BIM - Benchmark Interpretability Method

This repository contains dataset, models, and metrics for benchmarking interpretability methods (BIM) described in paper:

• Title: "BIM: Towards Quantitative Evaluation of Interpretability Methods with Ground Truth"
• Authors: Sherry (Mengjiao) Yang, Been Kim

Upon using this library, please cite:

@Article{BIM2019,
  title = {{BIM: Towards Quantitative Evaluation of Interpretability Methods with Ground Truth}},
  author = {Yang, Mengjiao and Kim, Been},
  year = {2019}
}

Setup

Run the following from the home directory of this repository to install python dependencies, download BIM models, download MSCOCO and MiniPlaces, and construct BIM dataset.

pip install bim
source scripts/download_models.sh
source scripts/download_datasets.sh
python scripts/construct_bim_dataset.py

Dataset

Images in data/obj and data/scene are the same but have object and scene labels respectively, as shown in the figure above. val_loc.txt records the top-left and bottom-right corner of the object and val_mask has the binary masks of the object in the validation set. Additional sets and their usage are described in the table below.

Models

Models in models/obj, models/scene, and models/scene_only are trained on data/obj, data/scene, and data/scene_only respectively. Models in models/scenei for i in {1...10} are trained on images where dog is added to i scene classes, and the rest scene classes do not contain any added objects. All models are in TensorFlow's SavedModel format.

Metrics

BIM metrics compare how interpretability methods perform across models (model contrast), across inputs to the same model (input dependence), and across functionally equivalent inputs (input independence).

Model contrast scores

Given images that contain both objects and scenes, model contrast measures the difference in attributions between the model trained on object labels and the model trained on scene labels.

Input dependence rate

Given a model trained on scene labels, input dependence measures the percentage of inputs where the addition of objects results in the region being attributed as less important.

Input independence rate

Given a model trained on scene-only images, input independence measures the percentage of inputs where a functionally insignificant patch (e.g., a dog) does not affect explanations significantly.

Evaluate saliency methods

To compute model contrast score (MCS) over randomly selected 10 images, you can run

python bim/metrics.py --metrics=MCS --num_imgs=10

To compute input dependence rate (IDR), change --metrics to IDR. To compute input independence rate (IIR), you need to first constructs a set of functionally insignificant patches by running

python scripts/construct_delta_patch.py

and then evaluate IIR by running

python bim/metrics.py --metrics=IIR --num_imgs=10

Evaluate TCAV

TCAV is a global concept attribution method whose MCS can be measured by comparing the TCAV scores of a particular object concept for the object model and the scene model. Run the following to compute the TCAV scores of the dog concept for the object model.

python bim/run_tcav.py --model=obj

Disclaimer

This is not an officially supported Google product.