Student guide

Guide for students participating in distributed GAN training as workers.

Your role

As a worker, you:

  • Contribute GPU/CPU processing power

  • Compute gradients on assigned image batches

  • Help train the shared GAN model

  • Learn about distributed systems and GANs

You don’t need to understand every detail - the worker handles most complexity automatically.

Getting started

1. Choose your setup path

Pick what works best for you:

2. Get credentials

Your instructor will provide:

  • Database host address

  • Database name

  • Your username

  • Your password

Keep these secure - don’t share publicly.

3. Configure and run

After installing dependencies:

# Copy template
cp config.yaml.template config.yaml

# Edit with your credentials
nano config.yaml  # or use any text editor

# Start contributing!
python src/worker.py

4. Set your name (optional)

Add your name to config.yaml so you appear on the dashboard leaderboard:

worker:
  name: Alice       # Your name here!
  batch_size: 32    # Increase to 64 if you have more GPU memory
  poll_interval: 5

This makes it easy to track your contributions and compete with classmates!

What the worker does

Automatic workflow

The worker runs in a loop:

  1. Poll database - Check for available work units

  2. Claim work - Atomically claim an unclaimed batch

  3. Load weights - Get latest model weights

  4. Compute gradients - Process assigned images

  5. Upload results - Send gradients to database

  6. Repeat - Continue until training completes

You just start it and let it run!

What you’ll see

Initializing worker...
Loaded dataset with 202599 images
Worker abc123 initialized successfully!
Name: Alice
GPU: NVIDIA GeForce RTX 3080
Batch size: 32
CPU cores: 8
RAM: 16.0 GB
GPU VRAM: 10.0 GB
Waiting for work units...

Processing work unit 42 (iteration 5)
Number of images: 320
Completed 10 batches (320 images)
G_loss: 2.345 | D_loss: 1.234 | D_real: 85.00% | D_fake: 80.00%
Extracting gradients...
Uploading gradients...
Work unit 42 completed successfully!

Processing work unit 43 (iteration 5)
...

Understanding the output

Initialization

Worker abc123 initialized successfully!
Name: Alice
GPU: NVIDIA GeForce RTX 3080
Batch size: 32
CPU cores: 8
RAM: 16.0 GB
Dataset size: 202599

  • Worker ID: Unique identifier for your worker

  • Name: Your name (from config, or hostname if not set)

  • GPU: Your hardware (or “CPU” if no GPU)

  • Batch size: Images per training batch (from your config)

  • System info: CPU cores, RAM, GPU VRAM (shown on dashboard)

During training

Processing work unit 42 (iteration 5)
Number of images: 320
Completed 10 batches (320 images)
G_loss: 2.345 | D_loss: 1.234 | D_real: 85.00% | D_fake: 80.00%
Work unit 42 completed successfully!

  • Work unit: Unique batch assignment

  • Iteration: Current training iteration

  • Images: How many images this work unit contains

  • Losses: Training loss values (lower is generally better)

  • Accuracy: How well discriminator distinguishes real/fake

Monitoring your contribution

In the console

The worker prints:

  • Number of work units completed

  • Total images processed

  • Processing time per unit

Database queries

Your instructor can show your stats:

SELECT worker_id, total_images, total_work_units, last_heartbeat
FROM workers
WHERE worker_id = 'YOUR_WORKER_ID';

Leaderboard (if available)

Your instructor may set up a dashboard showing:

  • Top contributors by name

  • Total work units processed

  • Active workers

Run it yourself with streamlit run src/dashboard.py

Best practices

Maximize contribution

  • Let it run - Keep worker running as long as possible

  • Stable connection - Ensure reliable internet

  • Avoid interruptions - Close unnecessary applications

  • Monitor occasionally - Check it hasn’t crashed

Resource management

  • GPU memory - Close other GPU applications

  • CPU usage - Worker uses one CPU core mostly

  • Disk space - Needs ~10GB for dataset

  • Network - Upload/download gradients and weights

When to stop

You can stop anytime:

  • Press Ctrl+C in terminal

  • Worker will finish current work unit gracefully

  • No data loss - training state is in database

Troubleshooting

No work units available

Problem: Worker keeps polling but finds no work

Solutions:

  • Training may not have started yet

  • Wait for coordinator to create work units

  • Check with instructor

Connection errors

Problem: Can’t connect to database

Solutions:

  • Verify credentials in config.yaml

  • Check database host is accessible

  • Test connection: ping DATABASE_HOST

  • Contact instructor

Out of memory

Problem: GPU runs out of memory

Solutions:

  • Close other GPU applications

  • Modify config.yaml: reduce batch_size

  • Try CPU-only mode

Worker crashes

Problem: Worker stops unexpectedly

Solutions:

  • Check error message

  • Verify dataset downloaded completely

  • Try reducing batch size

  • Restart worker - it will resume automatically

Slow performance

Problem: Processing work units very slowly

Solutions:

  • Check GPU utilization: nvidia-smi

  • Verify using GPU not CPU

  • Close background applications

  • Check network speed

FAQ

Q: How long should I run the worker?
A: As long as you can! Even 30 minutes helps. Ideally several hours.

Q: Will this harm my GPU?
A: No. GPUs are designed for this. Monitor temperature if concerned (should stay under 85°C).

Q: Can I use my computer while the worker runs?
A: Yes, but close other GPU applications. CPU work is fine.

Q: What if I need to stop?
A: Just press Ctrl+C. Worker stops gracefully. You can restart anytime.

Q: Do I get credit for contribution?
A: Your instructor tracks contributions. Check your course requirements.

Q: Can I run multiple workers?
A: Yes, if you have multiple GPUs. Each needs separate config.

Q: What if training finishes?
A: Worker will detect completion and stop. Check with instructor.

Learning outcomes

By participating as a worker, you learn:

Distributed systems:

  • How workers coordinate without direct communication

  • Database as message queue

  • Atomic operations and race conditions

  • Fault tolerance

Deep learning:

  • GAN architecture

  • Gradient computation

  • Data parallel training

  • The role of batch processing

Practical skills:

  • Environment setup

  • Configuration management

  • Monitoring processes

  • Debugging distributed applications

Next steps