Awesome

Reproducible Source Codes for the $M^4$ XAI Benchmark

Note that this code is for the reproducible purpose. A wrapper for these modulars is on the plan.

This code is implemented using PaddlePaddle and InterpretDL. A demo using HuggingFace models and InterpretDL can be found here. A demo using Pytorch models and InterpretDL can be found here.

Paper:

Xuhong Li, Mengnan Du, Jiamin Chen, Yekun Chai, Himabindu Lakkaraju, Haoyi Xiong. “M4: A Unified XAI Benchmark for Faithfulness Evaluation of Feature Attribution Methods across Metrics, Modalities and Models.” Neurips 2023, Dataset and Benchmark Track. https://openreview.net/forum?id=6zcfrSz98y.

Pipeline:

A first glance of the source codes:

• benchmark_data: stores the data that are used by the benchmark.

• benchmark-cv: benchmark for the image modality.

• benchmark-nlp: benchmark for the text modality.

• training: the process for training on the synthetic dataset of yellow patches.

• training_nlp: Finetuning the NLP models on the MovieReview dataset.

Steps

To perform the benchmark, it is required to follow the steps as below:

1. Prepare the dataset. ImageNet or MovieReview. Note that when using these datasets, you agree to the terms from ImageNet or MovieReview respectively.

(Step 2 is optional. One can download or prepare a trained model.)

2. Train models (only required for NLP models). See training_nlp/train.sh.

(Step 3 is only required for the SynScore calculation.)

3. Train models on the synthetic dataset. See training/train.sh.

4. Compute explanations. See benchmark-cv/run_expl.sh for image modality. See benchmark-nlp/run_expl.sh for text modality. We compute all the explanation results at once and save them locally. This can avoid repeating the computations.

5. Eval MoRF,ABPC. See benchmark-cv/run_eval.sh for image modality. See benchmark-nlp/run_eval.sh for text modality.

6. Eval Others. See benchmark-cv/run_eval-pgs.sh for image modality. See benchmark-nlp/run_eval2.sh for text modality.

Feel free to open issues.