🚀 Feature

Add automatic clipping support to Opacus with DPAutomaticClippingOptimizer and DPPerLayerAutomaticClippingOptimizer classes that adaptively scale gradients without manual threshold tuning.

Motivation

Manual tuning of clipping threshold is time-consuming. Different layers may need different clipping thresholds, but current Opacus requires either a global threshold or manual per-layer configuration. This can result in poor model utility.

Pitch

Implement automatic clipping based on Automatic Clipping: Differentially Private Deep Learning Made Easier and Stronger:

Formula: clip_factor = max_grad_norm / (per_sample_norms + 0.01) instead of min(1.0, max_grad_norm / per_sample_norms)

Usage:

privacy_engine = PrivacyEngine()
model, optimizer, dataloader = privacy_engine.make_private(
    module=model,
    optimizer=optimizer,
    data_loader=dataloader,
    noise_multiplier=1.0,
    max_grad_norm=1.0,
    clipping="automatic" or "automatic_per_layer",  # New options
)

• Eliminates hyperparameter tuning
• Adaptive per-sample clipping improves utility
• Maintains same privacy guarantees
• Easy API integration by following the docs

Additional context

I'm happy to submit a PR for this feature if it aligns with the project's goals. I'm implementing automatic clipping as part of my thesis work, and I believe making it available in the main library could benefit the broader Opacus community.