WANDB_SETUP.md

═══════════════════════════════════════════════════════════════════════

WandB Configuration for Medical VQA Training Monitoring

═══════════════════════════════════════════════════════════════════════

QUICK START:

1. Create WandB Account

Go to: https://wandb.ai/ Sign up with GitHub or Email

2. Get API Key

Go to: https://wandb.ai/settings/profile Copy your API key

3. Set Environment Variable

export WANDB_API_KEY="your_api_key_here"

Or in Jupyter:

import os os.environ['WANDB_API_KEY'] = 'your_api_key_here'

4. Run Training

python train_medical.py --variant A1

Automatically logs to WandB!

WHAT GETS LOGGED:

✅ Training Metrics (per epoch):

• train_loss
• train_accuracy
• train_bleu
• train_rouge
• train_bertscore

✅ Validation Metrics (per epoch):

• val_loss
• val_accuracy
• val_bleu
• val_rouge
• val_bertscore

✅ Model Info:

• Number of parameters
• Model architecture
• Config settings

✅ Hardware:

• GPU usage
• Memory
• Training time

✅ Learning Rate:

• Current LR per epoch
• Warmup schedule

MONITORING DASHBOARD:

View live at: https://wandb.ai/QuangVoAI/MedicalVQA-Vietnam

Features:

• Real-time loss graphs
• Metric comparison across variants
• Training progress
• System resource monitoring
• Hyperparameter tracking
• Model checkpoints

ADVANCED:

Save Checkpoints to WandB: wandb.save('checkpoint.pt')

Log Custom Metrics: wandb.log({'custom_metric': value, 'epoch': epoch})

Compare Models: Visit: https://wandb.ai/QuangVoAI/MedicalVQA-Vietnam/reports

OFFLINE MODE:

If you don't have internet: export WANDB_MODE=offline python train_medical.py --variant A1

Saves locally, can sync later

TIPS:

1. Set descriptive run names: wandb.init(..., name="A2_50epochs_final")

2. Add tags for easy filtering: wandb.init(..., tags=["production", "50-epochs"])

3. Create reports with charts: Use WandB UI to create custom reports

4. Compare multiple runs: Group runs by config/variant

═══════════════════════════════════════════════════════════════════════