Update README.md

README.md

@@ -1,3 +1,31 @@
----
-license: mit
----
+---
+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).