einspace: Searching for Neural Architectures from Fundamental Operations
Authors: Linus Ericsson, Miguel Espinosa Minano, Chenhongyi Yang, Antreas Antoniou, Amos J. Storkey, Shay Cohen, Steven McDonagh, Elliot J. Crowley
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Using this search space, we perform experiments to find novel architectures as well as improvements on existing ones on the diverse Unseen NAS datasets. We show that competitive architectures can be obtained by searching from scratch, and we consistently find large improvements when initialising the search with strong baselines. |
| Researcher Affiliation | Academia | Linus Ericsson1 Miguel Espinosa1 Chenhongyi Yang1 Antreas Antoniou2 Amos Storkey2 Shay B. Cohen2 Steven Mc Donagh1 Elliot J. Crowley1 1 School of Engineering 2 School of Informatics University of Edinburgh |
| Pseudocode | No | The paper describes its grammar and operations in detail but does not provide structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Project page: https://linusericsson.github.io/einspace Code: https://github.com/linusericsson/einspace |
| Open Datasets | Yes | We adopt a diverse benchmark suite from the recent paper on Unseen NAS [16], containing datasets at different difficulties across vision, language, audio, and further modalities. While Unseen NAS forms the basis of this section, we run additional experiments on the diverse NASBench360 benchmark [55] in Appendix C.1 |
| Dataset Splits | Yes | We run individual searches on these datasets, that are each split into train, validation and test sets. Add NIST: 45,000 are used for training, 15,000 are used for validation, and 10,000 images are used for testing. |
| Hardware Specification | Yes | All our experiments ran on our two internal clusters with the following infrastructure: AMD EPYC 7552 48-Core Processor with 1000GB RAM and 8 NVIDIA RTX A5500 with 24GB of memory AMD EPYC 7452 32-Core Processor with 400GB RAM and 7 NVIDIA A100 with 40GB of memory |
| Software Dependencies | No | The paper mentions software like the SGD optimizer and Python packages like einops, but it does not specify version numbers for these or other key software components, which is required for a reproducible description of ancillary software. |
| Experiment Setup | Yes | Table 5: Hyperparameters for each dataset, taken from Geada et al. [16] for Unseen NAS and Tu et al. [55] for NASBench360. We set the number of search epochs to be one eighth of the evaluation epochs to speed up the search stage without significantly compromising on signal quality. ... All networks are trained using the SGD optimizer with momentum of 0.9. The values for learning rate, weight decay, batch size and more can be found in Table 5. |