[ PROMPT_NODE_22698 ]
sweeps
[ SKILL_DOCUMENTATION ]
# Comprehensive Hyperparameter Sweeps Guide
Complete guide to hyperparameter optimization with W&B Sweeps.
## Table of Contents
- Sweep Configuration
- Search Strategies
- Parameter Distributions
- Early Termination
- Parallel Execution
- Advanced Patterns
- Real-World Examples
## Sweep Configuration
### Basic Sweep Config
python
sweep_config = {
'method': 'bayes', # Search strategy
'metric': {
'name': 'val/accuracy',
'goal': 'maximize' # or 'minimize'
},
'parameters': {
'learning_rate': {
'distribution': 'log_uniform',
'min': 1e-5,
'max': 1e-1
},
'batch_size': {
'values': [16, 32, 64, 128]
}
}
}
# Initialize sweep
sweep_id = wandb.sweep(sweep_config, project="my-project")
### Complete Config Example
python
sweep_config = {
# Required: Search method
'method': 'bayes',
# Required: Optimization metric
'metric': {
'name': 'val/f1_score',
'goal': 'maximize'
},
# Required: Parameters to search
'parameters': {
# Continuous parameter
'learning_rate': {
'distribution': 'log_uniform',
'min': 1e-5,
'max': 1e-1
},
# Discrete values
'batch_size': {
'values': [16, 32, 64, 128]
},
# Categorical
'optimizer': {
'values': ['adam', 'sgd', 'rmsprop', 'adamw']
},
# Uniform distribution
'dropout': {
'distribution': 'uniform',
'min': 0.1,
'max': 0.5
},
# Integer range
'num_layers': {
'distribution': 'int_uniform',
'min': 2,
'max': 10
},
# Fixed value (constant across runs)
'epochs': {
'value': 50
}
},
# Optional: Early termination
'early_terminate': {
'type': 'hyperband',
'min_iter': 5,
's': 2,
'eta': 3,
'max_iter': 27
}
}
## Search Strategies
### 1. Grid Search
Exhaustively search all combinations.
python
sweep_config = {
'method': 'grid',
'parameters': {
'learning_rate': {
'values': [0.001, 0.01, 0.1]
},
'batch_size': {
'values': [16, 32, 64]
},
'optimizer': {
'values': ['adam', 'sgd']
}
}
}
# Total runs: 3 × 3 × 2 = 18 runs
**Pros:**
- Compre
粤公网安备44030002003366号