Skip to main content

This package is a plug-n-play PyTorch reimplementation of Influence Functions. Influence Functions were introduced in the paper Understanding Black-box Predictions via Influence Functions by Pang Wei Koh and Percy Liang (ICML2017).

Project description

Influence Functions for PyTorch

This is a PyTorch reimplementation of Influence Functions from the ICML2017 best paper: Understanding Black-box Predictions via Influence Functions by Pang Wei Koh and Percy Liang. The reference implementation can be found here: link.

Why Use Influence Functions?

Influence functions help you to debug the results of your deep learning model in terms of the dataset. When testing for a single test image, you can then calculate which training images had the largest result on the classification outcome. Thus, you can easily find mislabeled images in your dataset, or compress your dataset slightly to the most influential images important for your individual test dataset. That can increase prediction accuracy, reduce training time, and reduce memory requirements. For more details please see the original paper linked here.

Influence functions can of course also be used for data other than images, as long as you have a supervised learning problem.


  • Python 3.6 or later
  • PyTorch 1.0 or later
  • NumPy 1.12 or later

To run the tests, further requirements are:

  • torchvision 0.3 or later
  • PIL


You can either install this package directly through pip:

pip3 install --user pytorch-influence-functions

Or you can clone the repo and import it as a package after it's in your PATH.


Calculating the influence of the individual samples of your training dataset on the final predictions is straight forward.

The most barebones way of getting the code to run is like this:

import pytorch_influence_functions as ptif

# Supplied by the user:
model = get_my_model()
trainloader, testloader = get_my_dataloaders()

config = ptif.get_default_config()

influences, harmful, helpful = ptif.calc_img_wise(config, model, trainloader, testloader)

# do someting with influences/harmful/helpful

Here, config contains default values for the influence function calculation which can of course be changed. For details and examples, look here.

Background and Documentation

The precision of the output can be adjusted by using more iterations and/or more recursions when approximating the influence.


config is a dict which contains the parameters used to calculate the influences. You can get the default config by calling ptif.get_default_config().

I recommend you to change the following parameters to your liking. The list below is divided into parameters affecting the calculation and parameters affecting everything else.

Misc parameters

  • save_pth: Default None, folder where to save s_test and grad_z files if saving is desired
  • outdir: folder name to which the result json files are written
  • log_filename: Default None, if set the output will be logged to this file in addition to stdout.

Calculation parameters

  • seed: Default = 42, random seed for numpy, random, pytorch
  • gpu: Default = -1, -1 for calculation on the CPU otherwise GPU id
  • calc_method: Default = img_wise, choose between the two calculation methods outlined here.
  • DataLoader object for the desired dataset
    • train_loader and test_loader
  • test_sample_start_per_class: Default = False, per class index from where to start to calculate the influence function. If False, it will start from 0. This is useful if you want to calculate the influence function of a whole test dataset and manually split the calculation up over multiple threads/ machines/gpus. Then, you can start at various points in the dataset.
  • test_sample_num: Default = False, number of samples per class starting from the test_sample_start_per_class to calculate the influence function for. E.g. if your dataset has 10 classes and you set this value to 1, then the influence functions will be calculated for 10 * 1 test samples, one per class. If False, calculates the influence for all images.
  • recursion_depth: Default = 5000, recursion depth for the s_test calculation. Greater recursion depth improves precision.
  • r: Default = 1, number of s_test calculations to take the average of. Greater r averaging improves precision.
  • damp: Default = 0.01, damping factor during s_test calculation.
  • scale: Default = 25, scaling factor during s_test calculation.

Modes of computation

This packages offers two modes of computation to calculate the influence functions. The first mode is called calc_img_wise, during which the two values s_test and grad_z for each training image are computed on the fly when calculating the influence of that single image. The algorithm moves then on to the next image. The second mode is called calc_all_grad_then_test and calculates the grad_z values for all images first and saves them to disk. Then, it'll calculate all s_test values and save those to disk. Subsequently, the algorithm will then calculate the influence functions for all images by reading both values from disk and calculating the influence base on them. This can take significant amounts of disk space (100s of GBs) but with a fast SSD can speed up the calculation significantly as no duplicate calculations take place. This is the case because grad_z has to be calculated twice, once for the first approximation in s_test and once to combine with the s_test vector to calculate the influence. The paper has a ton of more detail on that.

TL;DR: The recommended way is using calc_img_wise unless you have a crazy fast SSD and lots of free storage space.

Output variables


Is a dict/json containting the influences calculated of all training data samples for each test data sample. The dict structure looks similiar to this:

    "0": {
        "label": 3,
        "num_in_dataset": 0,
        "time_calc_influence_s": 129.6417362689972,
        "influence": [
        "harmful": [
        "helpful": [
    "1": {
        "label": 8,
        "num_in_dataset": 1,
        "time_calc_influence_s": 121.8709237575531,
        "influence": [


Harmful is a list of numbers, which are the IDs of the training data samples ordered by harmfulness. If the influence function is calculated for multiple test images, the harmfulness is ordered by average harmfullness to the prediction outcome of the processed test samples.


Helpful is a list of numbers, which are the IDs of the training data samples ordered by helpfulness. If the influence function is calculated for multiple test images, the helpfulness is ordered by average helpfulness to the prediction outcome of the processed test samples.



  • [x] makes variable names etc. dataset independent
  • [x] remove all dataset name checks from the code
  • [ ] ability to disable shell output eg for display_progress from the config
  • [ ] add proper result plotting support
  • [ ] add a dataloader for training on the most influential samples only


  • [ ] make the config a class, so that it can readjust itself, for example when the r and recursion_depth values can be lowered without big impact
  • [ ] check killing data augmentation!?
  • [ ] in in load_s_test, load_grad_z don't hard code the filenames


  • [ ] integrate myPy type annotations (static type checking)
  • [ ] Use multiprocessing to calc the influence
  • [ ] use r"doc" docstrings like pytorch

Project details

Release history Release notifications

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for pytorch-influence-functions, version 0.1.0
Filename, size File type Python version Upload date Hashes
Filename, size pytorch_influence_functions-0.1.0.tar.gz (16.2 kB) File type Source Python version None Upload date Hashes View hashes

Supported by

Elastic Elastic Search Pingdom Pingdom Monitoring Google Google BigQuery Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN SignalFx SignalFx Supporter DigiCert DigiCert EV certificate StatusPage StatusPage Status page