This code replicates the experiments from the following paper:
Pang Wei Koh*, Kai-Siang Ang*, Hubert H. K. Teo*, and Percy Liang
On the Accuracy of Influence Functions for Measuring Group Effects
We have a reproducible, executable, and Dockerized version of the paper on CodaLab.
Influence functions estimate the effect of removing particular training points on a model without needing to retrain it. They are based on a first-order approximation that is accurate for small changes in the model, and so are commonly used for studying the effect of individual points in large datasets. However, we often want to study the effects of large groups of training points, e.g., to diagnose batch effect or apportion credit between different data sources. Removing such large groups can result in significant changes to the model. Are influence functions still accurate in this setting? In this paper, we find that across many different types of groups and in a range of real-world datasets, the influence of a group correlates surprisingly well with its actual effect, even if the absolute and relative error can be large. Our theoretical analysis shows that such correlation arises under certain settings but need not hold in general, indicating that real-world datasets have particular properties that keep the influence approximation well-behaved.
- Python 2.7
- NumPy 1.16.5
- SciPy 1.2.2
- matplotlib 2.2.4
- seaborn 0.9.0
- jupyter 1.0.0
- scikit-learn 0.20.4
- Pandas 0.24.2
- Spacy 2.0.0
- Tensorflow 1.14.0 (GPU or CPU)
- Tensorflow Probability 0.7.0
You can install the Python dependencies locally using pip and
requirements.txt. Note that our requirements.txt includes the GPU version
of Tensorflow. It is also possible to run the code on the CPU using
Tensorflow. We highly recommend isolating these dependencies with virtualenv
or some other Python environment manager.
virtualenv -p python2.7 ~/influence/
source ~/influence/bin/activate
pip install -r requirements.txt
Then, pre-download the Spacy model used for preprocessing the Enron dataset.
python -m spacy download en_core_web_sm
We also provide a Dockerfile for the execution environment. A pre-built image
bbbert/influence:5 can be found in
DockerHub.
bbbert/influence:5 is the image used in the CodaLab worksheet.
The code maintains and expects a particular folder structure, parameterized by two base directories: a data directory, and an output directory to store the
result of experiment runs. All scripts in scripts/ that use these directories will expect the --data-dir and --out-dir CLI arguments, which default
to [repository_root]/data/ and [repository_root]/output/ respectively. Then, we will download and preprocess datasets and save experiment outputs
in the following structure:
.
├─── data/
│ ├─── hospital/ (A data directory, not necessarily a dataset ID)
│ ├─── spam/
│ └─── ...
└─── output/
├─── ss_logreg/ (An experiment ID)
│ ├─── spam_ihvp-explicit_seed-0_sizes-0.0025-0.25_num-10_rel-reg-1e-07/ (One particular run of the experiment)
│ └─── ...
├─── counterexamples/
└─── ...
Notice that each experiment in experiments/ has a fixed experiment ID, and different runs of the same experiment are stored as subfolders
in [out_dir]/[experiment_id]/[run_id]/. The run ID is a string defined by the experiment that is meant to contain all the relevant
parameters for the run.
All experiments and operations are performed by executing the scripts in
scripts/. The code expects the repository root to be in the Python path. We
recommend appending the repository root to the PYTHONPATH environment
variable, and executing the scripts from the repository root. For example, to
download and preprocess all the datasets, run
export PYTHONPATH=${PYTHONPATH}:$(pwd)
python scripts/load_all_datasets.py
For convenience, the codalab/ folder contains scripts used in our
CodaLab
worksheet. The sequence of runs that generates the worksheet can be found in
codalab/codalab_run.sh, except for the MultiNLI dataset, which we have
preprocessed and uploaded as a CodaLab bundle.