Configuration Options

KKT-HardNet accepts training options either when constructing KKTHardNet(name=..., train=...) or when calling model(), optimize(), or estimate().

Parameter	Description	Default Value	Required
epochs	Number of training epochs.	1200	No
batch_size	Number of samples per training batch.	32	No
learning_rate	Learning rate for Adam.	1e-3	No
hidden_size	Width of hidden MLP layers.	64	No
hidden_layers	Number of hidden MLP layers.	2	No
train_frac	Fraction of samples used for training.	0.8	No
seed	Random seed for splitting and initialization.	42	No
dtype	Numerical precision, usually "float64".	"float64"	No
print_every	Frequency of printed epoch logs.	1	No
drop_last	Whether to drop incomplete mini-batches.	False	No
eta	Optional loss threshold for switching from MLP-only to projection training.	None	No
epoch_mlp	Optional epoch at which projection training begins.	None	No
cons_alpha	Weight for the consistency loss between raw and projected predictions.	0.0	No

Projection settings can be supplied with the projection argument or nested training configuration.

Parameter	Description	Default Value
fb_eps	Fischer-Burmeister smoothing value.	1e-8
gn_max_iters	Maximum Gauss-Newton iterations in the projection solve.	30
gn_tol	Projection residual tolerance.	1e-6
gn_reg	Regularization in the normal-equation solve.	1e-3
newton_step_length	Initial line-search step length.	0.5
armijo_alpha	Armijo sufficient decrease parameter.	1e-4
armijo_beta	Backtracking contraction factor.	0.5
max_backtrack_iter	Maximum line-search backtracking steps.	10
backward_reg	Regularization used by the custom VJP backward solve.	1e-8

Example

from kkthn import KKTHardNet

TRAIN = {
    "epochs": 1000,
    "batch_size": 32,
    "learning_rate": 1e-3,
    "train_frac": 0.8,
    "hidden_size": 64,
    "hidden_layers": 2,
    "seed": 42,
    "dtype": "float64",
    "print_every": 50,
    "cons_alpha": 1.0,
}

PROJECTION = {
    "gn_max_iters": 30,
    "gn_tol": 1e-6,
    "gn_reg": 1e-3,
    "newton_step_length": 0.5,
}

model = KKTHardNet(name="Example_Model", train=TRAIN)

The output directory is named:

<model_name>_<YYYYMMDD>_<HHMMSS>

For example:

• Example_Model_20260505_104512

• kkthardnet_20260505_104512