tensorboard integration
yorkie opened this issue · comments
To integrate gracefully the tensorboard, we need the followings:
- reimplement the tensorboard/backend in
@pipcook/boa
- integrate the tensorboard/webapp.
- support real-time and remote logging by Pipcook SDK.
tensorboard/backend
Tensorboard uses backend as the service for the tensorboard frontend website, it uses the low-level Python modules and exports them by HTTP APIs. To integrate the backend in Pipboard, what we need is to create those HTTP APIs via vercel FaaS framework, and use boa to get functions from the low-level Python packages.
tensorboard/webapp
Tensorboard uses Polymer framework to write the web application, to integrate the tensorboard/webapp in Pipboard, we have the following choices:
- Build it separately, and integrate that via iframe.
- Build it with the current framework(iceworks).
real-time & remote logging
What we want to integrate the tensorboard is to show the training progress on Pipboard in real-time, thus it requires Pipcook SDK(Pipboard-side) is able to fetch the tensorboard-compatible logs in real-time. @FeelyChau, is there a way to add a custom event? IMHO -1 on adding a special API for tensorboard logs.