pore-codec-rsqf42c12a-510

PoreCodec-RSQF42C12A

A lightweight neural codec for nanopore electrical signals based on convolutional feature extraction and Residual Finite Scalar Quantization (Residual FSQ). The model converts continuous nanopore current signals into hierarchical discrete token sequences that can be directly consumed by language models, sequence models, retrieval systems, or downstream bioinformatics applications.

Overview

PoreCodec-RSQF42C12A bridges continuous nanopore electrical signals and discrete sequence modeling. It learns compact discrete representations using a convolutional encoder followed by Residual Finite Scalar Quantization (Residual FSQ), enabling efficient token-based modeling of nanopore signals without the codebook collapse commonly associated with VQ-VAE methods.

The model consists of three major components:

1. CNN Encoder

   • Extracts latent representations from raw electrical current signals.
   • Downsampling factor: 4×
   • Output feature dimension: 512

2. Residual Finite Scalar Quantizer (Residual FSQ)

   • Multi-stage scalar quantization.
   • Hierarchical residual coding.
   • Produces discrete token representations suitable for transformer-based models.

3. CNN Decoder

   • Reconstructs normalized nanopore signals from discrete tokens.

---

Model Architecture

Raw Signal
      │
      ▼
Feature Extractor
      │
      ▼
 CNN Encoder
      │
      ▼
 Linear Projection
      │
      ▼
 Residual FSQ
      │
      ▼
Discrete Tokens
      │
      ▼
 Linear Projection
      │
      ▼
 CNN Decoder
      │
      ▼
Reconstructed Signal

---

CNN Encoder

The encoder consists of:

• Conv1D
• BatchNorm
• SiLU activation
• Two stride-2 downsampling stages

---

Residual FSQ

Residual quantization is performed using multiple Finite Scalar Quantizers (FSQ).

Each quantizer encodes the residual error from the previous stage, producing hierarchical discrete representations while maintaining high reconstruction fidelity.

---

Signal Preprocessing

The accompanying feature extractor performs automatic preprocessing before inference.

Pipeline:

1. Physical boundary correction
2. Spike removal
3. Robust Median-MAD normalization
4. Optional median filtering
5. Smooth nonlinear clipping

Two preprocessing strategies are available.

apple (default)

• Error repair
• Spike removal
• Median-MAD normalization
• Median filtering
• Smooth clipping

mongo

• Error repair
• Spike removal
• Median-MAD normalization
• Smooth clipping

---

Quick Start

import numpy as np
from transformers import AutoFeatureExtractor, AutoModel

model_name = "ShuaiAnwo/pore-codec-rsqf42c12a-510"

feature_extractor = AutoFeatureExtractor.from_pretrained(
    model_name,
    trust_remote_code=True,
)

model = AutoModel.from_pretrained(
    model_name,
    trust_remote_code=True,
)

model.eval()

raw_signal = np.random.normal(
    loc=70.0,
    scale=8.0,
    size=1855,
).astype(np.float32)

signal = feature_extractor(
    raw_signal,
    return_tensors="pt",
)["signal"]

token_ids = model.encode_signal(
    signal,
    layer=2,
)

reconstructed = model.decode_token(
    token_ids,
    layer=2,
)

---

Usage

Load the Model

from transformers import (
    AutoConfig,
    AutoFeatureExtractor,
    AutoModel,
)

model_name = "ShuaiAnwo/pore-codec-rsqf42c12a-510"

config = AutoConfig.from_pretrained(
    model_name,
    trust_remote_code=True,
)

feature_extractor = AutoFeatureExtractor.from_pretrained(
    model_name,
    trust_remote_code=True,
)

model = AutoModel.from_pretrained(
    model_name,
    trust_remote_code=True,
)

model.eval()

print(model.num_quantizers)

Expected output

2

---

Prepare Input

The feature extractor expects a one-dimensional NumPy array containing raw nanopore current measurements.

import numpy as np

raw_signal = np.random.normal(
    loc=70.0,
    scale=8.0,
    size=1855,
).astype(np.float32)

signal = feature_extractor(
    raw_signal,
    return_tensors="pt",
)["signal"]

print(signal.shape)

Expected output

torch.Size([1, 1, 1855])

---

Encode Signals

encode_signal() converts normalized nanopore signals into hierarchical discrete token sequences.

token_ids = model.encode_signal(
    signal,
    layer=2,
)

print(token_ids.shape)
print(token_ids[0, :20])

