📱 HyPockeTuner

Bringing Hyperparameter Optimization to Mobile Devices

A mobile system for monitoring, steering, and reflecting on HPO experiments from smartphones.

HyPockeTuner supports long-running HPO experiments with a notification-driven workflow: Configure → Optimize → Analyze → Refine.

---

✨ Key Features

• EventCrumb Visualization

  • A mobile-optimized event sequence visualization for tracking experiment history.
  • Shows completed trials with performance metrics, bracket progression, and user intervention points.

• Notification-Driven Workflow

  • Subscribe to events: performance improvements, round completion, exceptions.
  • Get notified instead of constant monitoring.

• Interactive Refinement Operations

  • Narrow the search space while preserving learned weights.
  • Redefine the space to launch independent experiments with adjusted configurations.
  • Inject User Trials to test hypotheses directly in the optimization loop.

---

🔧 Supported Kernels

Kernel	Dataset	Task	Key Hyperparameters
MNIST	MNIST	Image Classification	lr, batch_size, optimizer, hidden_dim
CIFAR-10	CIFAR-10	Image Classification	lr, architecture, dropout, weight_decay
NLP	GLUE / K-MHaS	Text Classification	lr, warmup_steps, model_type, max_seq_len
Vision-Language	MS-COCO	Image-Text Retrieval	lr, batch_size, encoder, embedding_dim

---

🔌 Add Your Own Kernel

Integrate your model training code via the Python API to use with the BOHB optimizer.

See server/kernels/ for template examples.

---

⚡ Quick Start

Backend

# 1. Navigate to server
cd server

# 2. Install dependencies
pip install -r requirements.txt

# 3. Run the API server (Terminal 1)
python main.py --host 0.0.0.0 --port 8080

# 4. Run the worker process (Terminal 2)
python worker.py

Note: Both main.py (API server) and worker.py (trial executor) must be running simultaneously.

Frontend

# 1. Navigate to client
cd client

# 2. Install dependencies
npm install

# 3. Launch the UI
npm run dev

---

⚙️ Configuration

Push Notifications Setup

To enable push notifications on your smartphone:

1. Generate VAPID Keys

npx web-push generate-vapid-keys

This will output:

Public Key: BEl62iUYgU...
Private Key: UUxI4oKwbA...

2. Configure Server

# server/.env
VAPID_PUBLIC_KEY=BEl62iUYgU...      # Public key from step 1
VAPID_PRIVATE_KEY=UUxI4oKwbA...     # Private key from step 1
VAPID_CLAIMS_EMAIL=your@email.com

3. Configure Client

# client/.env.local
VITE_VAPID_PUBLIC_KEY=BEl62iUYgU... # Same public key from step 1

Note: Push notifications require HTTPS in production. For local development, notifications are simulated on localhost.

BOHB Parameters

Configure the optimization in experiment settings:

Parameter	Description	Example
η (eta)	Dividing factor for successive halving	3
b_min	Minimum budget (epochs)	1
b_max	Maximum budget (epochs)	81

Notification Conditions

Type	Description
Progress	Alert when brackets or rounds complete
Performance	Alert when accuracy reaches target or improves by Δ
Timeout	Alert when no improvement within time window
Emergency	Alert for high GPU temp, low utilization, or exceptions

---

📊 System Architecture

┌──────────────────┐     Socket.IO     ┌──────────────────┐
│  React Frontend  │◄─────────────────►│  FastAPI Backend │
│                  │                   │                  │
│  • EventCrumb    │                   │  • BOHB Engine   │
│  • Zustand Store │                   │  • ML Kernels    │
│  • Push Notif.   │                   │  • GPU Monitor   │
└──────────────────┘                   └──────────────────┘

---

🔗 Related Resources

• BOHB Paper → Robust and Efficient Hyperparameter Optimization at Scale
• ATMSeer → Increasing Transparency and Controllability in AutoML
• HyperTendril → User-driven Hyperparameter Optimization

---

📚 Citation

TBU

---

👥 Authors

Name	Affiliation	Email
Donghee Hong	Sungkyunkwan University	dh.hong@skku.edu
Bongshin Lee	Yonsei University	b.lee@yonsei.ac.kr
Jinwook Seo	Seoul National University	jseo@snu.ac.kr
Jaemin Jo*	Sungkyunkwan University	jmjo@skku.edu

* Corresponding author

---

🙏 Acknowledgments

This work was supported by:

• National Research Foundation of Korea (NRF) - RS-2025-24873100, RS-2023-NR077081
• Institute of Information and Communications Technology Planning and Evaluation (IITP) - AI Graduate School Program at SKKU, Yonsei, and SNU

---