Reproducible machine learning with PyTorch and Quilt

tutorial

#1

In this article, we'll train a PyTorch model to perform super-resolution imaging, a technique for gracefully upscaling images. We'll use the Quilt data registry to snapshot training data and models as versioned data packages.


This is a companion discussion topic for the original entry at https://blog.paperspace.com/reproducible-data-with-pytorch-and-quilt/

#2

As an option to Quilt, consider using and open-source DVC. It does not provide Python API yet, it’s working more like git-lfs (metafiles with pointers), but is much more flexible in terms of the remote storage one could use, for example on-premise via SSH, Azure, S3, GCP, etc. What’s nice is that for DVC versioning of data files and models is just a basic scenario.

Would be interesting to discuss and write another article on how to run DVC projects using paperspace.

Disclaimer: I’m one of the authors of the DVC project :slight_smile:


#3

@shcheklein DVC looks really interesting. If you have the time, it would be great if you could post a quick tutorial here.

Quick note: Be sure to add a disclaimer if you represent the topic you are writing about :slight_smile:


#4

The link is broken for the training project on paperspace — looks like the path needs to change from /workspace to /files.


#5

@colllin Which link is broken exactly?