Awesome

quanda is currently under active development. Note the release version to ensure reproducibility of your work. Expect changes to API.

📑 Shortcut to paper!

🐼 Library overview

Training data attribution (TDA) methods attribute model output on a specific test sample to the training dataset that it was trained on. They reveal the training datapoints responsible for the model's decisions. Existing methods achieve this by estimating the counterfactual effect of removing datapoints from the training set (Koh and Liang, 2017; Park et al., 2023; Bae et al., 2024) tracking the contributions of training points to the loss reduction throughout training (Pruthi et al., 2020), using interpretable surrogate models (Yeh et al., 2018) or finding training samples that are deemed similar to the test sample by the model (Caruana et. al, 1999; Hanawa et. al, 2021). In addition to model understanding, TDA has been used in a variety of applications such as debugging model behavior (Koh and Liang, 2017; Yeh et al., 2018; K and Søgaard, 2021; Guo et al., 2021), data summarization (Khanna et al., 2019; Marion et al., 2023; Yang et al., 2023), dataset selection (Engstrom et al., 2024; Chhabra et al., 2024), fact tracing (Akyurek et al., 2022) and machine unlearning (Warnecke et al., 2023).

Although there are various demonstrations of TDA’s potential for interpretability and practical applications, the critical question of how TDA methods should be effectively evaluated remains open. Several approaches have been proposed by the community, which can be categorized into three groups:

Library Features

• Unified TDA Interface: quanda provides a unified interface for various TDA methods, allowing users to easily switch between different methods.
• Metrics: quanda provides a set of metrics to evaluate the effectiveness of TDA methods. These metrics are based on the latest research in the field.
• Benchmarking: quanda provides a benchmarking tool to evaluate the performance of TDA methods on a given model, dataset and problem. As many TDA evaluation methods require access to ground truth, our benchmarking tools allow to generate a controlled setting with ground truth, and then compare the performance of different TDA methods on this setting.

Supported TDA Methods

Metrics

• Linear Datamodeling Score (Park et al., 2023): Measures the correlation between the (grouped) attribution scores and the actual output of models trained on different subsets of the training set. For each subset, the linear datamodeling score compares the actual model output to the sum of attribution scores from the subset using Spearman rank correlation.

• Identical Class / Identical Subclass (Hanawa et al., 2021): Measures the proportion of identical classes or subclasses in the top-1 training samples over the test dataset. If the attributions are based on similarity, they are expected to be predictive of the class of the test datapoint, as well as different subclasses under a single label.

• Model Randomization (Hanawa et al., 2021): Measures the correlation between the original TDA and the TDA of a model with randomized weights. Since the attributions are expected to depend on model parameters, the correlation between original and randomized attributions should be low.

• Top-K Cardinality (Barshan et al., 2020): Measures the cardinality of the union of the top-K training samples. Since the attributions are expected to be dependent on the test input, they are expected to vary heavily for different test points, resulting in a low overlap (high metric value).

• Mislabeled Data Detection (Koh and Liang, 2017): Computes the proportion of noisy training labels detected as a function of the percentage of inspected training samples. The samples are inspected in order according to their global TDA ranking, which is computed using local attributions. This produces a cumulative mislabeling detection curve. We expect to see a curve that rapidly increases as we check more of the training data, thus we compute the area under this curve

• Shortcut Detection (Yolcu et al., 2024): Assuming a known shortcut, or Clever-Hans effect has been identified in the model, this metric evaluates how effectively a TDA method can identify shortcut samples as the most influential in predicting cases with the shortcut artifact. This process is referred to as Domain Mismatch Debugging in the original paper.

• Mixed Datasets (Hammoudeh and Lowd, 2022): In a setting where a model has been trained on two datasets: a clean dataset (e.g. CIFAR-10) and an adversarial (e.g. zeros from MNIST), this metric evaluates how well the model ranks the importance (attribution) of adversarial samples compared to clean samples when making predictions on an adversarial example.

Benchmarks

quanda comes with a few pre-computed benchmarks that can be conveniently used for evaluation in a plug-and-play manner. We are planning to significantly expand the number of benchmarks in the future. The following benchmarks are currently available:

🔬 Getting Started

Installation

To install the latest release of quanda use:

pip install quanda
pip install captum@git+https://github.com/pytorch/captum

quanda requires Python 3.7 or later. It is recommended to use a virtual environment to install the package.

Basic Usage

In the following usage examples, we will be using the SimilarityInfluence data attribution from Captum.

Using Metrics

To begin using quanda metrics, you need the following components:

• Trained PyTorch Model (model): A PyTorch model that has already been trained on a relevant dataset. As a placeholder, we used the layer name "avgpool" below. Please replace it with the name of one of the layers in your model.
• PyTorch Dataset (train_set): The dataset used during the training of the model.
• Test Dataset (eval_set): The dataset to be used as test inputs for generating explanations. Explanations are generated with respect to an output neuron corresponding to a certain class. This class can be selected to be the ground truth label of the test points, or the classes predicted by the model. In the following we will use the predicted labels to generate explanations. Next, we demonstrate how to evaluate explanations using the Model Randomization metric.

from torch.utils.data import DataLoader
from tqdm import tqdm

from quanda.explainers.wrappers import CaptumSimilarity
from quanda.metrics.heuristics import ModelRandomizationMetric

We now create our explainer. The device to be used by the explainer and metrics is inherited from the model, thus we set the model device explicitly.

