p-j-r-1-2-3/N-Caltech101-SNN-models

N-Caltech101 Event-Driven SNN Models

UWU Research Project

Team: Ramyanath, Suranjaya, Weerakon Supervisor: Dr. K.P.P.S. Pathirana Institution: Uva Wellassa University

Training Results

Best Model Performance

• —Best Test Accuracy: 38.86%
• —Final Test Accuracy: 38.46%
• —Average Inference Time: 250.30ms/batch
• —Training Time: 0.32 hours
• —Platform: Kaggle T4 GPU

Model Architecture

• —Type: 3-layer Fully-Connected SNN (Baseline)
• —Parameters: 720,869
• —Hidden Size: 128
• —Beta (LIF): 0.9
• —Temporal Steps: 150

Training Configuration

• —Batch Size: 64
• —Epochs: 8
• —Learning Rate: 0.0005
• —Weight Decay: 0.0001 (L2 regularization)
• —Label Smoothing: 0.1
• —Optimizer: Adam with L2 regularization
• —Scheduler: ReduceLROnPlateau (patience=5)
• —Mixed Precision: True

Anti-Overfitting Techniques Applied

• —L2 Weight Regularization (weight_decay=0.0001)
• —Label Smoothing (0.1)
• —Lower Learning Rate (0.0005)
• —Adaptive LR Scheduler (reduces LR when validation plateaus)
• —Early Stopping (if train-test gap > 25%)
• —No Dropout (as requested - using weight decay instead)

Event-Driven Processing

• —Event Timestep: 2000µs (2.0ms)
• —Spatial Resolution: 60x45
• —Processing: Vectorized (100x faster than loop-based)
• —Caching: Enabled (preprocessed tensors)

Files in This Repository

• —checkpoints/best_model.pth - Best performing model
• —checkpoints/last_model.pth - Most recent checkpoint
• —results/results_kaggle.json - Training metrics
• —logs/training_history.json - Complete training history
• —results/training_plots.png - Accuracy/loss curves

Usage

import torch
from model import BaselineEventSNN

# Load checkpoint
checkpoint = torch.load('checkpoints/best_model.pth')

# Initialize model
model = BaselineEventSNN(
    input_size=5400,
    hidden_size=128,
    output_size=101,
    beta=0.9
)

# Load weights
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

Deployment Target

NVIDIA Jetson Nano (ARM Architecture) This baseline model is designed for edge deployment on neuromorphic hardware.

Last updated: 2026-01-07 11:59:59