msa-inr 0.1.0

Multi-Scale Sine Activation and INR network modules for PyTorch

MSA-INR

MSA-INR is a lightweight PyTorch package for using the Multi-Scale Sine Activation (MSA) in implicit neural representation (INR) models.

The package provides:

• SineActivation: a log-spaced multi-scale sine activation.
• MSANet: a simple MLP-style INR network using MSA.
• get_model: a convenient function for constructing MSANet.

---

Installation

After the package is published to PyPI:

pip install msa-inr

For local development:

git clone https://github.com/your-username/MSA_INR.git
cd MSA_INR
pip install -e .

---

Quick Start

Use the activation function

import torch
from msa_inr import SineActivation

activation = SineActivation(s_min=1.0, s_max=5.0)

x = torch.randn(8, 256)
y = activation(x)

print(y.shape)

Use the MSA-INR network

import torch
from msa_inr import MSANet

model = MSANet(
    in_features=2,
    hidden_features=256,
    hidden_layers=3,
    out_features=1,
    s_min=[1.0, 1.0, 1.0],
    s_max=[5.0, 5.0, 5.0],
)

coords = torch.rand(1024, 2)
pred = model(coords)

print(pred.shape)

Use get_model

from msa_inr import get_model

model = get_model(
    in_features=3,
    hidden_features=256,
    hidden_layers=2,
    out_features=1,
)

---

API

SineActivation

SineActivation(s_min: float, s_max: float)

For an input tensor with feature dimension C, the activation builds C log-spaced frequencies:

freqs = logspace(s_min, s_max, steps=C)

and returns:

sin(freqs * x)

The input tensor is expected to have shape:

[batch_size, channels]

---

MSANet

MSANet(
    in_features=3,
    hidden_features=256,
    hidden_layers=2,
    out_features=1,
    s_min=None,
    s_max=None,
)

If s_min and s_max are not provided, each MSA layer uses:

s_min = 1.0
s_max = 5.0

---

Build and Upload to PyPI

Install build tools:

pip install build twine

Build the package:

python -m build

Check the package:

twine check dist/*

Upload to PyPI:

twine upload dist/*

After uploading, users can install the package with:

pip install msa-inr

---

Citation

If this activation function or package is useful for your research, please cite the corresponding MSA-INR paper or repository.

@misc{msa_inr,
  title  = {MSA-INR: Multi-Scale Sine Activation for Implicit Neural Representations},
  author = {Han, Jufeng},
  year   = {2026},
  note   = {PyTorch package for multi-scale sine activation and INR models}
}