Neural network monitoring
This project is based on Reportr, the open source dashboard. For instructions specific to Reportr, please see the project homepage.
Mri-server constitutes the web-based monitoring portion of Mri. When used together with the Mri-app for Caffe or the Mri-python-client, it allows you to watch the progress of your networks as they train from anywhere, as well as test multiple hyperparameters or architectures at once.
The project is entirely open source and you can host your own Mri-server instance on your own server or Heroku.
For installation instructions, see the documentation
Reportr is configured using environment variables.
| Name | Description |
|---|---|
| PORT | Port for running the application, default is 5000 |
| MONGODB_URL | Url for the mongoDB database |
| REDIS_URL | (Optional) Url for a redis database when using worker mode |
| AUTH_USERNAME | Username for authentication |
| AUTH_PASSWORD | Password for authentication |
The Mri clients already know how to talk to the server, and will automatically create reports and visualizations as you train networks. Simply modify the Mri-app configuration file to properly interface with the server as a dispatch. See API_REFERENCE.md for full API specifications.
Reportr can easily be scaled on Heroku (and compatibles), use the REDIS_URL to enable a task queue between workers and web processes.