model = build_audio_model(cfg["model"])

variables = from_bytes(freeze({"params": {}, "vq": {}}), (repo_dir / "flax_model.msgpack").read_bytes())

x = jnp.zeros((1, cfg["input_signal"]["segment_samples"]), dtype=jnp.float32)
rng = jax.random.PRNGKey(0)
out = model.apply(
    variables,
    x,
    offset=0,
    rng=rng,
    collect_codebook_stats=False,
)
reconstructed = out["wave_hat"]

Notes

This model expects normalized nanopore signal chunks with shape (batch, samples). The exported config records the expected chunk size and sample rate.

Citation

No citation metadata has been provided yet.