Skip to content

weights.md

Latest commit

 

History

History
61 lines (44 loc) · 2 KB

weights.md

File metadata and controls

61 lines (44 loc) · 2 KB
Raw

Reference Weights

Reference checkpoints are distributed as GitHub Release assets instead of being committed to git.

Stage 4 Release Checkpoint

Field Value
Stage 4
Model Reset-based sparse SNN
Dataset MNIST
Hidden dim 128
Time steps 4
Epochs 100
Seed 35
Best test accuracy 97.75%
Download stage4_mnist_hidden128_seed35_v0.1.1.pt
Metrics CSV stage4_mnist_hidden128_seed35_v0.1.1_metrics.csv
SHA256 78683332db3c7b1f0c60b5a434e07020daa923c1d0f8ec5579cfd84d153e1d72

This checkpoint was trained with:

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

It is a standalone Stage 4 reference checkpoint. The Stage 0-4 comparison table in the README is produced by examples/compare_stages.py, while this checkpoint is produced by train.py with checkpointing enabled.

Because the comparison table reports the final epoch from a five-stage benchmark script and this checkpoint records the best Stage 4 epoch from a standalone training run, the reported Stage 4 numbers are expected to be close but not identical.

The checkpoint is saved by train.py as a dictionary with:

  • model_state_dict
  • best_accuracy
  • config

Load it with:

import torch

from snn_structural_evolution import get_model

checkpoint = torch.load("stage4_mnist_hidden128_seed35_v0.1.1.pt", map_location="cpu")
model = get_model(stage=4, hidden_dim=128, time_steps=4)
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()
print(checkpoint["best_accuracy"])

Recreate the checkpoint from scratch:

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

Verify the downloaded checkpoint:

Get-FileHash -Algorithm SHA256 stage4_mnist_hidden128_seed35_v0.1.1.pt