| | --- |
| | 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). |
