Configuration¶

Complete reference for all configuration options in DeepAugment.

DeepAugment Initialization¶

from deepaugment import DeepAugment

aug = DeepAugment(
    X_train, y_train, X_val, y_val,
    # Phase 1: Essential
    model="simple",
    device="auto",
    random_state=42,
    experiment_name=None,
    # Phase 2: Useful
    method="bayesian",
    save_history=True,
    # Phase 3: Advanced
    transform_categories=None,
    custom_reward_fn=None,
    resume_from=None,
    # Core parameters
    n_operations=4,
    train_size=2000,
    val_size=500,
)

Essential Parameters¶

modelstr or nn.Module class, default=”simple”

Model architecture to use for evaluation.

"simple": SimpleCNN (default, 1.2M parameters, fast)
Custom PyTorch model class (see Advanced Usage)

devicestr, default=”auto”

Device for training.

"auto": Automatically select best device (CUDA > MPS > CPU)
"cuda": Force NVIDIA GPU
"mps": Force Apple Silicon GPU
"cpu": Force CPU

random_stateint or None, default=None

Random seed for reproducibility. If None, uses default from config (42).

experiment_namestr or None, default=None

Name for the experiment. If None, generates timestamp-based name.

Useful Parameters¶

methodstr, default=”bayesian”

Search strategy.

"bayesian": Bayesian Optimization (recommended)
"random": Random search (baseline)

save_historybool, default=True

Whether to save optimization history to disk. Saved in experiments/ directory.

Advanced Parameters¶

transform_categorieslist or None, default=None

Filter transforms by category. If None, uses all transforms.

Available categories:

"geometric": rotate, flip, affine, shear, perspective, elastic, random_crop
"color": brightness, contrast, saturation, hue, color_jitter
"advanced_color": sharpen, autocontrast, equalize, invert, solarize, posterize, grayscale
"blur_noise": blur, gaussian_noise
"occlusion": erasing, cutout
"advanced": channel_permute, photometric_distort

Example:

transform_categories=["geometric", "color"]

custom_reward_fncallable or None, default=None

Custom reward function for optimization. Function signature:

def my_reward(entry: dict) -> float:
    """
    Args:
        entry: Dict with keys:
            - 'policy': List of (transform, magnitude) tuples
            - 'score': Validation accuracy
            - 'iteration': Current iteration number

    Returns:
        Custom reward value
    """
    return entry['score']  # Default behavior

resume_fromstr or None, default=None

Path to checkpoint file to resume optimization from.

Core Parameters¶

n_operationsint, default=4: Number of augmentation operations per policy. Higher values create more complex policies but increase search space.
train_sizeint, default=2000: Number of training samples to use for optimization. Using a subset speeds up optimization.
val_sizeint, default=500: Number of validation samples to use for evaluation.

Optimize Method¶

best_policy = aug.optimize(
    iterations=50,
    epochs=10,
    samples=1,
    batch_size=64,
    learning_rate=0.001,
    early_stopping=False,
    patience=10,
    verbose=True,
)

Parameters¶

iterationsint, default=50

Number of policies to try. More iterations generally lead to better policies but take longer.

Quick experiments: 25-50
Standard: 50-100
Thorough: 100-300

epochsint, default=10

Training epochs per policy evaluation. More epochs give more accurate evaluation but are slower.

Quick: 5-10
Standard: 10-15
Thorough: 15-20

samplesint, default=1

Number of training runs per policy. Results are averaged to reduce noise.

Fast: 1
Stable: 3
Very stable: 5

batch_sizeint, default=64

Training batch size.

learning_ratefloat, default=0.001

Learning rate for training.

early_stoppingbool, default=False

Stop optimization if no improvement is seen.

patienceint, default=10

Number of iterations without improvement before stopping (only if early_stopping=True).

verbosebool, default=True

Show progress bar and updates.

Default Configuration¶

DeepAugment uses sensible defaults from deepaugment.config.defaults:

from deepaugment.config import defaults

print(f"Default iterations: {defaults.iterations}")
print(f"Default epochs: {defaults.epochs}")
print(f"Default n_operations: {defaults.n_operations}")

# View all defaults
print(defaults)

These defaults are optimized for:

Balance between speed and quality
Quick iteration during development
Reasonable computational cost

You can override any default by passing it explicitly to DeepAugment() or optimize().

Environment Variables¶

DEEPAUGMENT_EXPERIMENTS_DIR

Override default experiments directory. Default: ./experiments

export DEEPAUGMENT_EXPERIMENTS_DIR=/path/to/experiments

DEEPAUGMENT_DEVICE

Override default device selection. Values: cuda, mps, cpu

export DEEPAUGMENT_DEVICE=cuda

Examples¶

Quick Experiment¶

aug = DeepAugment(
    X_train, y_train, X_val, y_val,
    train_size=500,
    val_size=100,
)
best = aug.optimize(iterations=25, epochs=5)

Thorough Optimization¶

aug = DeepAugment(
    X_train, y_train, X_val, y_val,
    train_size=5000,
    val_size=1000,
    save_history=True,
    experiment_name="thorough_run"
)
best = aug.optimize(
    iterations=200,
    epochs=20,
    samples=3,
    early_stopping=True,
    patience=20
)

Reproducible Run¶

aug = DeepAugment(
    X_train, y_train, X_val, y_val,
    random_state=42,
    device="cuda",
)
best = aug.optimize(iterations=50, epochs=10)
# Same results every time with same seed

Custom Configuration¶

aug = DeepAugment(
    X_train, y_train, X_val, y_val,
    model=MyCustomModel,
    transform_categories=["geometric", "color"],
    custom_reward_fn=my_reward,
    n_operations=6,
    method="bayesian",
    save_history=True,
)
best = aug.optimize(
    iterations=100,
    epochs=15,
    samples=2,
    batch_size=128,
    learning_rate=0.002,
)