DeepSample

This repository provides the artifacts for DeepSample execution. The experiments on 9 DNN for classification and 2 DNN for regression can be replicated by running the script 'run_DeepSample.sh'. Once the experiments are completed, all the results will be generated in the 'Results' folder.

DeepSample folder

The folder "DeepSample" contains all the ".jar" files and the source code. The source code is organized as follows:

• The "main" folder contains all the java files for the jar generation.
• "selector.classification" contains the source code of the implemented sampling algorithms for classification, plus DeepEST.
• "selector.regression" contains the source code of the implemented sampling algorithms for regression, plus DeepEST.
• "utility" and "utility.regression" contain utilities and data structures useful for the selectors.

The code has been developed with Eclipse.

Other samplers

CES and SRS implementations are available at: 'https://github.com/Lizenan1995/DNNOpAcc'. The results can be replicated by:

1. cloning the repository;
2. importing the models in 'CE method/MNIST/normal';
3. importing and running in 'CE method/MNIST/normal' the files available in the 'CES_SRS' folder of our repository ('crossentropy*.py' files).

In the following, we reported the chosen layers (LAY parameter in 'crossentropy*.py') required to run CES for each model:

• A: -2
• B: -4
• C: -2
• D: -4
• E: -4
• F: -2
• G: -2
• H: -4
• I: -4
• Dave_orig: -3
• Dave_dropout: -3

Neural Networks availability

The trained models are included in the 'models' folder. Model G, Dave_orig and Dave_dropout models are available at 'https://file.io/0405r74sgwqV'. The folder 'dataset' contains the dataset with the predictions and the auxiliary variables for all models. The source code for the classification models is available at: 'https://github.com/ICOS-OAA/ICOS.git'

Paper results

The folder 'Results_paper' contains the results reported in the paper.

A Python notebook is provided for the practitioners to query the results, for instance to ask for the best performing technique (based on their occurrences in the top-3 rankings) given a main objective (low RMSE, RMedSE or large failure exposure ability) and a configuration as input (e.g., small sample size, MNIST dataset, confidence as auxiliary variable) (notebook)

Requirements and Dependencies

The provided code requires Java 8. The "libs" folder contains all the libraries required to run the experiments.