Holistic Robust DRO

Training customization

HRNNDRO.fit(...) accepts paired train_loader and val_loader arguments, a PyTorch optimizer, and an optional scheduler. Select scheduler_step="epoch" (default) or "batch". Build the optimizer after calling update_user_mode(...); that call also rebinds the adversarial attack to the replacement network. The loader batch size can differ from train_batch_size: HR-DRO rebuilds its weighting problem when the size changes. See the shared training rules for validation and defaults.

class dro.neural_model.hrdro_nn.HRNNDRO(input_dim, num_classes, task_type='classification', model_type='mlp', alpha=0.1, r=0.01, epsilon=0.1, learning_approach='HD', adversarial_params=None, train_batch_size=64, device=device(type='cpu'))

Bases: BaseNNDRO

Holistic Robust DRO.

Initialize Holistic Robust DRO.

Parameters:

  • input_dim (int) – Input feature dimension \(d \geq 1\)

  • num_classes (int) –

    Output dimension:

    • Classification: \(K \geq 2\) (number of classes)

    • Regression: Automatically set to 1

  • task_type (str) –

    Learning task type. Supported:

    • 'classification': Cross-entropy loss

    • 'regression': MSE loss

  • model_type (str) –

    Neural architecture type. Supported:

    • 'mlp': Multi-Layer Perceptron (default)

    • linear

    • resnet

    • alexnet

  • device (torch.device) – Target computation device, defaults to CPU

  • alpha (float) – Robustness regularization coefficient \(\alpha > 0\), controls model conservativeness. Defaults to 0.1.

  • r (float) – Statistical error radius \(r > 0\) for distribution shift, defaults to 0.01.

  • epsilon (float) – Contamination ratio \(\epsilon \in [0,1)\), proportion of adversarial samples. Defaults to 0.1.

  • learning_approach (str) –

    Robust optimization method. Options:

    • "HR": Huberian Robust optimization

    • "HD": Hybrid DRO (default)

  • adversarial_params (dict) – Adversarial training configuration. norm is 'l2' or 'l-inf'; method is 'PGD', 'FGSM', or 'FFGSM'. FGSM and FFGSM require 'l-inf'. Torchattacks clips attack inputs to [0, 1]. Defaults to {"steps":10, "step_size":0.02, "norm":"l2", "method":"PGD"}.

  • train_batch_size (int) – Training batch size \(B \geq 1\), defaults to 64.

Raises:

ValueError –

  • If alpha ≤ 0

  • If r ≤ 0

  • If ε ∉ [0,1)

  • If invalid learning_approach

Example (CIFAR-10)::
>>> adv_params = {"steps":5, "step_size":0.03, "norm":"l2", "method":"PGD"}
>>> model = HRNNDRO(
...     input_dim=3072, 
...     num_classes=10,
...     alpha=0.5,
...     epsilon=0.2,
...     adversarial_params=adv_params
... )
update_user_mode(input_dim, num_classes, model, task_type='classification', device=device(type='cpu'))

Replace the network and bind both attacks to the new network.

Return type:

None

Parameters:
fit(X=None, y=None, train_ratio=0.8, lr=0.001, batch_size=None, epochs=100, verbose=True, *, train_loader=None, val_loader=None, optimizer=None, scheduler=None, scheduler_step='epoch')

Fit HR-DRO using arrays or caller-supplied training components.

Return type:

Dict[str, float]

Parameters:
predict(X)

Adversarial-robust prediction

Return type:

ndarray

Parameters:

X (ndarray | Tensor)