circadia-sleep-student

Conv1dStack_T2_a0.3 — Distilled Sleep Stage Classifier

A tiny (103KB, 25,957 params) sleep stage classifier distilled from SleepFM for real-time edge deployment on NVIDIA Jetson TK1 and similar constrained devices.

Model Details

Student Architecture Config

Conv1dStack:
  hidden_channels: 32
  kernel_size: 5

Distillation Setup

• Teacher: SleepFM (SleepEventLSTMClassifier) — biLSTM, 128-dim embeddings
• Sweep: 3 temperatures × 3 alphas × 3 architectures = 27 experiments
• Training: 50 epochs, AdamW lr=0.001, early stopping (patience=10)
• Data: Sleep-EDF (5 train / 1 val / 1 test subjects)

Target Hardware

Usage

ONNX Runtime

import numpy as np
import onnxruntime as ort

session = ort.InferenceSession("Conv1dStack_T2_a0.3.onnx")

# Input: pre-pooled embeddings (batch, seq_len, 128)
embeddings = np.random.randn(1, 120, 128).astype(np.float32)
logits = session.run(None, {"input": embeddings})[0]
predicted_stages = np.argmax(logits, axis=-1)

# Stage mapping: 0=Wake, 1=REM, 2=N1, 3=N2, 4=N3
print(predicted_stages)

TorchScript

import torch

model = torch.jit.load("Conv1dStack_T2_a0.3.pt")
embeddings = torch.randn(1, 120, 128)
logits = model(embeddings)
predicted_stages = logits.argmax(dim=-1)

Files

Limitations

• Trained on Sleep-EDF only (7 subjects) — may not generalize to other PSG datasets
• Expects pre-pooled 128-dim embeddings from SleepFM's encoder, not raw EEG
• No per-class metrics reported (overall accuracy only)
• Distilled from a single teacher checkpoint

Citation

@misc{circadia-distill-2026,
  title={Distilled Sleep Stage Classifier for Edge Deployment},
  year={2026},
  url={https://github.com/circadia}
}