Toolkit for quantitative evaluation of data attribution methods in PyTorch.

  [](https://github.com/psf/black)   [](https://codecov.io/gh/dilyabareeva/quanda)  [](https://quanda.readthedocs.io/en/latest/?badge=latest) [](https://arxiv.org/abs/2410.07158) **quanda** _is currently under active development. Note the release version to ensure reproducibility of your work. Expect changes to API._ [📑 Shortcut to paper!](https://arxiv.org/pdf/2410.07158) ## 🐼 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](https://proceedings.mlr.press/v70/koh17a.html); [Park et al., 2023](https://proceedings.mlr.press/v202/park23c.html); [Bae et al., 2024](https://arxiv.org/abs/2405.12186)) tracking the contributions of training points to the loss reduction throughout training ([Pruthi et al., 2020](https://proceedings.neurips.cc/paper/2020/hash/e6385d39ec9394f2f3a354d9d2b88eec-Abstract.html)), using interpretable surrogate models ([Yeh et al., 2018](https://proceedings.neurips.cc/paper/2018/hash/8a7129b8f3edd95b7d969dfc2c8e9d9d-Abstract.html)) or finding training samples that are deemed similar to the test sample by the model ([Caruana et. al, 1999](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2232607/); [Hanawa et. al, 2021](https://openreview.net/forum?id=9uvhpyQwzM_)). In addition to model understanding, TDA has been used in a variety of applications such as debugging model behavior ([Koh and Liang, 2017](https://proceedings.mlr.press/v70/koh17a.html); [Yeh et al., 2018](https://proceedings.neurips.cc/paper/2018/hash/8a7129b8f3edd95b7d969dfc2c8e9d9d-Abstract.html); [K and Søgaard, 2021](https://arxiv.org/abs/2111.04683); [Guo et al., 2021](https://aclanthology.org/2021.emnlp-main.808)), data summarization ([Khanna et al., 2019](https://proceedings.mlr.press/v89/khanna19a.html); [Marion et al., 2023](https://openreview.net/forum?id=XUIYn3jo5T); [Yang et al., 2023](https://openreview.net/forum?id=4wZiAXD29TQ)), dataset selection ([Engstrom et al., 2024](https://openreview.net/forum?id=GC8HkKeH8s); [Chhabra et al., 2024](https://openreview.net/forum?id=HE9eUQlAvo)), fact tracing ([Akyurek et al., 2022](https://aclanthology.org/2022.findings-emnlp.180)) and machine unlearning ([Warnecke et al., 2023](https://arxiv.org/abs/2108.11577)). 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:

Ground Truth

As some of the methods are designed to approximate LOO effects, ground truth can often be computed for TDA evaluation. However, this counterfactual ground truth approach requires retraining the model multiple times on different subsets of the training data, which quickly becomes computationally expensive. Additionally, this ground truth is shown to be dominated by noise in practical deep learning settings, due to the inherent stochasticity of a typical training process (Basu et al., 2021; Nguyen et al., 2023).

Downstream Task Evaluators

To remedy the challenges associated with ground truth evaluation, the literature proposes to assess the utility of a TDA method within the context of an end-task, such as model debugging or data selection (Koh and Liang, 2017; Khanna et al., 2019; Karthikeyan et al., 2021).

Heuristics

Finally, the community also used heuristics (desirable properties or sanity checks) to evaluate the quality of TDA techniques. These include comparing the attributions of a trained model and a randomized model (Hanawa et al., 2021) and measuring the amount of overlap between the attributions for different test samples (Barshan et al., 2020).

quanda is designed to meet the need of a comprehensive and systematic evaluation framework, allowing practitioners and researchers to obtain a detailed view of the performance of TDA methods in various contexts. ### 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 | Method Name | Repository | Reference | |----------------------------|------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------| | Similarity Influence | [Captum](https://github.com/pytorch/captum/tree/master) | [Caruana et al., 1999](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2232607/) | | Arnoldi Influence Function | [Captum](https://github.com/pytorch/captum/tree/master) | [Schioppa et al., 2022](https://arxiv.org/abs/2112.03052); [Koh and Liang, 2017](https://proceedings.mlr.press/v70/koh17a.html) | | TracIn | [Captum](https://github.com/pytorch/captum/tree/master) | [Pruthi et al., 2020](https://proceedings.neurips.cc/paper/2020/hash/e6385d39ec9394f2f3a354d9d2b88eec-Abstract.html) | | TRAK | [TRAK](https://github.com/MadryLab/trak) | [Park et al., 2023](https://proceedings.mlr.press/v202/park23c.html) | | Representer Point Selection | [Representer Point Selection](https://github.com/chihkuanyeh/Representer_Point_Selection) | [Yeh et al., 2018](https://proceedings.neurips.cc/paper/2018/hash/8a7129b8f3edd95b7d969dfc2c8e9d9d-Abstract.html) | ### Metrics - **Linear Datamodeling Score** ([Park et al., 2023](https://proceedings.mlr.press/v202/park23c.html)): 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](https://openreview.net/forum?id=9uvhpyQwzM_)): 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](https://openreview.net/forum?id=9uvhpyQwzM_)): 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](http://proceedings.mlr.press/v108/barshan20a/barshan20a.pdf)): 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](https://proceedings.mlr.press/v70/koh17a.html)): 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](https://arxiv.org/abs/2402.12118)): Assuming a known [shortcut](https://www.nature.com/articles/s42256-020-00257-z), or [Clever-Hans](https://www.nature.com/articles/s41467-019-08987-4) 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](https://dl.acm.org/doi/abs/10.1145/3548606.3559335)): 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:

