robert1003 / heldout-influence-estimation

Heldout Influence Estimation

This repository contains code to load and inspect the models trained with a random heldout subset of the training set. The predictions of those models can be used to estimate useful statistics such as memorization, influence, and consistency scores, as demonstrated in the following papers.

• Vitaly Feldman*, Chiyuan Zhang*. What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation. *equal contribution. 2020. (project website)
• Ziheng Jiang*, Chiyuan Zhang*, Kunal Talwar, Michael C. Mozer. Characterizing Structural Regularities of Labeled Data in Overparameterized Models. *equal contribution. 2020. (project website)

In addition to the pre-computed statistics released in the project website of each respective paper, we also release the model checkpoints from those pre-trained models to facilitate future research.

Dependencies

This code depends on the following open source libraries

• Tensorflow 2
• absl
• Tensorflow Datasets
• Sonnet

Model Checkpoints

The model checkpoints can be downloaded from Google Cloud Storage. You can download them with the gsutil tool or directly via HTTP.

• Inception models trained on CIFAR10, 239GB: gs://gresearch/heldout-training-checkpoints/cifar10-inception.tar.gz
• Inception models trained on CIFAR100, 244GB: gs://gresearch/heldout-training-checkpoints/cifar100-inception.tar.gz
• Resnet50 models trained on ImageNet, 367GB: gs://gresearch/heldout-training-checkpoints/imagenet-resnet50.tar.gz

The MD5 sum of the files are:

bc3b75203381709f8224810946be14b1  cifar10-inception.tar.gz
f87218368528afa6e2407c8410f1ffa1  cifar100-inception.tar.gz
64c39ef7702462b83959af9de622b4b8  imagenet-resnet50.tar.gz

Each experiment consists of a collection of checkpoints for a certain neural network architecture trained (with the heldout procedure) on a certain dataset. The files are organized in the following way:

<experiment_name (e.g. cifar10-inception)>
  +-- <subset_ratio (e.g. 0.1)
  |     +-- <run_id (e.g. 0)>
  |     |     +-- aux_arrays.npz
  |     |     +-- checkpoints
  |     |           +-- <ckpt-epoch>
  |     +-- 1
  |     +-- ...
  +-- 0.2
  +-- ...

Depending on the experiment, there might be multiple subset ratios, and a different number of runs for each subset ratio. The <ckpt-epoch> is the checkpoint files in the Tensorflow format for the final training epoch. The aux_arrays.npz is a numpy array exported file containing the following information of each experiment:

• correctness_<split>: a binary indicator array of the prediction correctness of the final trained model, where the split could be train, removed and test. Here train + removed is the full training set of the original dataset.
• index_<split>: an integer array for the index of each example in the correctness_<split> array. This is useful when counting statistics across different experiments where the evaluation might be done in different orders. Note for train and removed, the index indicates the id of each example in the original full training set, therefore compatible across different experiments.
• subsample_idx: the list of indices of examples from the original training set that is selected to be used in this run.

Code Demo

Code for loading and inspecting checkpoints

In the file demo.py, we include some demo code showing how to construct models, load checkpoints, run the models for evaluation and cross check with the results found in aux_arrays.npz.

Code used for training

In the mnist-example directory, we provide a self contained code example showing the whole pipeline of training with subsets of training examples, and estimating the influence and memorization based on the training statistics collected on multiple runs with different random subsets.

Disclaimer

This is not an officially supported Google product.