Expected output

torch.Size([1, 464])

tensor([
1132546210, 766619217, 767357412, 743923015,
744668606, 916327422, 1599713621, 1419557526,
1293561428, 963177038, 1086246378, 1270061142,
1280753282, 949579165, 1132645766, 1098553593,
1097767990, 961814960, 1269315414, 1600419564
])

Supported layers

---

Decode Tokens

Discrete token sequences can be reconstructed back into normalized nanopore signals.

reconstructed = model.decode_token(
    token_ids,
    layer=2,
)

print(reconstructed.shape)

Expected output

torch.Size([1, 1, 1856])

---

Forward API

The forward interface performs end-to-end encoding and decoding in a single call.

reconstruction, level_indices = model(signal)

print(reconstruction.shape)

for i, indices in enumerate(level_indices):
    print(f"Quantizer {i}: {indices.shape}")

Returns

• reconstruction: reconstructed normalized signal
• level_indices: a list of quantization indices produced by each residual quantizer

Expected output

torch.Size([1, 1, 1856])

Quantizer 0: torch.Size([1, 464])
Quantizer 1: torch.Size([1, 464])

---

Complete Example

import numpy as np
from transformers import (
    AutoFeatureExtractor,
    AutoModel,
)

model_name = "ShuaiAnwo/pore-codec-rsqf42c12a-510"

feature_extractor = AutoFeatureExtractor.from_pretrained(
    model_name,
    trust_remote_code=True,
)

model = AutoModel.from_pretrained(
    model_name,
    trust_remote_code=True,
)

model.eval()

raw_signal = np.random.normal(
    loc=70.0,
    scale=8.0,
    size=1855,
).astype(np.float32)

signal = feature_extractor(
    raw_signal,
    return_tensors="pt",
)["signal"]

token_ids = model.encode_signal(
    signal,
    layer=2,
)

reconstructed = model.decode_token(
    token_ids,
    layer=2,
)

print("Signal shape :", signal.shape)
print("Token shape  :", token_ids.shape)
print("Recon shape  :", reconstructed.shape)

Expected output

Signal shape : torch.Size([1, 1, 1855])
Token shape  : torch.Size([1, 464])
Recon shape  : torch.Size([1, 1, 1856])

---

Validation Output

The following output was generated using the provided validation script and a real nanopore read.

============================================================
🚀 Starting PoreRSQCodec Closed-Loop Validation Pipeline
🎯 Target Analytical Layer (TARGET_LAYER): 2
============================================================

-> Safetensors structural keys count: 61

[Step 1] Loading local auto-mapped components...

-> AutoConfig loaded successfully!
-> AutoFeatureExtractor loaded successfully!

Loading weights: 100%|████████████████████████████████| 61/61

-> AutoModel (Safetensors weights) loaded successfully!
-> Total residual quantizer layers (num_quantizers): 2

✅ Component decoupling test passed!

[Step 2] Parsing real nanopore raw signals...

Read ID:
250F600084012_1_202_674_11920089_12557

Signal length:
1855

Current range:
[34.10, 99.00] pA

[Step 3] Feature preprocessing

Strategy:
apple

Normalized tensor:
torch.Size([1, 1, 1855])

✅ FeatureExtractor pipeline verification passed!

[Step 4] Encoding

Token shape:
torch.Size([1, 464])

Token value range:
[562424211, 1818974092]

✅ Quantization completed!

[Step 5] Decoding

Layer 1 reconstruction:
torch.Size([1, 1, 1856])

Layer 0 reconstruction:
torch.Size([1, 1, 1856])

Layer 2 reconstruction:
torch.Size([1, 1, 1856])

Ground Truth Mean:
68.2291

Layer 1 Mean:
-0.1262

Layer 0 Mean:
-0.1333

Layer 2 Mean:
-0.1333

✅ Reconstruction completed!

---

Applications

This model can be used for:

• Nanopore signal tokenization
• Neural signal compression
• Discrete representation learning
• Foundation models for nanopore sequencing
• Biological sequence modeling
• Token-based pretraining for genomic language models
• Retrieval and indexing of nanopore signals

---

Model Configuration

---

License

Please refer to the repository license for usage terms.

---

Citation

If you use this model in your research, please cite the repository or the corresponding publication.

@misc{porecodec2026,
  title={PoreCodec: Residual Finite Scalar Quantization for Nanopore Signal Tokenization},
  author={Shuai Jiao},
  year={2026}
}