DEVICE = "cpu"
model.to(DEVICE)

explainer_kwargs = {
    "layers": "avgpool",
    "model_id": "default_model_id",
    "cache_dir": "./cache"
}
explainer = CaptumSimilarity(
    model=model,
    train_dataset=train_set,
    **explainer_kwargs
)

The ModelRandomizationMetric needs to instantiate a new explainer to generate explanations for a randomized model. These will be compared with the explanations of the original model. Therefore, explainer_cls is passed directly to the metric along with initialization parameters of the explainer for the randomized model.

explainer_kwargs = {
    "layers": "avgpool",
    "model_id": "randomized_model_id",
    "cache_dir": "./cache"
}
model_rand = ModelRandomizationMetric(
    model=model,
    train_dataset=train_set,
    explainer_cls=CaptumSimilarity,
    expl_kwargs=explainer_kwargs,
    correlation_fn="spearman",
    seed=42,
)

We now start producing explanations with our TDA method. We go through the test set batch-by-batch. For each batch, we first generate the attributions using the predicted labels, and we then update the metric with the produced explanations to showcase how to concurrently handle the explanation and evaluation processes.

test_loader = DataLoader(eval_set, batch_size=32, shuffle=False)
for test_tensor, _ in tqdm(test_loader):
    test_tensor = test_tensor.to(DEVICE)
    target = model(test_tensor).argmax(dim=-1)
    tda = explainer.explain(
        test_tensor=test_tensor,
        targets=target
    )
    model_rand.update(test_data=test_tensor, explanations=tda, explanation_targets=target)

print("Randomization metric output:", model_rand.compute())

Using Pre-assembled Benchmarks

The pre-assembled benchmarks allow us to streamline the evaluation process by downloading the necessary data and models, and running the evaluation in a single command. The following code demonstrates how to use the mnist_subclass_detection benchmark:

from quanda.explainers.wrappers import CaptumSimilarity
from quanda.benchmarks.downstream_eval import SubclassDetection

DEVICE = "cpu"
model.to(DEVICE)

explainer_kwargs = {
    "layers": "avgpool",
    "model_id": "default_model_id",
    "cache_dir": "./cache"
}

subclass_detect = SubclassDetection.download(
    name="mnist_subclass_detection",
    cache_dir=cache_dir,
    device="cpu",
)
score = subclass_detect.evaluate(
    explainer_cls=CaptumSimilarity,
    expl_kwargs=explain_fn_kwargs,
    batch_size=batch_size,
)["score"]
print(f"Subclass Detection Score: {score}")

Assembling a benchmark from existing components

Next, we demonstrate assembling a benchmark with assets that the user has prepared. As in the Using Metrics section, we will assume that the user has already trained model on train_set, and a corresponding eval_set to be used for generating and evaluating explanations.

from quanda.explainers.wrappers import CaptumSimilarity
from quanda.benchmarks.ground_truth import TopKCardinality

DEVICE = "cpu"
model.to(DEVICE)

explainer_kwargs = {
    "layers": "avgpool",
    "model_id": "default_model_id",
    "cache_dir": "./cache"
}

We now have everything we need, we can just assemble the benchmark and run it. This will encapsulate the process of instantiating the explainer, generating explanations and using the TopKCardinalityMetric to evaluate them.

topk_cardinality = TopKCardinality.assemble(
    model=model,
    train_dataset=train_set,
    eval_dataset=eval_set,
)
score = topk_cardinality.evaluate(
    explainer_cls=CaptumSimilarity,
    expl_kwargs=explain_fn_kwargs,
    batch_size=batch_size,
)["score"]
print(f"Top K Cardinality Score: {score}")

Generating the benchmark object from scratch

Some evaluation strategies require a controlled setup or a different strategy of using attributors to evaluate them. For example, the MislabelingDetectionMetric requires a dataset with known mislabeled examples. It computes the self-influence of training points to evaluate TDA methods. Therefore, it is fairly complicated to train a model on a mislabeled dataset, and then using the metric object or assembling a benchmark object to run the evaluation. While pre-assembled benchmarks allow to use pre-computed assets, quanda Benchmark objects provide the generate interface, which allows the user to prepare this setup from scratch.

