sanet 1.0.11

Spiking Attention Network model

sanet

Spiking Attention Network (sanet) model package.

Paper: Spiking Attention Network: A Hybrid Neuromorphic Approach to Underwater Acoustic Localization and Zero-shot Adaptation

Install

pip3 install sanet

Or

uv add sanet

Usage

Minimal

import torch
import sanet

model = sanet.SA_NET()

batch_size = 2
time_steps = 1500
x = torch.randn(batch_size, time_steps, 21)

with torch.no_grad():
	y = model(x)
	# expected torch.Size([2, 1])

print(y.shape)

All Parameters

import torch
import sanet

model = sanet.SA_NET(
	input_channels=21,
	output_channels=1,
	middle_channels=11,
	seed=42,
	spike_slope=25,
	lif1_beta=0.9956,
	lif2_beta=0.9821,
	lif3_beta=0.930,
	conformer_dim=512,
	conformer_depth=2,
	conformer_dim_head=64,
	conformer_heads=8,
	conformer_ff_mult=4,
	conformer_conv_expansion_factor=2,
	conformer_conv_kernel_size=24,
	conformer_attn_dropout=0.1,
	conformer_ff_dropout=0.1,
	conformer_conv_dropout=0.1,
	dropout_p1=0.0,
	dropout_p2=0.1,
	dropout_p3=0.1,
	dropout_p4=0.1,
)

batch_size = 2
time_steps = 1500
x = torch.randn(batch_size, time_steps, 21)

with torch.no_grad():
    y = model(x)
    # expected torch.Size([2, 1])

print(y.shape)

Model Notes

• Input tensor shape: [batch, time, channels]
• The forward pass applies per-channel standardization before the backbone.
• The network uses ResNet-style 1D blocks, spiking neurons, and Conformer layers.
• Model initialization sets a deterministic seed (Python, NumPy, and PyTorch) and enables deterministic CUDA behavior.

Architecture Diagram

flowchart TB
	A["Input (batch, time, channels)"] --> B["Per-channel Standardization"]

	subgraph Backbone
		C["ResNet1 + MaxPool + Dropout (p1)"] --> D["LIF1 Spiking"]
		D --> E["ResNet2 + MaxPool + Dropout (p2)"]
		E --> F["LIF2 Spiking"]
		F --> G["ResNet3 + MaxPool + Dropout (p3)"]
		G --> H["LIF3 Spiking"]
		H --> I["ResNet4 + MaxPool + Dropout (p4)"]
	end

	B --> C
	I --> J["Conformer Stack"]
	J --> K["Multilayer Perceptron"]
	K --> L["Output"]

Parameters

• input_channels (default: 21): Number of input channels.
• output_channels (default: 1): Number of output channels.
• middle_channels (default: 11): Number of intermediate channels before the final projection.
• seed (default: 42): Random seed for reproducibility.
• spike_slope (default: 25): Slope for the surrogate spike gradient.
• lif1_beta, lif2_beta, lif3_beta (default: 0.9956, 0.9821, 0.930): Decay rates for spiking neurons.
• conformer_dim (default: 512): Conformer model dimension.
• conformer_depth (default: 2): Number of Conformer blocks.
• conformer_dim_head (default: 64): Attention head dimension.
• conformer_heads (default: 8): Number of attention heads.
• conformer_ff_mult (default: 4): Feedforward expansion multiplier.
• conformer_conv_expansion_factor (default: 2): Conformer conv expansion factor.
• conformer_conv_kernel_size (default: 24): Conformer conv kernel size.
• conformer_attn_dropout (default: 0.1): Attention dropout in Conformer.
• conformer_ff_dropout (default: 0.1): Feedforward dropout in Conformer.
• conformer_conv_dropout (default: 0.1): Convolution dropout in Conformer.
• dropout_p1, dropout_p2, dropout_p3, dropout_p4 (default: 0.0, 0.1, 0.1, 0.1): Dropout probabilities for the 1D dropout layers.

Citation

If you find this package useful, please consider citing our paper:

@inproceedings{vo2026sa-net,
	title={Spiking Attention Network: A Hybrid Neuromorphic Approach to Underwater Acoustic Localization and Zero-shot Adaptation},
	author={Vo, Quoc Thinh and Han, David K},
	booktitle={2026 51st IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP)},
	pages={1--5},
	year={2026},
	organization={IEEE}
}