vyntri.config

`Config`

Validated configuration with describe(), JSON export/import.

Field	Default	Description
backbone	"mobilenet_v3_small"	Feature extractor backbone (mobilenet_v3_small, resnet18, resnet50, efficientnet_b0, convnext_tiny, or auto)
adaptation	"fk"	Adaptation method: "fk" (discriminative whitening), "slce" (centroid encoder), or "none"
candidates	None	List of backbone names to evaluate when backbone="auto". None = evaluate all supported backbones.

Field	Default	Description
shrinkage	"diagonal"	Covariance regularization: "diagonal", "ledoit_wolf", or "none"
shrinkage_alpha	0.5	Diagonal shrinkage intensity
regularization	1e-4	Ridge regularization strength
projection_dim	None	Projection dimension (None auto-resolves to min(128, feature_dim, C-1))

Field	Default	Description
seed	42	Random seed
device	"auto"	"auto", "cpu", or "cuda"
dtype	"float32"	Feature storage dtype (solves always use float64)
batch_size	32	Feature extraction batch size
num_workers	0	DataLoader workers for extraction (0 = main process)
input_size	224	Center-crop size fed to the backbone

Field	Default	Description
cache_enabled	True	Enable disk caching of extracted features
cache_dir	".vyntri_cache"	Directory for the feature cache

Field	Default	Description
eigenvalue_floor	None	Absolute eigenvalue floor for FK inverse square root (None = auto)
floor_ratio	1e-4	Relative eigenvalue floor for FK when eigenvalue_floor is unset

Returned by model.predict(image).

Returned by model.predict_batch(images).

Returned by model.evaluate(test_dataset).

Field	Type	Description
accuracy	float	Overall accuracy
macro_f1	float	Macro-averaged F1 score
weighted_f1	float	Weighted F1 score
balanced_accuracy	float	Balanced accuracy (recall per class averaged)
confusion_matrix	np.ndarray	Confusion matrix (rows=true, cols=pred)
class_names	list[str]	Class names in order
per_class	dict	Per-class precision, recall, f1, support
n_samples	int	Total samples evaluated
timings	dict	Timing breakdown

Returned by model.update(dataset).

Field	Type	Description
classes_before	list[str]	Classes before update
classes_after	list[str]	Classes after update
new_classes	list[str]	Newly added classes (empty if only adding data to existing classes)
n_samples	int	Samples added
samples_per_class	dict	Samples added per class
in_sample_accuracy	float \| None	Accuracy on the update batch itself (not held-out)

Returned by model.analyze(dataset).

Returned by model.select_backbone(dataset).

Field	Type	Description
candidates	list[str]	All candidate backbone names
scores	dict	LogME scores per candidate
ranked	list[str]	Candidates ranked by score
validated	list[str]	Top-k candidates validated
validation_accuracies	dict	Validation accuracy per validated candidate
selected	str	Winning backbone name
selected_accuracy	float	Winning validation accuracy
timings	dict	Timing breakdown

Returned by model.fine_tune(...).

Field	Type	Description
scope	str	Scope used ("last_layer", "last_block", or "full")
epochs	int	Number of epochs requested
gradient_steps	int	Total gradient steps taken
trainable_parameters	int	Number of trainable parameters
optimizer	str	Optimizer used ("adam")
learning_rate	float	Learning rate used
weight_decay	float	Weight decay used
best_validation_accuracy	float \| None	Best validation accuracy during training
history	list[dict]	Per-epoch training history
n_train	int	Training samples used
n_val	int	Validation samples used
timings	dict	Timing breakdown