As in previous examples, we assume that train_set refers to a vanilla training dataset, without any modifications for evaluation. Furthermore, we assume model refers to a torch Module, but in this example we do not require that model is trained. Finally, n_classes is the number of classes in the train_set.

import torch

from quanda.explainers.wrappers import CaptumSimilarity
from quanda.benchmarks.downstream_eval import MislabelingDetection

DEVICE = "cpu"
model.to(DEVICE)

explainer_kwargs = {
    "layers": "avgpool",
    "model_id": "default_model_id",
    "cache_dir": "./cache"
}

For mislabeling detection, we will train a model from scratch. quanda allows to use Lightning Trainer objects. If you want to use Lightning trainers, model needs to be an instance of a Lightning LightningModule. Alternatively, you can use an instance of quanda.utils.training.BaseTrainer. In this example, we use a very simple training setup via the quanda.utils.training.Trainer class.

trainer = Trainer(
    max_epochs=100,
    optimizer=torch.optim.SGD,
    lr=0.01,
    criterion=torch.nn.CrossEntropyLoss(),
)

We can now call the generate method to instantiate our MislabelingDetection object and directly start the evaluation process with it. The generate method takes care of model training using trainer, generation of explanations and their evaluation.

mislabeling_detection = MislabelingDetection.generate(
    model=model,
    base_dataset=train_set,
    n_classes=n_classes,
    trainer=trainer,
)
score = mislabeling_detection.evaluate(
    explainer_cls=CaptumSimilarity,
    expl_kwargs=explain_fn_kwargs,
    batch_size=batch_size,
)["score"]
print(f"Mislabeling Detection Score: {score}")

More detailed examples can be found in the tutorials folder.

Custom Explainers

In addition to the built-in explainers, quanda supports the evaluation of custom explainer methods. This section provides a guide on how to create a wrapper for a custom explainer that matches our interface.

Your custom explainer should inherit from the base Explainer class provided by quanda. The first step is to initialize your custom explainer within the __init__ method.

from quanda.explainers.base import Explainer

class CustomExplainer(Explainer):
    def __init__(self, model, train_dataset, **kwargs):
        super().__init__(model, train_dataset, **kwargs)
        # Initialize your explainer here

The core of your wrapper is the explain method. This function should take test samples and their corresponding target values as input and return a 2D tensor containing the influence scores.

• test: The test batch for which explanations are generated.
• targets: The target values for the explanations.

Ensure that the output tensor has the shape (test_samples, train_samples), where the entries in the train samples dimension are ordered in the same order as in the train_dataset that is being attributed.

def explain(
  self,
  test_tensor: torch.Tensor,
  targets: Union[List[int], torch.Tensor]
) -> torch.Tensor:
    # Compute your influence scores here
    return influence_scores

By default, quanda includes a built-in method for calculating self-influence scores. This base implementation computes all attributions over the training dataset, and collects the diagonal values in the attribution matrix. However, you can override this method to provide a more efficient implementation. This method should calculate how much each training sample influences itself and return a tensor of the computed self-influence scores.

def self_influence(self, batch_size: int = 1) -> torch.Tensor:
    # Compute your self-influence scores here
    return self_influence_scores

For detailed examples, we refer to the existing explainer wrappers in quanda.

⚠️ Usage Tips and Caveats

• Controlled Setting Evaluation: Many metrics require access to ground truth labels for datasets, such as the indices of the "shorcut samples" in the Shortcut Detection metric, or the mislabeling (noisy) label indices for the Mislabeling Detection Metric. However, users often may not have access to these labels. To address this, we recommend either using one of our pre-built benchmark suites (see Benchmarks section) or generating (generate method) a custom benchmark for comparing explainers. Benchmarks provide a controlled environment for systematic evaluation.

• Caching: Many explainers in our library generate re-usable cache. The cache_id and model_id parameters passed to various class instances are used to store these intermediary results. Ensure each experiment is assigned a unique combination of these arguments. Failing to do so could lead to incorrect reuse of cached results. If you wish to avoid re-using cached results, you can set the load_from_disk parameter to False.

• Explainers Are Expensive To Calculate: Certain explainers, such as TracInCPRandomProj, may lead to OutOfMemory (OOM) issues when applied to large models or datasets. In such cases, we recommend adjusting memory usage by either reducing the dataset size or using smaller models to avoid these issues.

📓 Tutorials

We have included a few tutorials to demonstrate the usage of quanda:

• Explainers: shows how different explainers can be used with quanda
• Metrics: shows how to use the metrics in quanda to evaluate the performance of a model
• Benchmarks: shows how to use the benchmarking tools in quanda to evaluate a data attribution method

To install the library with tutorial dependencies, run:

pip install quanda[tutorials]

👩‍💻Contributing

We welcome contributions to quanda! You could contribute by:

• Opening an issue to report a bug or request a feature.
• Submitting a pull request to fix a bug, add a new explainer wrapper, a new metric, or another feature.

A detailed guide on how to contribute to quanda can be found here.

🔗Citation

@misc{bareeva2024quandainterpretabilitytoolkittraining,
      title={Quanda: An Interpretability Toolkit for Training Data Attribution Evaluation and Beyond},
      author={Dilyara Bareeva and Galip Ümit Yolcu and Anna Hedström and Niklas Schmolenski and Thomas Wiegand and Wojciech Samek and Sebastian Lapuschkin},
      year={2024},
      eprint={2410.07158},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2410.07158},
}