Benchmark	Modality	Model	Metric	Type
mnist_top_k_cardinality	Vision	MNIST	TopKCardinalityMetric	Heuristic
mnist_mixed_datasets			MixedDatasetsMetric	Heuristic
mnist_class_detection			ClassDetectionMetric	Downstream-Task-Evaluator
mnist_subclass_detection			SubclassDetectionMetric	Downstream-Task-Evaluator
mnist_mislabeling_detection			MislabelingDetectionMetric	Downstream-Task-Evaluator
mnist_shortcut_detection			ShortcutDetectionMetric	Downstream-Task-Evaluator
mnist_linear_datamodeling_score			LinearDatamodelingMetric	Ground Truth

## 🔬 Getting Started ### Installation To install the latest release of **quanda** use: ```setup 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. ### Usage In the following, we provide a quick guide to **quanda** usage. To begin using **quanda**, ensure you have the following: - **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 Batches (`test_tensor`) and Explanation Targets (`target`)**: A batch of test data (`test_tensor`) and the corresponding explanation targets (`target`). Generally, it is advisable to use the model's predicted labels as the targets. In the following, we assume the existence of a `torch.utils.data.DataLoader` to load the test data in batches, with variable name `test_loader`. In the following usage examples, we will be using the `SimilarityInfluence` data attribution from `Captum`. #### Metrics Usage Next, we demonstrate how to evaluate explanations using the **Model Randomization** metric.

Step 1. Import dependencies and library components

```python import torch from torch.utils.data import DataLoader from tqdm import tqdm from quanda.explainers.wrappers import captum_similarity_explain, CaptumSimilarity from quanda.metrics.heuristics import ModelRandomizationMetric ```

Step 2. Create the explainer object

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. ```python 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 ) ```

Step 3. Initialize the metric

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. ```python 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, ) ```

Step 4. Iterate over test set and feed tensor batches first to explain, then to metric

```python for i, (test_tensor, target) in enumerate(tqdm(test_loader)): test_tensor, target = test_tensor.to(DEVICE), target.to(DEVICE) tda = explainer.explain( test_tensor=test_tensor, targets=target ) model_rand.update(test_data=test_tensor, explanations=tda, explanation_targets=target) print("Model heuristics metric output:", model_rand.compute()) ```

#### Benchmarks Usage 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. **Step 1** and **Step 2** from the previous section are still required to be executed before running the benchmark. The following code demonstrates how to use the `mnist_subclass_detection` benchmark:

Step 3. Load a pre-assembled benchmark and score an explainer

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

More detailed examples can be found in the [tutorials](tutorials) folder. ### Custom Explainers In addition to the built-in explainers, **quanda** supports the evaluatioon of custom explainer methods. This section provides a guide on how to create a wrapper for a custom explainer that matches our interface.

Step 1. Create an explainer class

Your custom explainer should inherit from the base [Explainer](quanda/explainers/base.py) class provided by **quanda**. The first step is to initialize your custom explainer within the `__init__` method. ```python 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 ```

Step 2. Implement the explain method

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

Step 3. Implement the self_influence method (Optional)

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. ```python 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](quanda/explainers/wrappers/captum_influence.py) [explainer](quanda/explainers/wrappers/representer_points.py) [wrappers](quanda/explainers/wrappers/trak_wrapper.py) 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](#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](tutorials) to demonstrate the usage of **quanda**: * [Explainers](tutorials/demo_explainers.ipynb): shows how different explainers can be used with **quanda** * [Metrics](tutorials/demo_metrics.ipynb): shows how to use the metrics in **quanda** to evaluate the performance of a model * [Benchmarks](tutorials/demo_benchmarks.ipynb): shows how to use the benchmarking tools in **quanda** to evaluate a data attribution method To install the library with tutorial dependencies, run: ```bash 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](CONTRIBUTING.md). ## 🔗Citation ```bibtex @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}, } ```