- Automated and standardized deep learning experiment management
- Based on GitHub and Google Sheets, accesible for individual and team
- Automatic experiment repeating, results synchronization and error recovery
- Using repository to manage code and log.
- Every project has a folder in the repo.
- Every project has a sheet to record experiment settings and results.
- The repo access is granted by Github access token
- Get personal access token: Creating a personal access token
- In "Select scopes", choose "repo"
- Sheets accesses are granted by credentials json file
- A Google Cloud Platform project with the API enabled. To create a project and enable an API, refer to Create a project and enable the API
- Authorization credentials for a desktop application. To learn how to create credentials for a desktop application, refer to Create credentials
- The detailed steps are listed below:
- Create a GCP Project
- Enable Google Sheets API in Dashboard of APIs and services
- Configure OAuth consent screen of APIs and services (External is OK)
- Add your google account to Test users
- Create a new OAuth client ID with type Desktop app in Credentials
- Download in OAuth 2.0 Client IDs
- You got the credentials json file
- At the first time of use, you need to get authorization code from a url. Then the auth information will be save into a token file and auth will be skipped next time.
-
1 Repo
Github repository for code and log
-
1 Sheet
Google Sheet for hyperparameters and results
-
N Worker
Training on any workers
- Add a row in Sheet
- Push code to Repo
- Wait for experiment result
- Fetch code from Github
- Training
- Push result to Sheet and push log to Repo