cloud-cv / digits

The version of DIGITS supported by CloudCV as a part of its services. Originally from the Deep Learning GPU Training System

DIGITS

DIGITS (the Deep Learning GPU Training System) is is a webapp for training deep learning models.

Installation

If you are using Ubuntu 14.04, you can install pre-built packages for the latest released version via apt-get. Visit docs/UbuntuInstall.md for instructions.

To build DIGITS and its requirements from source, visit docs/BuildDigits.md for instructions.

Once you have installed DIGITS, visit docs/GettingStarted.md for an introductory walkthrough.

Get help

Installation issues

• First, check out the instructions above
• Then, ask questions on our user group

Usage questions

• First, check out the Getting Started page
• Then, ask questions on our user group

Bugs and feature requests

• Please let us know by filing a new issue
• Bonus points if you want to contribute by opening a pull request!
  • You will need to send a signed copy of the Contributor License Agreement to [email protected] before your change can be accepted.

digits's Issues

Integrating changes back into NVIDIA/DIGITS

Hi guys, great discussion this morning!

I'd like to get a clear picture of (1) what you've done and (2) how feasible it would be to integrate each of those things back upstream. Here's what I've got so far:

Add support for workspaces

Is each DIGITS instance owned by a single user? Or is each DIGITS instance partitioned into multiple workspaces? Can a Job be owned by multiple users or by a group of users?

Why did you switch to Django? Did you replace Flask entirely? How much work was it?

Interface with Cloud storage

How did you handle authentication? Do you download the images locally or just read them directly from a URL?

When saving things back to the cloud, did you save an entire job directory or just specific files? Did you save the status.pickle files?

Add pretrained model as a new ModelJob

We're discussing doing a re-architecture which would make this easier. The ModelZoo download is an especially cool feature that we'd like to have in DIGITS.

Store activations to disk

Did you save the weights, too? Did you save the whole activations even for fully connected layers? How big do those files get?

Sorry this is an absurd number of questions. I guess we could break the discussion into separate threads if that's helpful? Pointers to specific commits with clear commit messages for each of these features would be much appreciated.