Quick Start

This page walks through a minimal forward pass, a training command, and instance-level inference.

Minimal Forward Pass

import torch
from spikeseg.models import SpikeSEG

# Instantiate from the IGARSS 2023 paper configuration
model = SpikeSEG.from_paper("igarss2023", n_classes=1)

# Simulate an event frame: (batch, channels, height, width)
input_events = torch.randn(1, 1, 128, 128)

model.reset_state()
for t in range(10):
    saliency, encoder_output = model(input_events)
    if encoder_output.has_spikes:
        print(f"t={t}: {encoder_output.n_classification_spikes} classification spikes")

saliency is a pixel-level map; encoder_output holds per-layer spikes, pooling indices, and membrane potentials.

Train with One Command

python scripts/train.py --config configs/config.yaml

Training is layer-wise: Conv1 uses fixed DoG filters, Conv2 and Conv3 are trained via STDP with WTA competition. The script saves checkpoints, TensorBoard logs, and a final metrics JSON. See Training Guide for details.

Instance Segmentation with HULK-SMASH

import torch
from spikeseg.models import SpikeSEG
from spikeseg.algorithms import HULKDecoder, group_instances_to_objects

model = SpikeSEG.from_paper("igarss2023", n_classes=1)
model.load_state_dict(torch.load("checkpoint.pth"))

hulk = HULKDecoder.from_encoder(model.encoder)

model.reset_state()
encoder_output = model.encode(input_events)

instances = []
for spike_loc in encoder_output.get_spike_locations():
    inst = hulk.unravel_spike(
        spike_location=spike_loc,
        pool1_indices=encoder_output.pooling_indices.pool1_indices,
        pool2_indices=encoder_output.pooling_indices.pool2_indices,
        pool1_output_size=encoder_output.pooling_indices.pool1_output_size,
        pool2_output_size=encoder_output.pooling_indices.pool2_output_size,
    )
    instances.append(inst)

objects = group_instances_to_objects(instances, smash_threshold=0.1)
print(f"Detected {len(objects)} objects from {len(instances)} instances")

Evaluate a Checkpoint

python scripts/evaluate.py \
  --checkpoint checkpoints/best.pt \
  --data-root /path/to/EBSSA \
  --volume-based \
  --output metrics.json

This runs volume-based evaluation (IGARSS 2023 methodology) and writes all metrics to metrics.json. See Evaluation Guide.

What Next

Goal	Page
Understand the theory	Spiking Neurons
See the full architecture	Architecture Overview
Configure training	Configuration
Browse the API	API Reference