dataeval.shift.OODReconstruction¶

class dataeval.shift.OODReconstruction(model, device=None, model_type='auto', use_gmm=None, config=None)¶

Autoencoder (AE) or Variational Autoencoder (VAE) based out-of-distribution detector.

Supports standard autoencoders and variational autoencoders with optional Gaussian Mixture Model (GMM) in the latent space for enhanced detection. Model type can be auto-detected from the model structure or explicitly specified.

Parameters:¶

model : torch.nn.Module¶: An autoencoder or VAE model to use for encoding and reconstruction of images for detection of out-of-distribution samples. Model type will be auto-detected if not specified.
device : DeviceLike or None, default None¶: The hardware device to use if specified, otherwise uses the DataEval default or torch default.
model_type : {"ae", "vae", "auto"} or None, default "auto"¶: Type of model: “ae” for standard autoencoder, “vae” for variational autoencoder, or “auto” to auto-detect based on model structure. If None, defaults to “auto”.
use_gmm : bool or None, default None¶: Whether to use Gaussian Mixture Model in the latent space for enhanced OOD detection. If None, will be auto-detected based on whether model has gmm_density_net attribute. When True, the model must output (reconstruction, z, gamma) where z is the latent representation and gamma is the mixture assignment probabilities.
config : OODReconstruction.Config or None, default None¶: Optional configuration object with default training parameters. Parameters specified in fit() will override these defaults.

Example

Auto-detection (recommended):

>>> from dataeval.utils.models import AE, VAE, GMMDensityNet
>>> from dataeval.shift import OODReconstruction
>>>
>>> train_data = torch.rand(10, 1, 28, 28)
>>>
>>> # Auto-detect AE
>>> ae = AE(input_shape=(1, 28, 28))
>>> ood = OODReconstruction(ae)  # Automatically detects as "ae", use_gmm=False
>>>
>>> # Auto-detect VAE
>>> vae = VAE(input_shape=(1, 28, 28))
>>> ood = OODReconstruction(vae)  # Automatically detects as "vae", use_gmm=False
>>>
>>> # Auto-detect GMM
>>> ae_gmm = AE(input_shape=(1, 28, 28), gmm_density_net=GMMDensityNet(latent_dim=256, n_gmm=3))
>>> ood = OODReconstruction(ae_gmm)  # Automatically detects as "ae", use_gmm=True

Using configuration:

>>> config = OODReconstruction.Config(epochs=10, batch_size=128, threshold_perc=99.0)
>>> ood = OODReconstruction(vae, config=config)
>>> ood.fit(train_data)  # Uses config defaults

Explicit specification:

>>> ood = OODReconstruction(vae, model_type="vae", use_gmm=False)
>>> ood.fit(train_data, threshold_perc=95, epochs=20)

fit(x_ref, threshold_perc=None, loss_fn=None, optimizer=None, epochs=None, batch_size=None)¶

Train the model and infer the threshold value.

Parameters specified here override defaults from the config object.

Parameters:¶

x_ref : ArrayLike¶: Training data.
threshold_perc : float or None, default None¶: Percentage of reference data that is normal (0-100). If None, uses config.threshold_perc.
loss_fn : ReconstructionLossFn | VAELossFn | Callable | None, default None¶: Loss function used for training. Can be: - A ReconstructionLossFn for AE (e.g., torch.nn.MSELoss()) - A VAELossFn for VAE (e.g., ELBOLoss(beta=1.0)) - Any callable with appropriate signature If None, uses config.loss_fn, or auto-selects based on model type: - For AE/GMM-AE: uses MSELoss() - For VAE: uses ELBOLoss()
optimizer : torch.optim.Optimizer | None, default None¶: Optimizer used for training. If None, uses config.optimizer or Adam with lr=0.001.
epochs : int or None, default None¶: Number of training epochs. If None, uses config.epochs (default 20).
batch_size : int or None, default None¶: Batch size used for training. If None, uses config.batch_size (default 64).

Examples

Using config defaults (recommended):

>>> from dataeval.shift import OODReconstruction
>>> from dataeval.utils.models import AE, VAE
>>> input_shape = (1, 28, 28)
>>> train_data = torch.rand(20, *input_shape)
>>> config = OODReconstruction.Config(epochs=10, threshold_perc=95)
>>> ood = OODReconstruction(AE(input_shape), config=config)
>>> ood.fit(train_data)  # Uses config defaults

Overriding specific parameters:

>>> ood = OODReconstruction(VAE(input_shape))
>>> ood.fit(train_data, epochs=20, batch_size=10)  # Override defaults

Using custom loss:

>>> config = OODReconstruction.Config(loss_fn=ELBOLoss(beta=2.0))
>>> ood = OODReconstruction(VAE(input_shape), config=config)
>>> ood.fit(train_data)

predict(x, batch_size=int(10000000000.0), ood_type='instance')¶

Predict whether instances are out of distribution or not.

Parameters:¶

x : ArrayLike¶: Input data for out-of-distribution prediction.
batch_size : int, default 1e10¶: Number of instances to process in each batch.
ood_type : "feature" | "instance", default "instance"¶: Predict out-of-distribution at the ‘feature’ or ‘instance’ level.

Raises:¶

ValueError – x input data must be unit interval [0-1].

Returns:¶

Dictionary containing the outlier predictions for the selected level, and the OOD scores for the data including both ‘instance’ and ‘feature’ (if present) level scores.

Return type:¶

OODOutput

score(x, batch_size=int(10000000000.0))¶

Compute the out of distribution scores for a given dataset.

Parameters:¶

x : ArrayLike¶: Input data to score.
batch_size : int, default 1e10¶: Number of instances to process in each batch. Use a smaller batch size if your dataset is large or if you encounter memory issues.

Raises:¶

ValueError – x input data must be unit interval [0-1].

Returns:¶

An object containing the instance-level and feature-level OOD scores.

Return type:¶

OODScoreOutput

Classes¶

`Config`	Configuration for OODReconstruction detector training and threshold computation.