CVPR-NAS is an annual competition that compares how well Neural Search Architectures can apply machine learning methods to unseen datasets. Competitors must register with a website so that they can download a starting kit and datasets, then use these resources to complete their NAS algorithm.
Their code based entry is then uploaded to the same website which must then allocate resources and run their submitted code. This code run is time limited to one hour. There will only be a limited number of GPUs available to run the code, so jobs may be queued and executed in the order received.
Between 5-10 submissions are expected per day during the allowed submission period. Competitors will ideally have email notification of their code status. It is expected that code will be uploaded in a zip format.
- Stephen McGough - (stephen.mcgough@newcastle.ac.uk)
- Rob Geada - (rob@geada.net)
- David Towers - (d.towers2@newcastle.ac.uk)
- Nik Khadijah Nik Aznan - (nik.nik-aznan@newcastle.ac.uk)
- Amir Atapour-Abarghouei - (amir.atapour-abarghouei@durham.ac.uk)
Rebecca Osselton
Newcastle University
(rebecca.osselton@newcastle.ac.uk)
This website is Vue.js based front-end that connects to a back-end application created using Strapi. CVPR-NAS-API
The competition has 3 stages. Users are able to upload zip files which are received by the Strapi back-end and recorded. Files are then passed to an Azure batch service with a unique id and run using specific datasets. Once the submissions have been executed via the batch pool, results are returned to the back-end and stored against th rignal submission. The website is able to query the back-end for each submission results and display a ranking of competitors based on runtimes and scores.
Vue-cli, Node.js, yarn
Clone the repo into a suitable directory. Create an .env file containing value for BASE_URL e.g. BASE_URL=localhost:1338.
yarn serve to start the webserver.
yarn build to build the application ready for deployment. Source code is minified and compiled in a 'dist' directory. Contents of the dist folder can then form the basis of a static front-end. The application is currently manually deployed to Azure (RSE Team General -> Storage Accounts -> cvprnas -> Blob Containers -> $web) using Azure Storage Explorer. Files and folders are deleted and then re-uploaded from the lastest compiled contents of the dist folder.
Main and dev only