vihari / AAA

Active and Efficient Estimation of Accuracy Surfaces

Code for the NeurIPS 2021 paper:
Active Assessment of Prediction Services as Accuracy Surface Over Attribute Combinations.

Table of Contents

1. Package dependencies
2. Datasets
3. Code files and their utility
4. Instructions for running
   1. For Estimation only Experiments of Section 4.5 Reported in Table 2, Figure 3
   2. For Estimation + Exploration reported in Figure 4
   3. Impact of Calibration of Section 4.7 and Figure 5

Package dependencies

• Python 1.6
• Pytorch 1.6.0
• GPytorch
• Pickle
• tqdm

Datasets

The following datasets are used. If you wish to run the MF-* tasks, you are required to download the CelebA dataset from the below link and place it in datasets of your home folder. You do not have to download COCOS dataset for running AC-* tasks. See "Instructions for Running" section.

• CelebA (http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html)
• COCOS (https://github.com/nightrome/cocostuff)

Code files and their utility

The task and the corresponding file name is mentioned is as below.

Task	File name
MF-CelebA	celeba.py
MF-IMDB	imdb.py
AC-COCOS	cocos.py
AC-COCOS10K	cocos10k.py

Important source tree files and their utiltity is shown in table below.

File name	Utility
src/beta_gp_rloss_explorer.py	Implements BetaGP, BetaGP-SL, BetaGP-SLP
src/simple_explorer.py	Implements Beta-I
src/bern_gp_rloss_explorer.py	Implements BernGP
src/betaab_gp_rloss_explorer.py	BetaGPab
src/likelihoods/beta_gp_likelihoods.py	Implements different data likelihoods. The routines: `baseline`, `simplev3_rloss` correspond to BetaGP-SL, BetaGP-SLP
src/dataset.py	Abstract classes for Dataset, service model
src/dataset_fitter.py	Routines for calibration, sampling from arms, warm start etc.
notebooks/ToyGP.ipynb	Has the code for the simple setting described in the Appendix.

Instructions for running

1. Setup your environment by installing all the package dependencies listed above.
2. Download data.zip from this drive link (300MB). Unzip data.zip in the working directory.
3. (Optional) If you wish to run MF-* tasks, download CelebA dataset and place it in $HOME/datasets.
4. Follow the description below for the specific python commands.

For Estimation only Experiments of Section 4.5 Reported in Table 2, Figure 3

Beta-I python <task>.py --explorer simple --seed 0 --ablation

BernGP python <task>.py --explorer bern_gp_rloss --seed 0 --ablation

BetaGP
python <task>.py --explorer beta_gp_rloss --seed 0 --approx_type baseline --no_scale_loss --ablation

BetaGP-SL python <task>.py --explorer beta_gp_rloss --seed 0 --approx_type baseline --ablation

BernGP-SLP
python <task>.py --explorer beta_gp_rloss --seed 0 --approx_type simplev3 --ablation

<task> should be set to one of celeba, celeba_private, cocos3, cocos3_10k.

For Estimation + Exploration reported in Figure 4

Use the same commands described in the previous section but with --ablation flag removed.

Impact of Calibration of Section 4.7 and Figure 5

python <task>.py --explorer simple --seed 0 --ablation --sample_type [option]

Sample type options should be one of correctedwep, correctednoep, raw and correspond to Cal:Full, Cal:Temp, Cal:Simple respectively of the paper.
<task> should be set to one of celeba, celeba_private, cocos3, cocos3_10k.