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.
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
. - install it using
python setup.py install
- install it using
python setup.py develop
(if you want to edit the code)
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()
ptif.init_logging()
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.
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.
save_pth
: DefaultNone
, folder where to saves_test
andgrad_z
files if saving is desiredoutdir
: folder name to which the result json files are writtenlog_filename
: DefaultNone
, if set the output will be logged to this file in addition tostdout
.
seed
: Default = 42, random seed for numpy, random, pytorchgpu
: Default = -1,-1
for calculation on the CPU otherwise GPU idcalc_method
: Default = img_wise, choose between the two calculation methods outlined here.DataLoader
object for the desired datasettrain_loader
andtest_loader
test_sample_start_per_class
: Default = False, per class index from where to start to calculate the influence function. IfFalse
, it will start from0
. 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 thetest_sample_start_per_class
to calculate the influence function for. E.g. if your dataset has 10 classes and you set this value to1
, then the influence functions will be calculated for10 * 1
test samples, one per class. IfFalse
, calculates the influence for all images.
recursion_depth
: Default = 5000, recursion depth for thes_test
calculation. Greater recursion depth improves precision.r
: Default = 1, number ofs_test
calculations to take the average of. Greater r averaging improves precision.- Combined, the original paper suggests that
recursion_depth * r
should equal the training dataset size, thus the above values ofr = 10
andrecursion_depth = 5000
are valid for CIFAR-10 with a training dataset size of 50000 items. damp
: Default = 0.01, damping factor durings_test
calculation.scale
: Default = 25, scaling factor durings_test
calculation.
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. Most importantnly however, s_test
is only
dependent on the test sample(s). While one grad_z
is used to estimate the
initial value of the Hessian during the s_test
calculation, this is
insignificant. grad_z
on the other hand is only dependent on the training
sample. Thus, in the calc_img_wise
mode, we throw away all grad_z
calculations even if we could reuse them for all subsequent s_test
calculations, which could potentially be 10s of thousands. However, as stated
above, keeping the grad_z
s only makes sense if they can be loaded faster/
kept in RAM than calculating them on-the-fly.
TL;DR: The recommended way is using calc_img_wise
unless you have a crazy
fast SSD, lots of free storage space, and want to calculate the influences on
the prediction outcomes of an entire dataset or even >1000 test samples.
Visualised, the output can look like this:
The test image on the top left is test image for which the influences were calculated. To get the correct test outcome of ship, the Helpful images from the training dataset were the most helpful, whereas the Harmful images were the most harmful. Here, we used CIFAR-10 as dataset. The model was ResNet-110. The numbers above the images show the actual influence value which was calculated.
The next figure shows the same but for a different model, DenseNet-100/12. Thus, we can see that different models learn more from different images.
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": [
-0.00016939856868702918,
4.3426321099104825e-06,
-9.501376189291477e-05,
...
],
"harmful": [
31527,
5110,
47217,
...
],
"helpful": [
5287,
22736,
3598,
...
]
},
"1": {
"label": 8,
"num_in_dataset": 1,
"time_calc_influence_s": 121.8709237575531,
"influence": [
3.993639438704122e-06,
3.454859779594699e-06,
-3.5805194329441292e-06,
...
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.
- makes variable names etc. dataset independent
- 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
- add some visualisation of the outcome
- add recreation of some graphs of the original paper to verify implementation
- allow custom save name for the influence json
- make the config a class, so that it can readjust itself, for example
when the
r
andrecursion_depth
values can be lowered without big impact - check killing data augmentation!?
- in
calc_influence_function.py
inload_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