examples

dattri examples

This folder contains bite-sized examples that can help users build their own applications with dattri.

Noisy label detection

This section contains examples using different attributors to detect noisy labels in various datasets.

Use influence function to detect noisy labels in Mnist10 + Logistic regression.

Use TracIN to detect noisy labels in Mnist10 + MLP.

Use TRAK to detect noisy labels in CIFAR10 + ResNet-9.

Use pretrained checkpoints and pre-calculated ground truth

This section contains examples using the pretrained checkpoints and pre-calculated ground truth provided by dattri to evaluate the data attribution methods.

Use pre-trained Mnist10 + MLP benchmark setting and evaluate Influence Function (CG) algorithm by LDS

Use pre-trained Mnist10 + LR benchmark setting and evaluate TRAK algorithm by LOO correlation

Use pre-trained MNIST10 + MLP benchmark setting and evaluate TRAK + dropout ensemble by LDS

Use pre-trained MNIST10 + MLP benchmark setting and evaluate LoGra by LDS

Estimate the brittleness

This section contains examples using attribution scores to estimate the brittleness of a model.

Use influence function to estimate the brittleness of losigitc regression trained on Mnist10

Data cleaning

This section contains examples using attribution scores to find the data points that can be removed from the training set and improve the test performance.

Use influence function to find the low-quality data points in MNIST-10 and evaluate the performance

Customized retraining and ground truth calculation

This section shows how the user can customize training and target function and retrain in LDS settings.

Retrain MLP on MNIST-10 and obtain new ground truth, then compare it with the data attribution score from full MNIST-10 + MLP model

Trajectory-Specific LOO

This section contains examples using Trajectory-Specific LOO methods to estimate the temporal influence of training data.

Use DVEmb to estimate the temporal influence of training data in Mnist10 + MLP