Skip to main content

Configuration Guide

All training, model, and data parameters are specified in YAML files under configs/.

Loading Configuration​

from spikeseg.config import load_config, merge_configs

config = load_config("configs/config.yaml")

# Override specific keys
config = load_config("configs/config.yaml", overrides={"max_epochs": 100})

Full Configuration Reference​

Experiment​

experiment_name: "spikeseg_default"
output_dir: "./runs"
seed: 42
device: "cuda"           # "cuda", "cpu", "cuda:0"
max_epochs: 50
train_conv1: false        # Conv1 = fixed DoG filters
train_conv2: true
train_conv3: true

Model Architecture​

model:
  n_classes: 1
  conv1_channels: 4
  conv2_channels: 36
  kernel_sizes: [5, 5, 7]
  pool1_kernel: 2
  pool1_stride: 2
  pool2_kernel: 2
  pool2_stride: 2
  thresholds: [0.1, 0.1, 0.1]
  leaks: [0.09, 0.01, 0.0]      # 90%, 10%, 0% of threshold
  use_dog_filters: true

STDP Learning​

stdp:
  lr_plus: 0.04            # LTP rate (IGARSS 2023)
  lr_minus: 0.03           # LTD rate
  weight_min: 0.0
  weight_max: 1.0
  weight_init_mean: 0.8
  weight_init_std: 0.01
  use_soft_bounds: true     # Multiplicative STDP

Winner-Take-All​

wta:
  mode: "both"             # "global", "local", "both"
  local_radius: 2
  enable_homeostasis: true
  target_rate: 0.1
  homeostasis_lr: 0.001
  threshold_min: 1.0
  threshold_max: 100.0

Homeostasis​

homeostasis:
  enabled: true
  theta_rest: 0.1
  theta_plus: 0.02
  tau_theta: 500.0
  theta_max: 10.0
  dead_neuron_recovery: true
  dead_threshold: 0.01
  recovery_boost: 0.1

Data​

data:
  dataset: "ebssa"
  data_root: "/path/to/EBSSA"
  sensor: "all"             # "ATIS", "DAVIS", "all"
  batch_size: 1
  num_workers: 0
  n_timesteps: 10
  input_height: 128
  input_width: 128
  input_channels: 2         # ON/OFF polarity
  normalize: true
  train_ratio: 0.9
  augmentation:
    enabled: true
    flip_horizontal: true
    flip_polarity: false

Convergence​

convergence:
  min_wins_per_neuron: 10
  target_ratio: 0.90
  patience: 20
  delta_threshold: 0.0001
  check_interval: 100

Checkpointing​

checkpoint:
  save_dir: "checkpoints"
  save_interval: 1
  keep_last_n: 3
  save_best: true
  save_on_interrupt: true

Logging​

logging:
  log_dir: "logs"
  log_level: "INFO"
  log_interval: 100
  tensorboard: true
  wandb: false
  wandb_project: "spikeseg"