---
license: mit
task_categories:
- summarization
language:
- en
tags:
- nas
pretty_name: epsinas
size_categories:
- 100M<n<1B
---

# Release data for `epsinas`
Release data for the results reported in the *epsinas* paper: arxiv.org/abs/2302.04406.


# Contents
Contains the results of the experiments with NAS-Bench-101, NAS-Bench-201 and NAS-Bench NLP benchmark NAS spaces.
We provide pickle data files and figures for the main *epsinas* evaluation and ablation studies.

The files are split according to the search space.
Depending on the benchmark, the directories may contain:

- **evaluation**. Evaluation of *epsinas* for all neural architecture from the benchmark.
- **ablation**. Various ablation studies related to the given dataset.
- **integration**. Results of the integration with Evolutionary search and Random search.
- **average_performance**. Reports average accuracy of the architecture selected with *epsinas* metric when choosing from a subset of 100/1000 architectures over 500 runs (Table 3 in the paper). 
- **computational_efficiency**. Contains the results of the computational efficiency study.

For NAS-Bench-201, these subdirs are further split based on the dataset used for evaluation (CIFAR10, CIFAR100, or ImageNet).