ood_detectors 0.0.6

No project description provided

ood_detectors

OOD Detectors is a Python package that offers a suite of algorithms designed to identify out-of-distribution samples in datasets. This is crucial for maintaining the reliability and accuracy of machine learning models when faced with unfamiliar data.

---

Table of Contents

• Installation
• License

Installation

To install OOD Detectors, run the following command:

pip install ood_detectors

Usage

This package includes several OOD detection algorithms, each tailored to different aspects of OOD detection:

• Likelihood Based: SubSDE_DDM, VPSDE_DDM and VESDE_DDM are likelihood-based methods that use different variations stochastic differential equations for DDMS to detect OOD samples.

• Residual: This method employs the least significant eigen vector for OOD detection.

All detectors share a common interface:

1. Initialize the detector with necessary hyperparameters.
2. Fit the model using fit() with the training data.
3. Use predict() to obtain OOD scores for new data samples.

Example

import ood_detectors.likelihood as likelihood

ood_detector = likelihood.SubSDE_DDM(feat_dim).to('cuda')
train_loss = ood_detector.fit(train_data, n_epochs, batch_size)
scores = ood_detector.predict(test_data, batch_size)

from ood_detectors import Residual

ood_detector = Residual()
train_loss = ood_detector.fit(train_data)
scores = ood_detector.predict(test_data)

low-level interface

The low-level interface allows you to customize the training process and access the model's internal components.

import ood_detectors.likelihood as likelihood
import ood_detectors.sde as sde_lib 
import ood_detectors.models as models
import ood_detectors.losses as losses
...
sde = sde_lib.subVPSDE(beta_min=beta_min, beta_max=beta_max)

model = models.SimpleMLP(
    channels=feat_dim,
    bottleneck_channels=bottleneck_channels,
    num_res_blocks=num_res_blocks,
    time_embed_dim=time_embed_dim,
    dropout=dropout,
)

optimizer = functools.partial(
                torch.optim.Adam,
                lr=lr,
                betas=(beta1, beta2),
                eps=eps,
                weight_decay=weight_decay,
                )

ood_detector = likelihood.Likelihood(
    sde = sde,
    model = model,
    optimizer = optimizer,
    ).to(device)

update_fn = functools.partial(
    losses.SDE_EMA_LRS_BF16_GradClip, 
    ema_rate=ema_rate,
    total_steps=(len(train_blob['data'])//batch_size) * n_epochs,
    grad_clip=grad_clip,
    continuous=continuous,
    reduce_mean=reduce_mean,
    likelihood_weighting=likelihood_weighting,
    )

train_loss = ood_detector.fit(train_data, update_fn, batch_size)

Create a custom component

You can create a custom component by doing the same thing as the library does. Good luck!

Evaluate

To assess the performance of the OOD detectors, you can utilize the following metrics:

• AUC: Area under the ROC curve
• FPR95: False positive rate when the true positive rate is 95%

import ood_detectors.eval_utils as eval_utils
score_id = ood_detector.predict(train_data)
score_ref = ood_detector.predict(reference_data)
print(f"Train AUC: {eval_utils.auc(-score_ref, -score_id):.2%}")
print(f"Train FPR95: {eval_utils.fpr95(-score_ref, -score_id):.2%}")

results = eval_utils.eval_ood(ood_detector, train_data, reference_data, ood_data, batch_size, verbose=False)
plot_utils.plot(results, id_name, ood_names, encoder=embedding, model=ood_detector.name,
                train_loss=train_loss,
                config=conf,
                )

License

ood_detectors is distributed under the terms of the apache-2.0 license.